Research ArticleA Two-Stage Optimization Strategy forFuzzy Object-Based Analysis Using AirborneLiDAR and High-Resolution Orthophotos forUrban Road Extraction

Maher Ibrahim Sameen and Biswajeet Pradhan

Department of Civil Engineering Faculty of Engineering University Putra Malaysia 43400 Selangor Malaysia

Correspondence should be addressed to Biswajeet Pradhan biswajeet24gmailcom

Received 10 December 2016 Revised 24 January 2017 Accepted 8 February 2017 Published 28 February 2017

Academic Editor Eugenio Martinelli

Copyright copy 2017 Maher Ibrahim Sameen and Biswajeet Pradhan This is an open access article distributed under the CreativeCommons Attribution License which permits unrestricted use distribution and reproduction in any medium provided theoriginal work is properly cited

In the last decade object-based image analysis (OBIA) has been extensively recognized as an effective classificationmethod for veryhigh spatial resolution images or integrated data from different sources In this study a two-stage optimization strategy for fuzzyobject-based analysis using airborne LiDAR was proposed for urban road extraction The method optimizes the two basic steps ofOBIA namely segmentation and classification to realize accurate land cover mapping and urban road extraction This objectivewas achieved by selecting the optimum scale parameter to maximize class separability and the optimum shape and compactnessparameters to optimize the final image segments Class separability was maximized using the Bhattacharyya distance algorithmwhereas image segmentation was optimized using the Taguchi methodThe proposed fuzzy rules were created based on integrateddata and expert knowledge Spectral spatial and texture features were used under fuzzy rules by implementing the particle swarmoptimization techniqueThe proposed fuzzy rules were easy to implement and were transferable to other areas An overall accuracyof 82 and a kappa index of agreement (KIA) of 079 were achieved on the studied area when results were compared with referenceobjects created via manual digitization in a geographic information system The accuracy of road extraction using the developedfuzzy rules was 076 (producer) 085 (user) and 072 (KIA) Meanwhile overall accuracy was decreased by approximately 6when the rules were applied on a test site A KIA of 070 was achieved on the test site using the same rules without any changesTheaccuracy of the extracted urban roads from the test site was 072 (KIA) which decreased to approximately 016 Spatial information(ie elongation) and intensity from LiDAR were the most interesting properties for urban road extraction The proposed methodcan be applied to a wide range of real applications through remote sensing by transferring object-based rules to other areas usingoptimization techniques

1 Introduction

Urban road network information is essential for severalgeospatial and geographic modeling applications such asflood simulations [1] environmental studies [2] and trafficaccident analysis and prevention [3 4] Information onurban roads can also help decision makers make strategicand effective decisions which are important for sustainableurban planning and management LiDAR is one of the bestapproaches to gather 3D terrain information [5] Recently

LiDAR data have been used for road network extraction androad geometric modeling LiDAR provides highly accurateelevation data that can be used to extract 3D models ofdifferent types of objects along with their geometric char-acteristics Additional information such as image intensityand aerial orthophotos collected by LiDAR systems providesconsiderable opportunities for geometric road modeling

In the last decade object-based image analysis (OBIA)has been recognized as one of the most effective classifica-tion techniques for very high spatial resolution images or

HindawiJournal of SensorsVolume 2017 Article ID 6431519 17 pageshttpsdoiorg10115520176431519

2 Journal of Sensors

integrated data from different sources (eg aerial orthopho-tos and LiDAR data) [6] Compared with pixel-based clas-sification techniques OBIA provides additional informationthat can be used to improve the discrimination of land coverclasses [7] OBIA has two main steps First image segmenta-tion is the process of creating nonoverlapping homogeneousobjects from image pixels based on their spectral spatialand texture information In this step OBIA generally usesthe multiresolution segmentation (MRS) algorithm [8] Themain advantage of OBIA is that it processes satellite imagesin a manner similar to that of the human brain that isthis technique identifies landscape features as objects [9]The size and shape of image objects are controlled by threemain parameters namely scale shape and compactness asin the MRS algorithm [6] Scale controls the size of imageobjects whereas the other two parameters control the shapeof image objects in terms of geometry and smoothness Thescale parameter defines the allowed maximum heterogeneityof grid cell values within image objects Small-scale valuescreate small image objects which can be used to classifyspecific image objects with certain rule sets By contrastlarge values of the scale parameter generate relatively largeimage objects which typically lead to the loss of detailedinformation Therefore an appropriate scale parameter iscritical for accurate classification and information extractionfrom satellite imagesThe shape and compactness parametersare also essential The selection of their values directly affectsthe final image objects and the classification results

Selecting the values of segmentation parameters in OBIAclassification approach is achieved via a method called trial-and-error optimization [10] This visual estimation methodis subjective time-consuming and labor intensive Recentsegmentation parameter selection methods are based onsegmentation accuracy measures by comparing machine-created segments with hand-digitized ones Consideringthese advances in object-based analysis the development oftransferable rule sets remains a challenge mainly becausecreating appropriate image segments and identifyingmultipleoptimal scales for extraction for a particular feature aredifficult tasks Knowledge-based feature extraction can onlybe applied if the features of interest are segmented properlywhich require selecting multiple scales This technique willallow the use of spatial information As the shape andboundary of the considered features texture and contextualinformation for feature extraction are described efficiently bythe segmentation process Therefore appropriate segmenta-tion by selecting the optimum multiple scales in OBIA is acritical task for accurate feature extraction and for developingknowledge-based and transferable rule sets In this studya new strategy for selecting the optimum segmentationparameters by optimizing both image segments and classseparability is proposed

11 Related Studies Among the methods used to extract fea-tures from remote sensing data fuzzy object-based methodsare becoming increasingly popular Compared with that ofother approaches the accuracy of fuzzy-based classificationis less sensitive to thresholds [11] In addition fuzzy logicis considered an appropriate classification method for urban

feature extraction because it models uncertainties among theconsidered classes these uncertainties are typically found incomplex urban areas [12] Jabari and Zhang [11] developedfuzzy rules for delineating road networks from GeoEye-1and QuickBird images in an urban environment the resultsshowed that the roads could be extracted at 85 user accu-racy Spectral geometric and neighborhood informationwere used for road extraction Grote et al [13] developeda region-based method for extracting road networks fromLiDAR data in suburban areas Their method used veryhigh-resolution (VHR) color infrared orthophotos and adigital surfacemodel (DSM) Intensity normalized differencevegetation index area and length width elongation andconvexity and height from DSM were used to extract roadsusing the approach proposed by the authors The accuracyassessment demonstrated that correctness was approximately90 for extraction with DSM Sebari and He [14] developedfuzzy-based rules for OBIA classification approach to extractroads in an urban environment They used spectral andshape (ie elongation and size) information in the rule Acorrectness of 81 was achieved

In addition Li et al [15] provided a hierarchical methodfor urban road extraction Their method encompassed twomain steps The first step involved obtaining the roadregion of interest from a VHR image whereas the secondstep involved hierarchically representing this road regionof interest in a binary partition tree and extracting roadsat hierarchical levels Geometric features (ie compactnessand elongation) and structural features (ie histogramsand morphological profiles) were used to guide the regionmerging of the binary partition tree The method was testedon two different data sets and correctness rates of 85 and57 were achieved for the two data sets The low accuracy ofthe second study area was attributed to the dense urban areasHamedianfar and Shafri [16] developed fuzzy-based parame-ters for object-oriented classification for various impervioussurface and roofing materials The authors proposed a setof rules for road extraction from WorldView-2 images bycombining spectral and spatial features A user accuracy of92 was achieved following their road class rules

The literature review indicated that previous studies usedfuzzy object-based analysis without optimizing segmentationparameters As mentioned earlier the optimum values ofthe scale shape and compactness parameters are criticalfor accurate classification and feature extraction Althoughseveral studies have provided clear guidelines for selectingthe optimum values of segmentation parameters they dis-regarded class separability distance when selecting the opti-mum segmentation parameters In the present study a noveltwo-stage optimization strategy for OBIA is presented andtested by using it to extract road networks from LiDAR datacombined with high-resolution orthophotos In this strategythe Taguchi method is used to optimize the segmentationparameters whereas the Bhattacharyya distance is used tooptimize class separability

12 Study Areas The study areas were selected from someregions of the Universiti Putra Malaysia campus located inSelangor State Malaysia (Figure 1) Two pilot sites were

Journal of Sensors 3

Pahang

Perak

Johor

Kelantan

Kedah

Selangor

Trengganu

Negeri Sembilan

Melaka

Perlis

Pulau PinangPulau Pinang

Study area(method development)

Test site

6∘0㰀0㰀㰀N

4∘0㰀0㰀㰀N

2∘0㰀0㰀㰀N

6∘0㰀0㰀㰀N

4∘0㰀0㰀㰀N

2∘0㰀0㰀㰀N

100∘0㰀0㰀㰀E 102

∘0㰀0㰀㰀E 104

∘0㰀0㰀㰀E

100∘0㰀0㰀㰀E 102

∘0㰀0㰀㰀E 104

∘0㰀0㰀㰀E

Figure 1 Study sites used for method development and method evaluation

selected The first site was for method development whereasthe second site was for method transferability testing Geo-graphically the first study area is located in a region thatextends from the upper left longitude of 3∘00101584014481015840N andlatitude of 101∘421015840147E to the lower right longitude of3∘001015840007N and latitude of 101∘42101584044121015840E of the WGS84coordinate system Meanwhile the second study area islocated 4 km away from and to the right side of the first studyareaThe study and test areas are characterized by amixture ofurban asphalt road and vegetation features Buildings vary interms of roofing material shape and size asphalt roads varyin width and condition

2 Methodology

In this study a two-stage optimization strategy for fuzzyobject-based analysis was proposed to extract urban roadsfrom airborne LiDAR data The two-stage optimizationtechnique was used to optimize the image segmentation andimage classification processes simultaneously by selecting theoptimum scale parameter that maximized class separabilityand the optimum shape and compactness parameters thatoptimized the final image segments The method proposedin this study was developed through the following steps(Figure 2) First LiDAR point clouds and aerial orthophotoswere preprocessed and analyzed to prepare the input datafor urban road extraction Second the scale parameter wasoptimized using the Bhattacharyya distance algorithm [17]and the best ranges of the scale parameters were selectedfor further analysis Afterwards the shape and compactnessparameters were optimized using the Taguchi method [18]

After the three parameters of the segmentation algorithmwere optimized and selected the input image was segmentedusing the optimum values of scale shape and compactnessFinally fuzzy rules were developed using the decision treeand fuzzy 119888-mean algorithms which enabled the classifica-tion of image segments and the extraction of urban roads

21 LiDAR Data Acquisition and Processing The LiDAR dataused in this studywere collected onMarch 8 2013 using RieglLM Q5600 and Hassleblad 39Mp camera The device has aspatial resolution of 13 cm a laser-scanning angle of 60∘ and acamera angle of approximately minus45∘ In addition the postingdensity of the LiDAR data is 3-4 ptsm2

The collected LiDAR point clouds (ie LAS file) wereconverted into vector points in ArcGIS 103 software Thevector points created from the LiDAR point clouds had 2Dcoordinates (ie119883 and119884) and height information as attributedata To convert these vector points into raster format thefirst outliers from the point clouds were removed In thisresearch outliers were removed using ldquoLocate Outliersrdquo atool in ArcGIS 103 software This tool identifies anomalouselevation measurements from terrain TIN or LAS datasets that exceed a defined range of elevation values orhave slope characteristics inconsistent with the surroundingsurface (ESRI 2016) Afterwards the filtered point cloudswere converted into a raster surface at 013 cm spatial res-olution using a TIN-based interpolation method [4] Theproduced raster surface represented the DSM (Figure 3) witha spatial resolution the same as the orthophotos (013 cm)This allowed using the orthophoto and the produced DSMfor creating the image objects

4 Journal of Sensors

LiDAR Aerial orthophoto

DSM DEMIntensity G BR

Scale = 10 Shape = 05 Compactness = 05

Layer stacking

Input image

Calculate class separability (Bhattacharyya distance)

Create image objectsIterate

Optimal scale Shape = 05 Compactness = 05

Optimize shape and compactness parameters (Taguchi method)

Optimal scale Optimal shape Optimal compactness

Feature selection (particle swarm)

Develop fuzzy rules

Classify and extract urban roads

Create image objects

nDSM

i = 1 to 15

Figure 2 Overall steps for two-stage optimization and fuzzy object-based analysis for urban road extraction

The point clouds were then filtered to ground andnonground points using multiscale curvature classification(MCC) [19] MCC is an iterative multiscale algorithm forclassifying LiDAR returns into ground and nonground [19]The MCC algorithm was developed at the Moscow ForestrySciences Laboratory of theUnited States Forest Service RockyMountain Research Station It integrates curvature filteringwith a scale component and variable curvature toleranceDuring this stage a surface is interpolated at different res-olutions using the thin-plate spline method [19] and pointsare classified based on a progressive curvature thresholdparameter which increases as resolution coarsens to com-pensate for the slope effect because data are generalizedThe same interpolation technique was then used to generatethe digital elevation models (DEMs) for the study and testareas (Figure 4) Afterwards the DSM and DEM rasters

were used to produce the nDSM raster by subtracting DEMfrom DSM In addition the intensity raster was generatedby interpolating the ground point clouds with the relatedintensity attribute information of the point clouds (Figure 5)

Meanwhile the aerial orthophoto was first correctedgeometrically to fit exactly into the LiDAR point clouds usingthe intensity raster (Figure 6) Finally the LiDAR-derivedDEM andDSM intensity nDSM and aerial orthophoto wereused as a subset based on the boundary of the study areato prepare the final input data for method development andurban road extraction

22 Feature Subset Selection via Particle Swarm Optimization(PSO) In object-based or knowledgemodeling a large num-ber of features are present in the data sets however not all of

Journal of Sensors 5

0 04 0802(km)

DSM 1(meters)

High 9164

Low 3882

N

(a)

0 04 0802(km)

DSM 2(meters)

High 7787

Low 4794

N

(b)

Figure 3 DSMs derived from airborne LiDAR data (a) DSM of the study area used for method development and (b) DSM of test area

High 470157

Low 3918

0 04 0802(km)

N

DEM 1(meters)

(a)

High 6254

Low 4794

0 04 0802(km)

N

DEM 2(meters)

(b)

Figure 4 DEMs derived from LiDAR point clouds (a) DEM of the study area used for method development and (b) DEM of the test area

these features are useful for classifying and extracting infor-mation Therefore the first step in object-based analysis orknowledgemodeling is selecting features from those availablein the data sets Useful features from the data sets for a par-ticular object extraction process can be selected using opti-mization techniques Such techniques aim to select a smallnumber of relevant features to achieve similar or better clas-sification results than using all the available featuresThe finalfeatures to be selected should be relevant Including irrelevantand redundant features in the analysis may produce poorclassification results because of the course of dimensionality

Considering the aforementioned reason only relevant fea-tures should be selected for accurate modeling

Several optimization methods such as greedy search-based sequential forward selection and sequential backwardselection can be used for feature selection However thesetechniques suffer from various problems such as stagnationin local optima and high computational cost [20] An efficientglobal search technique such as evolutionary computation(EC) is required to address feature selection problems moreeffectively PSO is one of the EC techniques based onswarm intelligence [20] Compared with other techniques

6 Journal of Sensors

Intensity 1High 24203

Low 8

N

0 04 0802(km)

(a)

Intensity 2High 19267

Low 8

N

0 04 0802(km)

(b)

Figure 5 LiDAR intensity derived by interpolating the intensity attributes of LiDAR point clouds (a) intensity of the study area used formethod development and (b) intensity of the test area

N

0 04 0802(km)

(a)

0 04 0802(km)

N

(b)

Figure 6 Aerial orthophotos (a) study area used for method development and (b) test area

such as genetic algorithm and simulated annealing PSO iscomputationally less costly and converges more rapidly [20]Therefore this technique was used in this study

PSO was proposed by Kennedy and Eberhart in 1995[21 22] This technique is motivated by social behavior suchas birds flocking and fish schooling In PSO knowledge isoptimized through social interaction in the population inwhich thinking is not only personal but also social Theuseful features from the available features of eCognitionsoftware were selected using PSO optimization implementedvia MATLAB PSO was used to minimize the classificationerror rateThe fitness function is given in (1) which is used tominimize the classification error rate obtained by the selected

features during the evolutionary training process and thenumber of selected features [20 23]

119865 = 119908 times TrainER

119904

+ (1 minus 119908) times TestER119904

120572 times FeaturesAll Features

+ (1 minus 120572) times ER119904ERall

(1)

where119908 is a constant number (the weight of the classificationerror rate obtained from the training data) and 119908 isin [0 1]TrainER

119904

is the classification error rate obtained from theselected feature subset and the training subset data TestER

119904

isthe classification error rate obtained from the selected featuresubset and the testing subset data 120572 is a constant number (theweight of the number of selected features) Features stands

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 3

Pahang

Perak

Johor

Kelantan

Kedah

Selangor

Trengganu

Negeri Sembilan

Melaka

Perlis

Pulau PinangPulau Pinang

Study area(method development)

Test site

6∘0㰀0㰀㰀N

4∘0㰀0㰀㰀N

2∘0㰀0㰀㰀N

6∘0㰀0㰀㰀N

4∘0㰀0㰀㰀N

2∘0㰀0㰀㰀N

100∘0㰀0㰀㰀E 102

∘0㰀0㰀㰀E 104

∘0㰀0㰀㰀E

100∘0㰀0㰀㰀E 102

∘0㰀0㰀㰀E 104

∘0㰀0㰀㰀E

Figure 1 Study sites used for method development and method evaluation

selected The first site was for method development whereasthe second site was for method transferability testing Geo-graphically the first study area is located in a region thatextends from the upper left longitude of 3∘00101584014481015840N andlatitude of 101∘421015840147E to the lower right longitude of3∘001015840007N and latitude of 101∘42101584044121015840E of the WGS84coordinate system Meanwhile the second study area islocated 4 km away from and to the right side of the first studyareaThe study and test areas are characterized by amixture ofurban asphalt road and vegetation features Buildings vary interms of roofing material shape and size asphalt roads varyin width and condition

2 Methodology

In this study a two-stage optimization strategy for fuzzyobject-based analysis was proposed to extract urban roadsfrom airborne LiDAR data The two-stage optimizationtechnique was used to optimize the image segmentation andimage classification processes simultaneously by selecting theoptimum scale parameter that maximized class separabilityand the optimum shape and compactness parameters thatoptimized the final image segments The method proposedin this study was developed through the following steps(Figure 2) First LiDAR point clouds and aerial orthophotoswere preprocessed and analyzed to prepare the input datafor urban road extraction Second the scale parameter wasoptimized using the Bhattacharyya distance algorithm [17]and the best ranges of the scale parameters were selectedfor further analysis Afterwards the shape and compactnessparameters were optimized using the Taguchi method [18]

After the three parameters of the segmentation algorithmwere optimized and selected the input image was segmentedusing the optimum values of scale shape and compactnessFinally fuzzy rules were developed using the decision treeand fuzzy 119888-mean algorithms which enabled the classifica-tion of image segments and the extraction of urban roads

21 LiDAR Data Acquisition and Processing The LiDAR dataused in this studywere collected onMarch 8 2013 using RieglLM Q5600 and Hassleblad 39Mp camera The device has aspatial resolution of 13 cm a laser-scanning angle of 60∘ and acamera angle of approximately minus45∘ In addition the postingdensity of the LiDAR data is 3-4 ptsm2

The collected LiDAR point clouds (ie LAS file) wereconverted into vector points in ArcGIS 103 software Thevector points created from the LiDAR point clouds had 2Dcoordinates (ie119883 and119884) and height information as attributedata To convert these vector points into raster format thefirst outliers from the point clouds were removed In thisresearch outliers were removed using ldquoLocate Outliersrdquo atool in ArcGIS 103 software This tool identifies anomalouselevation measurements from terrain TIN or LAS datasets that exceed a defined range of elevation values orhave slope characteristics inconsistent with the surroundingsurface (ESRI 2016) Afterwards the filtered point cloudswere converted into a raster surface at 013 cm spatial res-olution using a TIN-based interpolation method [4] Theproduced raster surface represented the DSM (Figure 3) witha spatial resolution the same as the orthophotos (013 cm)This allowed using the orthophoto and the produced DSMfor creating the image objects

4 Journal of Sensors

LiDAR Aerial orthophoto

DSM DEMIntensity G BR

Scale = 10 Shape = 05 Compactness = 05

Layer stacking

Input image

Calculate class separability (Bhattacharyya distance)

Create image objectsIterate

Optimal scale Shape = 05 Compactness = 05

Optimize shape and compactness parameters (Taguchi method)

Optimal scale Optimal shape Optimal compactness

Feature selection (particle swarm)

Develop fuzzy rules

Classify and extract urban roads

Create image objects

nDSM

i = 1 to 15

Figure 2 Overall steps for two-stage optimization and fuzzy object-based analysis for urban road extraction

The point clouds were then filtered to ground andnonground points using multiscale curvature classification(MCC) [19] MCC is an iterative multiscale algorithm forclassifying LiDAR returns into ground and nonground [19]The MCC algorithm was developed at the Moscow ForestrySciences Laboratory of theUnited States Forest Service RockyMountain Research Station It integrates curvature filteringwith a scale component and variable curvature toleranceDuring this stage a surface is interpolated at different res-olutions using the thin-plate spline method [19] and pointsare classified based on a progressive curvature thresholdparameter which increases as resolution coarsens to com-pensate for the slope effect because data are generalizedThe same interpolation technique was then used to generatethe digital elevation models (DEMs) for the study and testareas (Figure 4) Afterwards the DSM and DEM rasters

were used to produce the nDSM raster by subtracting DEMfrom DSM In addition the intensity raster was generatedby interpolating the ground point clouds with the relatedintensity attribute information of the point clouds (Figure 5)

Meanwhile the aerial orthophoto was first correctedgeometrically to fit exactly into the LiDAR point clouds usingthe intensity raster (Figure 6) Finally the LiDAR-derivedDEM andDSM intensity nDSM and aerial orthophoto wereused as a subset based on the boundary of the study areato prepare the final input data for method development andurban road extraction

22 Feature Subset Selection via Particle Swarm Optimization(PSO) In object-based or knowledgemodeling a large num-ber of features are present in the data sets however not all of

Journal of Sensors 5

0 04 0802(km)

DSM 1(meters)

High 9164

Low 3882

N

(a)

0 04 0802(km)

DSM 2(meters)

High 7787

Low 4794

N

(b)

Figure 3 DSMs derived from airborne LiDAR data (a) DSM of the study area used for method development and (b) DSM of test area

High 470157

Low 3918

0 04 0802(km)

N

DEM 1(meters)

(a)

High 6254

Low 4794

0 04 0802(km)

N

DEM 2(meters)

(b)

Figure 4 DEMs derived from LiDAR point clouds (a) DEM of the study area used for method development and (b) DEM of the test area

these features are useful for classifying and extracting infor-mation Therefore the first step in object-based analysis orknowledgemodeling is selecting features from those availablein the data sets Useful features from the data sets for a par-ticular object extraction process can be selected using opti-mization techniques Such techniques aim to select a smallnumber of relevant features to achieve similar or better clas-sification results than using all the available featuresThe finalfeatures to be selected should be relevant Including irrelevantand redundant features in the analysis may produce poorclassification results because of the course of dimensionality

Considering the aforementioned reason only relevant fea-tures should be selected for accurate modeling

Several optimization methods such as greedy search-based sequential forward selection and sequential backwardselection can be used for feature selection However thesetechniques suffer from various problems such as stagnationin local optima and high computational cost [20] An efficientglobal search technique such as evolutionary computation(EC) is required to address feature selection problems moreeffectively PSO is one of the EC techniques based onswarm intelligence [20] Compared with other techniques

6 Journal of Sensors

Intensity 1High 24203

Low 8

N

0 04 0802(km)

(a)

Intensity 2High 19267

Low 8

N

0 04 0802(km)

(b)

Figure 5 LiDAR intensity derived by interpolating the intensity attributes of LiDAR point clouds (a) intensity of the study area used formethod development and (b) intensity of the test area

N

0 04 0802(km)

(a)

0 04 0802(km)

N

(b)

Figure 6 Aerial orthophotos (a) study area used for method development and (b) test area

such as genetic algorithm and simulated annealing PSO iscomputationally less costly and converges more rapidly [20]Therefore this technique was used in this study

PSO was proposed by Kennedy and Eberhart in 1995[21 22] This technique is motivated by social behavior suchas birds flocking and fish schooling In PSO knowledge isoptimized through social interaction in the population inwhich thinking is not only personal but also social Theuseful features from the available features of eCognitionsoftware were selected using PSO optimization implementedvia MATLAB PSO was used to minimize the classificationerror rateThe fitness function is given in (1) which is used tominimize the classification error rate obtained by the selected

features during the evolutionary training process and thenumber of selected features [20 23]

119865 = 119908 times TrainER

119904

+ (1 minus 119908) times TestER119904

120572 times FeaturesAll Features

+ (1 minus 120572) times ER119904ERall

(1)

where119908 is a constant number (the weight of the classificationerror rate obtained from the training data) and 119908 isin [0 1]TrainER

119904

is the classification error rate obtained from theselected feature subset and the training subset data TestER

119904

isthe classification error rate obtained from the selected featuresubset and the testing subset data 120572 is a constant number (theweight of the number of selected features) Features stands

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Journal of Sensors

LiDAR Aerial orthophoto

DSM DEMIntensity G BR

Scale = 10 Shape = 05 Compactness = 05

Layer stacking

Input image

Calculate class separability (Bhattacharyya distance)

Create image objectsIterate

Optimal scale Shape = 05 Compactness = 05

Optimize shape and compactness parameters (Taguchi method)

Optimal scale Optimal shape Optimal compactness

Feature selection (particle swarm)

Develop fuzzy rules

Classify and extract urban roads

Create image objects

nDSM

i = 1 to 15

Figure 2 Overall steps for two-stage optimization and fuzzy object-based analysis for urban road extraction

The point clouds were then filtered to ground andnonground points using multiscale curvature classification(MCC) [19] MCC is an iterative multiscale algorithm forclassifying LiDAR returns into ground and nonground [19]The MCC algorithm was developed at the Moscow ForestrySciences Laboratory of theUnited States Forest Service RockyMountain Research Station It integrates curvature filteringwith a scale component and variable curvature toleranceDuring this stage a surface is interpolated at different res-olutions using the thin-plate spline method [19] and pointsare classified based on a progressive curvature thresholdparameter which increases as resolution coarsens to com-pensate for the slope effect because data are generalizedThe same interpolation technique was then used to generatethe digital elevation models (DEMs) for the study and testareas (Figure 4) Afterwards the DSM and DEM rasters

were used to produce the nDSM raster by subtracting DEMfrom DSM In addition the intensity raster was generatedby interpolating the ground point clouds with the relatedintensity attribute information of the point clouds (Figure 5)

Meanwhile the aerial orthophoto was first correctedgeometrically to fit exactly into the LiDAR point clouds usingthe intensity raster (Figure 6) Finally the LiDAR-derivedDEM andDSM intensity nDSM and aerial orthophoto wereused as a subset based on the boundary of the study areato prepare the final input data for method development andurban road extraction

22 Feature Subset Selection via Particle Swarm Optimization(PSO) In object-based or knowledgemodeling a large num-ber of features are present in the data sets however not all of

Journal of Sensors 5

0 04 0802(km)

DSM 1(meters)

High 9164

Low 3882

N

(a)

0 04 0802(km)

DSM 2(meters)

High 7787

Low 4794

N

(b)

Figure 3 DSMs derived from airborne LiDAR data (a) DSM of the study area used for method development and (b) DSM of test area

High 470157

Low 3918

0 04 0802(km)

N

DEM 1(meters)

(a)

High 6254

Low 4794

0 04 0802(km)

N

DEM 2(meters)

(b)

Figure 4 DEMs derived from LiDAR point clouds (a) DEM of the study area used for method development and (b) DEM of the test area

these features are useful for classifying and extracting infor-mation Therefore the first step in object-based analysis orknowledgemodeling is selecting features from those availablein the data sets Useful features from the data sets for a par-ticular object extraction process can be selected using opti-mization techniques Such techniques aim to select a smallnumber of relevant features to achieve similar or better clas-sification results than using all the available featuresThe finalfeatures to be selected should be relevant Including irrelevantand redundant features in the analysis may produce poorclassification results because of the course of dimensionality

Considering the aforementioned reason only relevant fea-tures should be selected for accurate modeling

Several optimization methods such as greedy search-based sequential forward selection and sequential backwardselection can be used for feature selection However thesetechniques suffer from various problems such as stagnationin local optima and high computational cost [20] An efficientglobal search technique such as evolutionary computation(EC) is required to address feature selection problems moreeffectively PSO is one of the EC techniques based onswarm intelligence [20] Compared with other techniques

6 Journal of Sensors

Intensity 1High 24203

Low 8

N

0 04 0802(km)

(a)

Intensity 2High 19267

Low 8

N

0 04 0802(km)

(b)

Figure 5 LiDAR intensity derived by interpolating the intensity attributes of LiDAR point clouds (a) intensity of the study area used formethod development and (b) intensity of the test area

N

0 04 0802(km)

(a)

0 04 0802(km)

N

(b)

Figure 6 Aerial orthophotos (a) study area used for method development and (b) test area

such as genetic algorithm and simulated annealing PSO iscomputationally less costly and converges more rapidly [20]Therefore this technique was used in this study

PSO was proposed by Kennedy and Eberhart in 1995[21 22] This technique is motivated by social behavior suchas birds flocking and fish schooling In PSO knowledge isoptimized through social interaction in the population inwhich thinking is not only personal but also social Theuseful features from the available features of eCognitionsoftware were selected using PSO optimization implementedvia MATLAB PSO was used to minimize the classificationerror rateThe fitness function is given in (1) which is used tominimize the classification error rate obtained by the selected

features during the evolutionary training process and thenumber of selected features [20 23]

119865 = 119908 times TrainER

119904

+ (1 minus 119908) times TestER119904

120572 times FeaturesAll Features

+ (1 minus 120572) times ER119904ERall

(1)

where119908 is a constant number (the weight of the classificationerror rate obtained from the training data) and 119908 isin [0 1]TrainER

119904

is the classification error rate obtained from theselected feature subset and the training subset data TestER

119904

isthe classification error rate obtained from the selected featuresubset and the testing subset data 120572 is a constant number (theweight of the number of selected features) Features stands

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 5

0 04 0802(km)

DSM 1(meters)

High 9164

Low 3882

N

(a)

0 04 0802(km)

DSM 2(meters)

High 7787

Low 4794

N

(b)

Figure 3 DSMs derived from airborne LiDAR data (a) DSM of the study area used for method development and (b) DSM of test area

High 470157

Low 3918

0 04 0802(km)

N

DEM 1(meters)

(a)

High 6254

Low 4794

0 04 0802(km)

N

DEM 2(meters)

(b)

Figure 4 DEMs derived from LiDAR point clouds (a) DEM of the study area used for method development and (b) DEM of the test area

these features are useful for classifying and extracting infor-mation Therefore the first step in object-based analysis orknowledgemodeling is selecting features from those availablein the data sets Useful features from the data sets for a par-ticular object extraction process can be selected using opti-mization techniques Such techniques aim to select a smallnumber of relevant features to achieve similar or better clas-sification results than using all the available featuresThe finalfeatures to be selected should be relevant Including irrelevantand redundant features in the analysis may produce poorclassification results because of the course of dimensionality

Considering the aforementioned reason only relevant fea-tures should be selected for accurate modeling

Several optimization methods such as greedy search-based sequential forward selection and sequential backwardselection can be used for feature selection However thesetechniques suffer from various problems such as stagnationin local optima and high computational cost [20] An efficientglobal search technique such as evolutionary computation(EC) is required to address feature selection problems moreeffectively PSO is one of the EC techniques based onswarm intelligence [20] Compared with other techniques

6 Journal of Sensors

Intensity 1High 24203

Low 8

N

0 04 0802(km)

(a)

Intensity 2High 19267

Low 8

N

0 04 0802(km)

(b)

Figure 5 LiDAR intensity derived by interpolating the intensity attributes of LiDAR point clouds (a) intensity of the study area used formethod development and (b) intensity of the test area

N

0 04 0802(km)

(a)

0 04 0802(km)

N

(b)

Figure 6 Aerial orthophotos (a) study area used for method development and (b) test area

such as genetic algorithm and simulated annealing PSO iscomputationally less costly and converges more rapidly [20]Therefore this technique was used in this study

PSO was proposed by Kennedy and Eberhart in 1995[21 22] This technique is motivated by social behavior suchas birds flocking and fish schooling In PSO knowledge isoptimized through social interaction in the population inwhich thinking is not only personal but also social Theuseful features from the available features of eCognitionsoftware were selected using PSO optimization implementedvia MATLAB PSO was used to minimize the classificationerror rateThe fitness function is given in (1) which is used tominimize the classification error rate obtained by the selected

features during the evolutionary training process and thenumber of selected features [20 23]

119865 = 119908 times TrainER

119904

+ (1 minus 119908) times TestER119904

120572 times FeaturesAll Features

+ (1 minus 120572) times ER119904ERall

(1)

where119908 is a constant number (the weight of the classificationerror rate obtained from the training data) and 119908 isin [0 1]TrainER

119904

is the classification error rate obtained from theselected feature subset and the training subset data TestER

119904

isthe classification error rate obtained from the selected featuresubset and the testing subset data 120572 is a constant number (theweight of the number of selected features) Features stands

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Journal of Sensors

Intensity 1High 24203

Low 8

N

0 04 0802(km)

(a)

Intensity 2High 19267

Low 8

N

0 04 0802(km)

(b)

Figure 5 LiDAR intensity derived by interpolating the intensity attributes of LiDAR point clouds (a) intensity of the study area used formethod development and (b) intensity of the test area

N

0 04 0802(km)

(a)

0 04 0802(km)

N

(b)

Figure 6 Aerial orthophotos (a) study area used for method development and (b) test area

such as genetic algorithm and simulated annealing PSO iscomputationally less costly and converges more rapidly [20]Therefore this technique was used in this study

PSO was proposed by Kennedy and Eberhart in 1995[21 22] This technique is motivated by social behavior suchas birds flocking and fish schooling In PSO knowledge isoptimized through social interaction in the population inwhich thinking is not only personal but also social Theuseful features from the available features of eCognitionsoftware were selected using PSO optimization implementedvia MATLAB PSO was used to minimize the classificationerror rateThe fitness function is given in (1) which is used tominimize the classification error rate obtained by the selected

features during the evolutionary training process and thenumber of selected features [20 23]

119865 = 119908 times TrainER

119904

+ (1 minus 119908) times TestER119904

120572 times FeaturesAll Features

+ (1 minus 120572) times ER119904ERall

(1)

where119908 is a constant number (the weight of the classificationerror rate obtained from the training data) and 119908 isin [0 1]TrainER

119904

is the classification error rate obtained from theselected feature subset and the training subset data TestER

119904

isthe classification error rate obtained from the selected featuresubset and the testing subset data 120572 is a constant number (theweight of the number of selected features) Features stands

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 7

Table 1 Selected features via PSO used for classification

Property Quantifiableattribute Descriptionmathematical formulation

Spectralattributes

Height nDSM nDSM = DSM minus DEM

Color Spectral bands ofaerial orthophoto Digital number of blue green and red bands

Spatialattributes

Elongated shapes Elongation index119868 = 2 sdot 120587 sdot 119860

1198752 where 119860 and 119875 are the area and the perimeter ofimage object

Smooth objects Shape indexSI = 119890

4radic119860

where SI is the shape index 119890 is the border lengthand 119860 is the area of the image object

Distribution of pixelsin space Density

Calculated by the area of the object divided by theapproximated average of the major and minor axesof the ellipse fitted to the object

Rectangular shape Rectangular fit

Describes the difference between a master rectangleand the considered object using the same measure ofarea width and length for the master rectangle (1 =complete fitting object 0 = no fit)

Texturalattributes Texture GLCM-

homogeneity HOM = 119899sum119894=0

119899sum119894=0

11 + (119894 minus 119895)2 sdot ℎ119888 (119894 119895)

for the number of selected features All Features representsthe total number of features available for classification ER119904is the classification error rate obtained from the selected fea-tures and ERall is the classification error rate obtained fromall the available features The error rate of the classificationresults can be calculated using (2) [23 24] as follows

Error Rate (ER) = FP + FNTP + TN + FP + FN

(2)

where TP TN FP and FN stand for true positives truenegatives false positives and false negatives respectively

23 Knowledge Modeling and Creation of Fuzzy Rule SetsKnowledge modeling is the process used to model theknowledge used by an expert to identify land coverlanduse objects in aerial photos or satellite images In thisstudy spectral spatial and texture properties were used todescribe land coverland use objects For each property aset of quantifiable attributes was determined using severalmathematical formulations The useful features (quantifiableattributes) selected by the PSO technique were used to createfuzzy rule sets The details of these attributes are provided inTable 1

Pure data-driven rules developed for object-based anal-ysis are generally difficult to transfer [24] Meanwhile fuzzylogic allows the formulation of knowledge given in a naturallanguage with vague and imprecise expressions [14] Fuzzylogic also avoids the use of hard boundaries between theconsidered classes and thus allows easy and efficient selectionof thresholds [14] That is classification accuracy is less sen-sitive to thresholds when fuzzy logic is used [11] In additionfuzzy logic is considered an appropriate technique for urbanfeature extraction because it models uncertainties among the

considered classes these uncertainties are typically foundin urban areas [12] The aforementioned reasons justify theuse of fuzzy logic in object-based analysis for urban roadextraction and to generate thematic maps for the study areas

In general fuzzy rules can be developed in two waysFirst fuzzy rules can be developed using training data andclustering algorithms Second they can be developed usingexpert knowledge on object identification Both approacheshave advantages and disadvantages For example data-drivenfuzzy rules are highly sensitive to the selection of trainingdata whereas the expert knowledge-based approach is rela-tively difficult to perform and knowledge may be difficult toconvert into quantifiable attributes In this study both meth-ods were integrated for fuzzy rule-based creation becauseruleswere initially generated based on training data using twoalgorithms (decision tree and fuzzy 119888-mean) and then therules were assessed andmodified based on expert knowledgeThis manner of assessment and modification ensures that therules are logical and can be potentially transferred to otherareas Such approach also ensures that quantifiable attributesare correctly assigned for each class

First a small number of training data (10 objects perclass) was generated for the considered classes includingurban roads A small number of training data were selected toavoid including incorrect samples and to ensure that the rulessatisfactorily defined the objects Afterwards a decision treealgorithm was applied on the training data and a set of crisprules was generated In addition the fuzzy 119888-mean algorithmwas applied in the same training data set to generate fuzzyrules that could recognize the considered classes Utilizingexpert knowledge (ie that of the authors of this article inthis case) the two sets of crisp and fuzzy rules were analyzedFinally by understanding the importance of each quantifiable

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Journal of Sensors

Table 2 Fuzzy rule sets developed by analyzing the decision tree and the fuzzy 119888-mean clustering algorithms

Land cover class Fuzzy rules

Building

IF nDSM is High AND 119877 is High THEN Object is BuildingIF nDSM is High AND Density is High AND 119866 is Low AND Rectangular Fit is High THEN Object is BuildingIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is High THEN Object is BuildingIF Intensity is High AND nDSM is High AND Rectangular Fit is High THEN Object is Building

Road

IF nDSM is Low AND 119877 is High AND Density is Low THEN Object is RoadIF nDSM is Low AND 119877 is Low AND GLCM-Hom is Low THEN Object is RoadIF Intensity is High AND nDSM is Low AND 119877 is Low AND Elongation is High THEN Object is RoadIF nDSM is Low AND 119877 is High AND Density is Low THEN Object is Road

Urban tree

IF nDSM is High AND 119877 is Low THEN Object is Urban TreeIF nDSM is High ANDDensity is High AND 119866 is Low AND Rectangular Fit is Low THEN Object is Urban TreeIF nDSM is High AND GLCM-Hom is High AND Rectangular Fit is Low THEN Object is Urban TreeIF Intensity is High AND nDSM is High AND Rectangular Fit is Low THEN Object is Urban Tree

Grass land IF Intensity is High AND nDSM is Low AND 119877 is High THEN Object is Grass LandIF Intensity is High AND nDSM is Low AND 119861 is Low THEN Object is Grass Land

Bare landIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare LandIF nDSM is Low AND GLCM-Hom is High AND Shape Index is Low THEN Object is Bare LandIF nDSM is Low AND 119877 is High AND Density is High THEN Object is Bare Land

Water IF nDSM is Low AND 119877 is Low AND GLCM-Hom is High AND 119861 is High THEN Object is WaterIF Intensity is High AND nDSM is Low AND 119877 is Low AND 119861 is Low THEN Object is Water

Shadow IF Intensity is Low AND nDSM is Low AND 119877 is High THEN Object is ShadowIF nDSM is Low AND GLCM-Hom is High THEN Object is Shadow

attribute for each class and by determining the number ofattributes required to recognize a particular class the finalversion of the fuzzy rules was formulated (Table 2)

24 Selection of the Optimum Range Values for the ScaleParameter In OBIA the accuracy levels of feature extractionand object classification are generally controlled through seg-mentation quality [8] High-quality segmentation (perfectlycoinciding with the reference objects) can be achieved byselecting the optimum values for its user-defined parameters(ie scale shape and compactness) However the value ofthe scale parameter directly affects the accuracy of objectclassification Therefore before optimizing the scale param-eter for segmentation selecting a range of scale values thatcan achieve high classification accuracy for image objects isnecessary The optimum range values of the scale parametercan be selected by choosing the values that maximize classseparability distance Figure 7 illustrates an experiment thatanalyzes the effect of the scale parameter on classificationaccuracy Scale parameters ranging from 10 to 150 wereanalyzed for the class separability distance which was mea-sured using the Bhattacharyya distance algorithm Figure 7shows three peaks in the graph which indicate that the scaleparameters that can achieve the best classification results onthe data set and the training samples are used

Image object hierarchies that incorporate spatial infor-mation describing each of these situations separately fre-quently provide better classification results than a spatialrepresentation constrained within a single operative scaleAlthough multiscale representations exhibit advantages oversingle-scale representations a corresponding increase indata processing and storage demands occurs Therefore the

000000100020003000400050006000700080

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

Bhat

tach

aryy

a dist

ance

Scale value

Figure 7 Class separability distance estimated using the Bhat-tacharyya distance algorithm used to optimize the scale parameters

minimum number and configuration of object scales (hier-archical levels) that maximize the potential final land coverclassification accuracy when they are combined should beidentified while minimizing excess data processing [25] Thisanalysis also determines the best number of segmentationlevels that can be used to classify land coverland use classesin the data set Therefore three segmentation levels withdifferent scale values were used In this analysis the scaleparameters for the three segmentation levels are determinedwithin a range of values The exact scale parameter foreach segmentation level should be determined by applyinganother optimization technique The selection of the exactscale parameters for the three segmentation levels is discussedin the next section

25 Selection of the Optimum Values for the Scale Shapeand Compactness Parameters After the range values for thescale parameter were selected the exact values for the scaleshape and compactness parameters should be optimized As

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 9

Table 3 Factors and their levels used to optimize image segmentation

Factor Level 1 Level 2 Level 3 Level 4Segmentation level 1

Scale 30 35 40 45Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 2Scale 105 110 115 120Shape 01 03 05 09Compactness 01 03 05 09

Segmentation level 3Scale 135 140 145 mdashShape 01 05 09 mdashCompactness 01 05 09 mdash

Table 4 1198719 and 11987116 orthogonal arrays and the POF used to optimize the shape and compactness parameters

Run Level 1 Level 2 Level 3MI Variance POF MI Variance POF MI Variance POF

(1) 0396 0930 1327 0599 0964 1563 1464 1011 2474(2) 0233 1000 1233 0795 0971 1765 0373 0885 1258(3) 0438 0865 1303 0647 0963 1610 0320 0000 0320(4) 0141 0411 0552 0523 0000 0523 1000 0947 1947(5) 0371 0962 1333 0859 0956 1815 0000 0904 0904(6) 0000 0967 0967 0704 0986 1690 0523 0616 1138(7) 1000 0697 1697 0692 0901 1593 0326 0945 1271(8) 0174 0000 0174 0698 0049 0746 0969 1000 1969(9) 0013 0954 0967 0668 1000 1668 0404 0332 0736(10) 0565 0801 1366 0113 0958 1072(11) 0578 0758 1336 0783 0998 1782(12) 0867 0274 1141 0738 0415 1153(13) 0329 0952 1280 0554 0978 1532(14) 0035 0891 0927 0000 0979 0979(15) 0307 0798 1105 0434 0948 1381(16) 0060 0207 0268 1000 0480 1480

mentioned earlier selecting these parameters using the trial-and-error method is time-consuming and labor intensiveConsequently Taguchi optimization was used in this study

The first step in implementing the Taguchi method isto design an orthogonal array to test a set of choices [26]An orthogonal array is related to the situation in which thecolumns for independent variables are ldquoorthogonalrdquo to oneanother These tables provide an easy and consistent designfor experiments Selecting the orthogonal array depends onthe number of levels and the number of parameters [25](Table 3) This technique is frequently used in cases inwhich various levels of parameters are available For exampleseveral combinations which require a significant amount oftime for testing can be used However by using the Taguchitechnique choices will be considerably reduced to a smallnumber of experiments

Taguchi optimization can be implemented by conductingthe following steps First the objective of the process should

be determined This step involves defining the possiblevalues of a particular parameter for the process Secondthe parameters that can influence the process are definedThese parameters exhibit variable values that can affectaccuracy or performance thus the level should be definedby the user depending on the effect of a parameter onthe process In this study the range values of the scaleparameter were determined using the Bhattacharyya distancealgorithm as mentioned earlier whereas the levels of theshape and compactness parameters were determined throughtheir minimum (min = 0) and maximum (max = 1)values Subsequently the experiments were performed afterthe appropriate orthogonal array was designed The effect ofeach parameter on performance was then measured

In the next step the plateau objective function (POF)was measured for each test to examine segmentation qualityby using each testing combination and to determine theoptimum segmentation parameters (Table 4) POF is the

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

10 Journal of Sensors

0 1 2 3Bare land

Grass landBuilding

Urban treeRoad

ShadowWater

Level of segmentation

Nam

e of t

he la

nd co

ver c

lass

Figure 8 Land cover classes of the study area used for methoddevelopment (with their assigned segmentation levels)

combination of Moranrsquos index and a variance indicatorThen signal-to-noise ratio (SNR) was measured to assess theresults of the segmentation tests A high SNR indicates highsegmentation accuracy Equation (3) is used to calculate SNRas follows

SNR = minus10 log10 (1119899 sum11199102119894 ) (3)

where 119899 is the number of repetitions under similar testsituations (119899 = 1 in this study) and 119910 denotes the POF valuesobtained from each segmentation test The SNR table wasthen completed and the optimum conditionwas determined

Selecting image segmentation parameters is complicatedbecause multiple spatial scales can convey important infor-mation on physical characteristics which define a specificclass or feature [27] A single image segmentation config-uration may not be optimal to distinguish all the featuresin a typical image Consequently optimization techniqueswere used in this study to select the appropriate combinationof user-defined parameters for segmentation The optimiza-tion techniques used also determined the best number ofsegmentation levels required to classify image objects andto extract the considered information accurately Once therequired number of segmentation levels is identified withtheir associated user-defined parameters each segmentationlevel should be assigned to its corresponding land coverlanduse classes In this research the visual interpretation of thesegmented images was used to determine the appropriatesegmentation level for each class According to the conductedanalysis shadow and urban tree classes were assigned tosegmentation level 1 urban road and building classes wereassigned to segmentation level 2 water body grasslandand bare land classes were assigned to segmentation level 3(Figure 8)

3 Results and Discussion

In this section the results obtained from the study on a two-stage optimization for fuzzy object-based analysis using air-borne LiDAR and orthophotos for urban road extraction arepresented and discussed The selection of the segmentationparameters is first described Then the results obtained fromthe classification of image objects based on the fuzzy rulesproposed in this study are provided and explained

minus10minus8minus6minus4minus2

0246

Sign

al-to

-noi

se ra

tio

Level 1Level 2

Level 3Level 4

Scal

e

Segmentation

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

Scal

e

Com

pact

ness

Shap

e

level 1Segmentation

level 2Segmentation

level 3

Figure 9 SNRs used to select the optimum scale shape andcompactness parameters for the fuzzy object-based analysis TheSNR measures how the response varies relative to the target valueunder different noise conditions The negative values are due tothe logarithmic transformation used in SNR formula The negativevalues indicate nonfavorable parameter values for the given task

31 Results of the Selection of the Segmentation ParametersUser-defined MRS parameters were optimized using theBhattacharyya distance and Taguchi approaches Figure 9shows the SNRs for each segmentation parameter and levelFor segmentation level 1 the highest SNR value of the scaleparameter suggests that the best value for the scale parameterin segmentation level 1 is 40 (scale level 3) Similarly Figure 9indicates that the shape value of 05 (shape level 3) is thebest for segmentation level 1 Meanwhile the best value forcompactness in segmentation level 1 is 03 (compactness level2)

For segmentation level 2 the analysis shows that thebest scale value is 115 (scale level 1) Meanwhile the bestvalues for shape and compactness are both 01 (shape level 1and compactness level 1) Furthermore the analysis indicatesthat the appropriate scale value for segmentation level 3 is05 (scale level 2) whereas the best values for shape andcompactness are both 01 (shape level 1 and compactness level1)

In the MRS algorithm the scale parameter defines thesize and heterogeneity of an object The scale value of 40was selected using the two-stage optimization method forsegmentation level 1 which included two classes (ie urbantree and shadow) because this threshold combined withthe other two parameters was the best for extracting theconsidered land cover classes Higher scale values cause over-segmentation in the shadow class whereas lower scale valuesresult in undersegmentation in the urban tree class Mean-while the two-stage optimization applied in this researchsuggested a shape parameter value of 05 for segmentationlevel 1 The sum of the weight values of the shape and colorparameters is 1 High shape values are recommended forregular objects whereas small-scale values are suggested forobjects with obvious spectral characteristics In segmentation

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 11

level 1 urban trees have nearly similar shapes (ie compact)whereas shadows have different shapes because they areformed by urban trees and tall buildings This result led tothe selection of the intermediate shape value of 05 to balancebetween shape and color properties In addition the com-pactness parameter refers to the compactness degree in eachsegmented object when weight is high the internal objectis compact Meanwhile the smooth parameter (smooth =1minuscompactness) represents the smoothness of the segmentedobject boundary In this study a low compactness value wasselected for the urban tree and shadow classes Therefore thestudy area has noncompact shadows from tall trees

Furthermore a scale value of 110 for segmentation level 2was selected for the two-stage optimization technique appliedin this study (Table 3) This segmentation level is consideredsuitable for extracting building and road features dependingon the scale parameter as determined by the Bhattacharyyadistance and Taguchi approaches However low values forboth shape and compactness were selected (01) whichindicated uncertainties in road boundary detection becausebuildingswere compact and had smooth shapes Accordinglyhigh shape and compactness values were selected Optimiza-tion was applied to select suitable parameters for both classes(building and road) simultaneously and thus low valueswereselected because of the effect of noncompact and nonsmoothroads

Water bodies bare lands and grasslands were assignedto segmentation level 3 based on the visual interpretationof the aerial orthophoto They have relatively large areasand are mostly homogenous In terms of segmentationparameters the scale value of 140 was determined as the bestthreshold to separate the three classes and to describe theobjects Meanwhile the shape and compactness values of 01were suitable using the proposed optimization method Thelow shape and compactness values can be justified becausethe three classes have noncompact and rough shapes Thiscondition is attributed to the presence of urban trees withingrasslands and bare lands (Figure 6(a))

32 Segmentation Assessment The assessment of the seg-mentation results and the optimization of the segmentationparameters may improve classification results and final roadnetworks In addition if image objects are created accuratelythen classification rules can be more efficiently optimizedand transferred to other study areas Therefore although thesegmentation parameters were selected using optimizationtechniques the segmentation results should still be assessedIn this study the segmentation results were evaluated bycomparing the extracted objects with the reference objectscreated via manual digitization in a geographic informationsystem (GIS) The geometric quality of segmentation wasevaluated using two indices area ratio (119877119860) and positionerror (119864119901) Area ratio yields accurate information on thepercentage of the area extracted for each object which isdefined as follows [14]

119877119860 = 119860 (Objectref capObjectext)119860 (Objectref) (4)

where 119860 is the area

Position error is related to the mean distance between theextracted objects and their corresponding objects in referencedata This error is determined according to the characteristicpoints in the two dataThe characteristic points for buildingscorrespond to building corners whereas those for roads areselected along road axes Position error is determined by thefollowing equation [14]

119864119901 =sum119899119896=1 (radic(119883ext minus 119883ref)2 minus (119884ext minus 119884ref)2)

119899 (5)

where 119899 is the number of characteristic points (119883ext119884ext) cor-respond to the coordinates of an extracted point and (119883ref 119884ref ) are the coordinates of the corresponding characteristicpoint on the reference layer

33 Classification Assessment Classification results wereassessed in eCognition software using its accuracy assessmenttools Results are shown in Tables 5 and 6 The overallaccuracy of the produced thematic map was first calculatedThen producer and user accuracies were calculated foreach class Finally the kappa index of agreement (KIA) wascalculated for thewhole classifiedmap and for each class KIAis calculated using the following equation [28]

KIA = (119860 minus 119864)(119873 minus 119864) (6)

where119860 is the number of samples in the accurately predictedclass 119864 is the number that will be predicted accurately viarandom class selection and119873 is the total number of samplesin that class

34 Results of Image Classification Figure 12 shows the classi-fication results on the study area used formodel developmentThequality of segmentation used to generate the classificationmap is presented in Figures 10 and 11 The average area ratioof all the classes was 8080This finding indicates that mostof the extracted objects are accurately matched with the ref-erence objectsMeanwhile theminimum andmaximum arearatios were 62 (for grass) and 99 (for water) respectivelyThe low accuracy of grass was ascribed to some of the grassobjects beingmerged together after the imagewas segmentedThe highest accuracy was obtained for the water class becauseonly a few water objects were present and all of these wereaccurately extracted The spectral information of the aerialorthophoto enabled the accurate extraction of water and theseparation of water from surrounding nonwater objects (iebare lands and grasslands) In addition the minimum andmaximum position errors in the segmented objects were 102(water class) and 38 (building class) respectively

In evaluating the classification results an overall accuracyof 82 and a KIA of 079 were achieved when the resultswere compared with the reference objects created via manualdigitization in GIS The producer user and KIA accuraciesfor each class and the detailed confusionmatrix are presentedin Tables 5 and 6 The KIA of the building class was 082This value indicated that most of the buildings were extracted

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

12 Journal of Sensors

Table 5 Classification assessment (confusion matrix per class accuracies and total accuracies) on the study area used for methoddevelopment

User classsample Building Road Tree Grass Bare Water Shadow SumConfusion matrix

Building 65 0 0 0 7 0 0 72Road 2 59 1 8 10 0 1 81Tree 1 0 68 2 0 0 1 72Grass 0 0 1 23 5 1 1 31Bare 7 7 0 2 66 0 0 82Water 0 0 0 0 0 4 0 4Shadow 0 0 5 1 0 0 24 30Sum 76 66 75 36 88 5 27

AccuracyProducer 085 089 090 063 075 080 088User 090 072 094 074 080 1 080KIA per class 082 086 088 060 068 080 088

Overall accuracy ()82

Kappa index of agreement (KIA)079

Table 6 Classification assessment (confusion matrix per class accuracies and total accuracies) on the test site

User classsample Tree Road Building Grass Bare Shadow SumConfusion matrix

Tree 58 0 1 7 0 18 84Road 3 54 1 0 2 3 63Building 2 6 116 15 4 22 165Grass 2 1 1 53 0 2 59Bare 5 10 5 3 57 3 83Shadow 0 0 0 1 0 37 38Sum 70 71 124 79 63 85

AccuracyProducer 082 076 093 067 090 043User 069 085 070 089 068 097KIA per class 079 072 090 062 088 038

Overall accuracy ()76

Kappa index of agreement (KIA)070

accurately and separated from other classes However mis-classifications between buildings and vehicles on the roadstypically occur because cars are rectangular and elevatedobjects (approximately 12m) These characteristics whichare similar to most of the buildings in the study area leadto a decrease in the KIA of the building class To address thisproblem in the future spectral information should be con-sidered or more advanced rules should be developed usingcontextual and geometric information Shadow objects wereextracted at a KIA accuracy of 088 This high accuracy wasachieved because shadow objects were dark and differed from

other natural classes (ie water bare land and grassland)They were distinguished from the building class because ofthe difference in height and were differentiated from roadsbecause of their spatial characteristics (roads were elongatedwhereas shadows had lower values in this property) Shadowsmostly occur because of tall buildings and urban trees Waterbodies were extracted at a KIA accuracy of 080 Spectralinformation such as the red band of the aerial orthophoto aswell as intensity and height information from LiDAR data isthe most important feature used to differentiate water bodiesfrom other classes in the study area Accuracy was reduced to

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 13

50

60

70

80

90

100

Building Road Tree Grass Bare Water Shadow

Are

a rat

io (

)

Name of class

0

1

2

3

4

5

Building Road Tree Grass Bare Water Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

Figure 10 Geometric quality assessment on the study area used for method development

0

1

2

3

4

5

Building Road Tree Grass Bare Shadow

Posit

ion

erro

r (m

)

Name of class

MinMeanMax

MinMeanMax

50

60

70

80

90

100

Building Road Tree Grass Bare Shadow

Are

a rat

io (

)

Name of class

Figure 11 Geometric quality assessment on the test site

0 140 280 420 56070(m)

BuildingRoadUrban treeGrass land

Bare landWaterShadow

N

Figure 12 Classified map produced by the fuzzy object-basedapproach (study area used for model development)

080 because one of the grass objects was classified as watergiven the similar texture and spectral information betweenwater and grass objects The rules developed for the waterclass indicate that spatial information is not as importantas spectral information because water bodies have spatialcharacteristics similar to those of some grassland and bareland classes In addition the shape of water bodies is notunique and depends on the study areaTherefore developingrules that use only spectral texture and height informationis important

Urban trees are easily differentiated from other classesbecause of the difference in height they are distinguishedfrom buildings through intensity and rectangular fit prop-erties Urban trees were extracted at a KIA accuracy of088 Given that urban trees are typically located within barelands or grasslands they are differentiated from surroundingobjects with different characteristics Texture (homogeneity)information and intensity information are also useful indifferentiating urban trees from some of buildings withspecific roofing materials The results demonstrated that theproposed fuzzy rules enabled the extraction of bare lands andgrasslands at an average KIA accuracy of 064 After the inter-pretation of the produced thematic map and the statisticalinformation of the confusion matrix the low accuracy was

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

14 Journal of Sensors

N

(a)

0 100 200 300 40050(m)

Road

(b)

Figure 13 Urban roads extracted using the proposed fuzzy rules (study area used for method development)

ascribed to some of the objectsmerging during segmentationThis situation led to the loss of the characteristics andproperties used to define these classes in rule developmentMisclassifications mostly occurred between bare lands androads grasslands and roads and bare and grasslands becausesome of the small bare land and grassland objects exhibitedsimilar spatial characteristics with roads (ie the elongationproperty of roadswas the same as those of small grassland andbare land objects) The interpretation of the results indicatedthat additional spectral and contextual information would benecessary to improve the accuracy of feature extraction

35 Results of Urban Road Detection Figure 13 shows theurban roads extracted using the proposed fuzzy rules Heightspectral spatial and texture information were used forroad extraction The accuracy of road extraction using thedeveloped fuzzy rules was 076 (producer) 085 (user) and072 (KIA) Spatial information (ie elongation) and intensityfrom LiDAR data were the most interesting properties fortheir extraction Several omissions were attributed to theheterogeneity of the surface of roads and the shadowsprojected onto roads by trees buildings or vehiclesThemostsignificant problems of road extraction were the presence oftrees along the roads and the undetected boundaries of roadsEven spatial information is not useful in solving these prob-lems because these problems occur during segmentation Toaddress these problems more advanced rules based on con-textual and probability theories can be integrated To improvethe quality of the extracted roads further postprocessingtechniques such as filtering removing nonroad featuresand smoothing the extracted polygons are necessary Thequality of the final road features is critical for the applicationsmentioned in Introduction Roads can also be converted intopolylines that represent their centerlines However the aimof this study is to extract roads in their complete geometry toincrease their use for a wide range of geospatial applications

Urban treeRoadBuilding

Grass landBare landShadow

0 130 260 390 52065(m)

N

Figure 14 Test site map produced by the fuzzy object-basedapproach

36 Evaluation of the Transferability of the Fuzzy Rule SetsA transferability test is essential to ensure that the proposedrules will extract the considered objects without any changesin the rule base Figure 14 shows the classified map usingthe same segmentation parameters and the proposed fuzzyrule sets The test site used to evaluate the efficiency of theproposed rules included the same classes except for waterbodies Therefore the test site was suitable for evaluating

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 15

Road0 130 260 390 52065

(m)

N

Figure 15 Roads extracted using the proposed fuzzy rules (test area)

the rules developed in this research The test site is a subsetof the data set used in this research but has a differenturban environment Assuming that most of the airborneLiDAR data have similar characteristics (ie similar featuresof nDSM intensity aerial orthophoto and spatial resolution)the proposed rules are most probably transferable to anyother study areas However slight changes in segmentationparameters may be required because these rules depend onthe scale of the study area and the resolution of the data sets

Overall accuracy was decreased by approximately 6when the rules were applied on the test site A KIA of 070was achieved on the test site using the same rules withoutany changes The first interpretation of this decrease inaccuracy suggested that some features in the new test site haddifferent properties which were disregarded when the ruleswere developed Therefore knowledge modeling should becarefully integrated into data-driven rules as applied in thisresearch Given the extensive variety of properties for naturaland man-made features these issues remain a challenge Inparticular the KIA accuracies of the building and grasslandclasses increased from 082 to 090 and from 068 to 088respectively because only one type of building structures wasfound in the test site By contrast the accuracies of the urbantree and shadow classes were decreased The KIA accuracyof the urban tree type decreased from 088 to 079 whereasthat of the shadow class decreased from 088 to 038 Thedecrease in the urban tree type was normal because the rulewas created based on data integration and expert knowledgeBy contrast the decrease in shadow accuracy was abnormal(KIA decreased by approximately 050) mainly because of theuncertainties in the testing samples The low contrast of theshadow features in the orthophoto made interpreting thosefeatures difficult and thus the creation of testing samples wasaffected

Figure 15 shows the extracted urban roads from the testsite using the proposed fuzzy rules The KIA accuracy of theextracted urban road was 072 which decreased to nearly016 Misclassificationsmostly occurred between urban roadsand buildings and between urban roads and bare lands Inthe test site many cars were located on the roads whichwere misclassified as buildings because of the similar spatialproperties of cars and buildings In addition the shadowsprojected from tall trees and high cars also contributedin decreasing road extraction accuracy To improve theresult of urban road extraction removing vehicles before theclassification process using several filtering techniques is sug-gested Further improvement can be achieved by developingadvanced rules for complex urban environments

4 Conclusion

In this study a two-stage optimization method for segmen-tation parameter selection was presented A fuzzy object-based analysis technique for urban road extraction from air-borne LiDAR data was also proposed Suitable segmentationparameters were selected using the two-stage optimizationmethod In the first stage the scale parameters were opti-mized via PSO to improve classification In the second stagesuitable shape and compactness parameters were selected toimprove object creation using the Bhattacharyya distanceThe proposed fuzzy rules are easy to implement and aretransferable The fuzzy rules were formulated based ondata samples and expert knowledge using the decision treeand fuzzy 119888-mean algorithms The results showed that theproposed method could extract urban roads from LiDARdata with a KIA accuracy of 072 on the test site

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

16 Journal of Sensors

Feature selection using PSO contributed to improvingknowledge modeling by selecting relevant and useful fea-tures and excluding nontransferable and problematic fea-tures Meanwhile the two optimization techniques namelyBhattacharyya distance and Taguchi enabled the selectionof the best combination of segmentation parameters toimprove object creation and classification results In additionintegrating fuzzy logic into object-based rules contributed tomanaging the complexity of classification in urban environ-ments by involving uncertainties in the boundaries of theused classes

Themain challenge in urban road extractionwas the pres-ence of urban trees along the roads which created difficultiesin accurately extracting road features with complete geom-etry The presence of vehicles on urban roads also createdproblems in urban road extraction In addition vehicles onthe roads formed shadows on the road surface which alsoaffected the quality of final road extraction Furthermore avariety of shadow shapes that could be present in the scenemade the extraction of some classes challenging To solvethese limitations and problems further research should beconducted by focusing on the factors that created a widerange and variety of problems

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This research is supported by research project withUPMVoteno 6300193

References

[1] M Pregnolato A Ford and R Dawson ldquoClimate and changesimulating flooding impacts on urban transport networkrdquo inProceedings of the EGUGeneral Assembly Vienna Austria April2015

[2] S Gumus H H Acar and D Toksoy ldquoFunctional forest roadnetwork planning by consideration of environmental impactassessment for wood harvestingrdquo Environmental Monitoringand Assessment vol 142 no 1ndash3 pp 109ndash116 2008

[3] N Dong H Huang and L Zheng ldquoSupport vector machine incrash prediction at the level of traffic analysis zones assessingthe spatial proximity effectsrdquo Accident Analysis and Preventionvol 82 pp 192ndash198 2015

[4] Q Zeng and H Huang ldquoBayesian spatial joint modeling oftraffic crashes on an urban road networkrdquoAccident Analysis andPrevention vol 67 pp 105ndash112 2014

[5] L Gong Y Zhang Z Li and Q Bao ldquoAutomated roadextraction from LiDAR data based on intensity and aerialphotordquo in Proceedings of the 2010 3rd International Congress onImage and Signal Processing (CISP rsquo10) October 2010

[6] T Blaschke ldquoObject based image analysis for remote sensingrdquoISPRS Journal of Photogrammetry and Remote Sensing vol 65no 1 pp 2ndash16 2010

[7] G Yan J-F Mas B H P Maathuis Z Xiangmin and P MVan Dijk ldquoComparison of pixelminusbased and objectminusoriented

image classification approachesmdasha case study in a coal fire areaWuda Inner Mongolia Chinardquo International Journal of RemoteSensing vol 27 no 18 pp 4039ndash4055 2006

[8] M Baatz and A Schape ldquoMultiresolution segmentation anoptimization approach for high quality multi-scale image seg-mentationrdquo in Angewandte Geographische Informationsverar-beitung XII pp 12ndash23 Wichmann 2000

[9] T Blaschke and J Strobl ldquoWhatrsquos wrong with pixels Somerecent developments interfacing remote sensing and GISrdquo Geo-Informations-Systeme vol 14 no 6 pp 12ndash17 2001

[10] X Li X ChengW Chen G Chen and S Liu ldquoIdentification offorested landslides using lidar data object-based image analysisand machine learning algorithmsrdquo Remote Sensing vol 7 no 8pp 9705ndash9726 2015

[11] S Jabari and Y Zhang ldquoVery high resolution satellite imageclassification using fuzzy rule-based systemsrdquo Algorithms vol6 no 4 pp 762ndash781 2013

[12] P A Longley M F Goodchild D J Maguire and D W RhindGeographic Information System and Science JohnWiley amp SonsLtd London UK 2001

[13] A Grote C Heipke and F Rottensteiner ldquoRoad networkextraction in suburban areasrdquoThePhotogrammetric Record vol27 no 137 pp 8ndash28 2012

[14] I Sebari and D-C He ldquoAutomatic fuzzy object-based analysisof VHSR images for urban objects extractionrdquo ISPRS Journal ofPhotogrammetry and Remote Sensing vol 79 pp 171ndash184 2013

[15] M Li A Stein W Bijker and Q Zhan ldquoRegion-based urbanroad extraction from VHR satellite images using binary par-tition treerdquo International Journal of Applied Earth Observationand Geoinformation vol 44 pp 217ndash225 2016

[16] A Hamedianfar and H Z M Shafri ldquoDevelopment of fuzzyrule-based parameters for urban object-oriented classificationusing very high resolution imageryrdquoGeocarto International vol29 no 3 pp 268ndash292 2014

[17] E Choi and C Lee ldquoFeature extraction based on the Bhat-tacharyya distancerdquo Pattern Recognition vol 36 no 8 pp 1703ndash1709 2003

[18] V Moosavi A Talebi and B Shirmohammadi ldquoProducing alandslide inventory map using pixel-based and object-orientedapproaches optimized by Taguchi methodrdquo Geomorphologyvol 204 pp 646ndash656 2014

[19] J S Evans and A T Hudak ldquoA multiscale curvature algorithmfor classifying discrete return LiDAR in forested environmentsrdquoIEEE Transactions on Geoscience and Remote Sensing vol 45no 4 pp 1029ndash1038 2007

[20] B Xue M Zhang and W N Browne ldquoParticle swarm opti-mization for feature selection in classification a multi-objectiveapproachrdquo IEEE Transactions on Cybernetics vol 43 no 6 pp1656ndash1671 2013

[21] J Kennedy ldquoParticle swarm optimizationrdquo in Encyclopedia ofMachine Learning pp 760ndash766 Springer 2010

[22] J Kennedy and R C Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks (ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE ServiceCenter Perth Australia December 1995

[23] M I Sameen and B Pradhan ldquoA novel built-up spec-tral index developed by using multiobjective particle-swarm-optimization techniquerdquo IOP Conference Series Earth andEnvironmental Science vol 37 no 1 Article ID 012006 2016

[24] R Sheikhpour M A Sarram and R Sheikhpour ldquoParticleswarm optimization for bandwidth determination and feature

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 17

selection of kernel density estimation based classifiers in diag-nosis of breast cancerrdquo Applied Soft Computing Journal vol 40pp 113ndash131 2016

[25] A Smith ldquoImage segmentation scale parameter optimizationand land cover classification using the random forest algo-rithmrdquo Journal of Spatial Science vol 55 no 1 pp 69ndash79 2010

[26] B Pradhan M N Jebur H Z M Shafri and M S TehranyldquoData fusion technique using wavelet transform and taguchimethods for automatic landslide detection from airborne laserscanning data and quickbird satellite imageryrdquo IEEE Transac-tions on Geoscience and Remote Sensing vol 54 no 3 pp 1610ndash1622 2016

[27] D E Jelinski and JWu ldquoThemodifiable areal unit problem andimplications for landscape ecologyrdquo Landscape Ecology vol 11no 3 pp 129ndash140 1996

[28] M J Aitkenhead and R Dyer ldquoImproving land-cover classifica-tion using recognition threshold neural networksrdquo Photogram-metric Engineering and Remote Sensing vol 73 no 4 pp 413ndash421 2007

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpswwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of