Upload
sk
View
214
Download
0
Embed Size (px)
Citation preview
The Egyptian Journal of Remote Sensing and Space Sciences (2013) 16, 163–169
National Authority for Remote Sensing and Space Sciences
The Egyptian Journal of Remote Sensing and Space
Sciences
www.elsevier.com/locate/ejrswww.sciencedirect.com
RESEARCH PAPER
An evolutionary computing frame work toward
object extraction from satellite images
P.V. Arun *, S.K. Katiyar
Dept. of Computer Science, National Institute of Technology, Bhopal, India
Received 9 April 2013; revised 20 May 2013; accepted 27 May 2013
Available online 17 July 2013
*
E
gm
Pe
Se
11
ht
KEYWORDS
Cellular neural network;
Feature detection;
Interpolation
Tel.: +91 4828249999.
-mail addresses: arunpv2601@
ail.com (S.K. Katiyar).
er review under responsibili
nsing and Space Sciences.
Production an
10-9823 � 2013 Production
tp://dx.doi.org/10.1016/j.ejrs.
gmail.co
ty of Na
d hostin
and hosti
2013.05.0
Abstract Image interpretation domains have witnessed the application of many intelligent method-
ologies over the past decade; however the effective use of evolutionary computing techniques for fea-
ture detection has been less explored. In this paper, we critically analyze the possibility of using cellular
neural network for accurate feature detection. Contextual knowledge has been effectively represented
by incorporating spectral and spatial aspects using adaptive kernel strategy. Developed methodology
has been compared with traditional approaches in an object based context and investigations revealed
that considerable success has been achieved with the procedure. Intelligent interpretation, automatic
interpolation, and effective contextual representations are the features of the system.� 2013 Production and hosting by Elsevier B.V. on behalf of National Authority for Remote Sensing and
Space Sciences.
1. Introduction
Earth observation data require chained transformations to be-come useful information products, and translation of imagesto interpretable form is a prerequisite in this context. Howeverdetection and identification of objects have been affected by
various factors such as geometrical complexity, noise, vagueboundaries, mixed pixel problems, and fine characteristics of
m (P.V. Arun), skatiyar1234@
tional Authority for Remote
g by Elsevier
ng by Elsevier B.V. on behalf of N
03
detailed structures (Daniel et al., 2006). Efficiency of pixel basedclassification techniques has been limited due to increased reso-
lution of images which popularized object based approaches(Vapnik, 1998). Different existing object extraction algorithmsare specific to the features and adopt computationally complex
methods (Yuan, 2009; Sunil et al., 2004; Zhang et al., 2012).Literature reveals a great deal of recent approaches toward
accuracy improvement of feature based strategies (Mladinich,
2010; Qi et al., 2010; Carlos et al., 2012). Computing tech-niques such as neural networks, genetic algorithms, and fuzzylogic followed by probabilistic concepts such as random fieldvariations have been extensively applied in this context (Lari
and Ebadi, 2011; Wang et al., 2009; Chi, 2004). Literaturehas also revealed many object enhancement filters as well asintensity-based approaches (Jenssen et al., 2003; Haralick
et al., 1983). N-dimensional classifiers as well as random fieldconcepts and different transformation techniques have beenalso applied for accurate detection (Hosseini and Homayoun,
2009; Kumar and Hebert, 2003; Chang and Kuo, 2006).
ational Authority for Remote Sensing and Space Sciences.
164 P.V. Arun, S.K. Katiyar
Contextual information is a key factor for real-time detectionto avoid ambiguity; knowledge-based classification approachessuch as predicate calculus have been recently used in this con-
text (Porway et al., 2008; Harvey, 2008; Qi et al., 2010).Different image interpretation features such as tone, tex-
ture, pattern and color are generally adopted for feature detec-
tion. Modeling of shape is less exploited in this context and is amajor factor in distinguishing different entities (Lindi, 2004;Mladinich, 2010). These studies have found that inverse map-
ping of cellular automata (CA) using genetic algorithm (GA)can be adopted for effective modeling of feature shapes(Orovas and Austin, 1998). CA is also found to be better forcontext representation than predicate calculus since the latter
lacks image compatible forms suitable for spatial relations(Mitchell et al., 1996; Porway et al, 2008). Spectral and spatialinformation can be combined using an adaptive kernel strategy
to improve effectiveness of the approach. Probabilisticpredicate rules in conjunction with evolutionary computingtechniques are found to be effective for contextual rule
representation.In this paper we present a frame work using CNN ap-
proaches along with adaptive kernel strategy and corset opti-
mization for accomplishing accurate detection. Automaticobject modeling, adaptive kernel mapping, automatic interpre-tation, and intelligent interpolation are salient features of thiswork. Accuracy of developed methodologies has been com-
pared with contemporary approaches using different satelliteimages of the study areas.
2. Theoretical back ground
2.1. Random modeling techniques
Evolutionary computing approaches such as CA, GA andtheir variants such as cellular neural network (CNN) and mul-
tiple attractor cellular automata (MACA), have been found tobe useful for modeling random features. CNN (Orovas andAustin, 1998; Mitchell et al., 1996) is an analogue parallel com-
puting paradigm defined in space and is characterized by local-ity of connections between processing elements (cells orneurons). Cell dynamics of this continuous dynamic systemmay be denoted using ordinary differential equations as given
in Eq. (1), where vector G is the gene which determines therandom nature.
XkðtÞ ¼ �X1þ fðG;Yk;UKÞ ð1Þ
CNN is effectively used for modeling object shape to facilitate
feature interpretation. Random rules governing the shape of afeature can be identified by evolving the feature from a singlestate using CNN and GA. Abstract representations of objects
are obtained by evolving features continuously until they canbe separated from the background.
MACA is a special type of CA with different local rules ap-
plied to different cells and will converge to certain attractorstates on execution (Sikdar et al, 2000). In an n-cell MACAwith 2m attractors, there exist m-bit positions at which attrac-tors give pseudo-exhaustive 2 m patterns. MACA is initialized
with an unknown pattern and operated for a maximum (depth)number of cycles until it converges to an attractor. PEF bitsafter convergence are extracted to identify the class of the pat-
tern and are compared with stored rules to interpret the object.
2.2. Mixture density kernel
Mercer kernel functions measure the similarity between twodata points that are embedded in a high, possibly infinite,dimensional feature space. mixture density kernel (MDK) is
a mercer kernel that measures the number of times an ensem-ble agrees that two points arise from same mode of probabilitydensity function (Srivastava, 2004). It may be described usingequation (2) where ‘M’ is the number of clusters and P (Cm/Xi)
is the probability that data point ‘Xi’ belongs to Cm.
KðXi;XjÞ ¼1
ZðXi;XjÞXMm¼1
XCm
Cm¼1Pm Cm=Xið ÞPm Cm=Xj
� �ð2Þ
Mixture density kernels are used to integrate an adaptive ker-nel strategy to the SVRF based clustering. This approach facil-itates learning of kernels directly from image data rather than
using a static approach.
2.3. Support vector random field (SVRF)
SVRF (Schnitzspan et al, 2008; Lee et al., 2005) is a discreterandom field (DRF) based extension for SVM. It considersinteractions in the labels of adjacent data points while preserv-
ing the same appealing generalization properties as the supportvector machine (SVM). The SVRF function is presented inequation (3), where Ci(X) is a function that computes features
from observations X for location i, O(yi, i(X)) is an SVM-based observation matching potential and V (yi, yj, and X) isa (modified) DRF pair wise potential.
PðYjXÞ ¼ 1
Zexp
Xi2S
logðOðyi;CiðXÞÞÞ þXi2S
Xj2Ni
Vðyi; yj;XÞ( )
ð3Þ
SVRF is used to implement initial clustering to segment vari-ous objects for accurate detection and interpretation.
2.4. Coreset
Coreset (Agarwal et al, 2001; Badoiu et al, 2002) is a small sub-set of a point set, which is used to compute a solution that
approximates solution of the entire set. Let l be a measurefunction (e.g., width of a point set) from subsets of Rd tonon-negative reals R+U{0} that is monotone, i.e., for P1 C
P2, l(P1) 6 l(P2). Given a parameter e > 0, we call a subsetQ C P as an e-Coreset of P (with respect to l) if (1 � e)l(P) 6 l(Q). Coreset optimization can be adopted to reducethe number of pixels required to represent an object by pre-
serving its shape. Hence it can be used to reduce the complex-ity of CA based inverse evolution.
2.5. Probabilistic decision strategy
Terrain features, specifically object parameters (tone, texture,type, and shape) are used for object specific sample point clus-
tering as well as selection of interpolation strategies. Rules forprediction are stored in the form of grammar which constitutesof cause effect relations and are updated dynamically. Proba-bilistic rules are used to produce most likely predictions based
on the previous experiences. This approach implements a fac-tored product model, with separate decision strategies, whose
An evolutionary computing frame work toward object extraction from satellite images 165
preferences are combined by efficient exact inference, using anA* algorithm (Marie-Catherine, 2006). Rules are used for aneffective learning strategy whereby the system performance is
improved through each detection.
3. Experiment
3.1. Dataset description
Different satellite images of Bhopal and Chandrapur havebeen used as test images for evaluating the system performancewith reference to various queries. Details of the satellite data
used for these investigations are summarized in (Table 1).The ground truthing information has been collected using aGlobal Navigational Satellite System (GNSS) survey con-
ducted over Bhopal and Chandrapur during October andNovember 2012, respectively.
3.2. Methodology
A schematic representation of the adopted methodology is pre-sented in Fig. 1. Different image features are detected usingCNN based shape modeling approaches and are further inter-
preted usingMACAbased pattern detection. Edges are detectedusing Canny operator and the information is used by variousstages. The image is clustered using adaptive SVRF approach
and the process is enhanced with boundary information as wellabstract object representation. Detected objects along withboundary information are optimized using Coreset approach
to reduce the complexity of shape modeling. The approximatedobjects along with edge information are used to model featureshapes using CNN andMACA. The interpreted objects are fur-ther interpolated using Interpolation rules.
3.2.1. Object extraction
Abstract representation of image features is initially obtainedusing edge detectors along with the CA based region growing
Table 1 Details of dataset.
S. No. Image Satellite
1 LISS 4 IRS-P6
2 LISS3 IRS-P6
3 LANDSAT-TM LANDSAT
4 PAN CARTOSAT
Object Detection (CA)
Objects
Clustering(SVRF + Adaptive Kernel)
Edge detection(Canny operator)
Image
Coresetoptimization
SegmoMA
Obje
Figure 1 Method
strategy. The image is then clustered using a mixture densitykernel based SVRF approach, and the process is enhancedusing abstract object information. Mixture density kernels are
used since they avoid the static nature of the usual kernel basedstrategies and facilitate adjusting the kernel parameters directlyfrom data. Parameters of mixture density kernels are adjusted
automatically based on ensembles, and are exploited to incor-porate contextual information as well as the adaptive kernelstrategy. Composite kernel concept is used to incorporate con-
textual information, preferably a weighted combination of ker-nels is adopted as discussed by Hosseini and Homayoun (2009)such that KðP;PiÞ ¼ lKxðP;PiÞ þ ð1� lÞKyðP;PiÞ. The valueof tuning parameter (l) is adjusted based on feature metadata
using GA based approach. Inverse mapping of CNN isexploited for the purpose, and CNN rules used to evolve a par-ticular feature are used to distinguish it. Rules corresponding to
various features are thus deducted and are mapped in a prologdata base.
3.2.2. Coreset optimization
Detected objects along with boundary information are opti-mized using the Coreset approach to reduce the complexityof shape modeling. Approximation of features using a Coreset
based approach will help to reduce the number of pixels con-siderably without losing the original shape. Thus features areapproximated without affecting the results of inverse modeling
(Back and Ron, 2005). GA is used along with Coreset to imple-ment a shape preserved approximation. The most suitable pos-sibilities of approximations are selected using GA that
measures the similarity of shapes by comparison of PseudoExhaustive Field (PEF) bits.
3.2.3. Object interpretation
Clustered objects along with edge information are utilized tomodel feature shapes using CNN and MACA. Detected ob-jects are further interpreted using shape-rule mapping thatmaps objects to corresponding MACA rules. Interpolation
Date of procurement Resolution (m)
November, 2012 5.8
November, 2012 23.5
November, 2012 30
November, 2012 2.5
InterpretationRules (Predicate calculus based)
Object interpretation
Rule-Shape mapping
mented object deling using CNN & CA
Interpolation Rules
ct Interpolation
ology adopted.
Table 2 Comparative analysis.
S. No Sensor Methodology Kappa statistics Overall accuracy (%)
1 LISS 3 Mahalanobis 0.90 91.4
Minimum distance 0.91 92.8
Maximum likelihood 0.93 94.8
Parallelepiped 0.93 94.6
Feature space 0.94 95.1
CNN Approach 0.95 96.3
2 LISS 4 Mahalanobis 0.91 92.1
Minimum distance 0.92 93.5
Maximum likelihood 0.94 95.1
Parallelepiped 0.94 95.8
Feature space 0.95 95.3
CNN 0.97 97.6
3 LANDSAT-TM Mahalanobis 0.78 81.4
Minimum distance 0.81 82.3
Maximum likelihood 0.84 85.8
Parallelepiped 0.83 84.6
Feature space 0.82 83.3
CNN 0.91 91.8
166 P.V. Arun, S.K. Katiyar
of features such as roads and rivers is accomplished using CA
rules integrated with stored predicate rules. In addition to fea-ture interpretation, these rules are also used to guide mutationand crossover of GA to increase efficiency. Coreset based opti-
mization is used for feature approximation so that objects canbe effectively mapped to lower cellular automata classes.MACA is automatically initialized with most likely patterns,to identify the class of an object in less than log (n) time.
3.2.4. Topology interpretation and interpolation
PCFG rule sets are used to govern the topology extraction andrelative positions are determined based on the coordinate infor-
mation associated with each feature. Comparisons of boundarypixel positions are adopted for determining relative positions ofrandom features. A relative rectangular co-ordinate system is
assumed for images if exact coordinate information is notavailable. Topology information, along with simple spatialbuffering, is adopted to process the proximity queries. Inter-
polations of features such as roads and rivers are accomplishedby using CA rules integrated with stored predicate rules.
4. Results
Investigations of feature extraction process over various satel-lite images (details in Table 1) revealed that considerable suc-
cess has been achieved with the approach. Kappa statistics(Congalton, 1991), overall accuracy (Stine et al., 2010) andareal extent comparisons are adopted for an object specificaccuracy evaluation. Ground truthing has been conducted by
means of Google earth and Global Navigation Satellite Sys-tems (GNSS) survey over the study areas. Different featuressuch as rivers, roads, lakes, and urban structures were consid-
ered for comparative evaluation. The results are of compara-tive analysis are summarized in (Table 2).
Table 2 observations indicate a considerable improvement
in performance using CNN based approach when comparedto its traditional counterparts.
Areal extents of features namely lakes, coal mines, and
parks were selected for comparative analysis, since these fea-tures are having well defined and fixed geometry. Original sur-face areas of various extracted features were calculated by
manual digitization and comparative analysis of results is pre-sented in (Table 3).
The comparison of areal extents also indicates that CNNapproach yields better results as the areas of features extracted
using CNN closely resemble the areas calculated throughground truthing. The visual results of a sample of features ex-tracted by the system are given below which also reveal the
accuracy of CNN approach.Road network as well as water body extracted from the
LISS 3 imagery of Bhopal is presented in (Fig. 2) which reveals
random modeling capability of the approach.(Fig. 3) Presents two alternative extractions of road net-
work from PAN image of Bhopal, i.e., some roads are clearlyvisible however some are not continuous and clear (feeble).
These figures illustrate the capability of system for customizedmodeling (semi automatic approach) by changing the contex-tual or interpolation/interpretation rules as discussed in
methodology.Effective interpolation of feeble road networks requires
manual interpretation. Fig. 4(a) shows a road network initially
extracted and Fig. 4(b) shows a refined detection after semiautomatic interpolation (interpolation threshold is selectedby human intervention). We can also see a falsely interpreted
road intersection (circled in Fig. 4(a)) is corrected by humanintervention (circled in Fig. 4(b)).
Detection of certain features such as roads and clouds re-quired a semi automatic approach for detection rather than
a complete automatic method, i.e. a priori training is requiredto distinguish poorly defined road networks. The main disad-vantage of the method is its computational complexity which
can be improved by Coreset optimization and similar approx-imation techniques. Complexity can be further reduced bystoring the detected rule variations; optimization methods such
as GA can be exploited to optimize the strategy. This research
Table 3 Comparative analysis of areal accuracy.
S. No. Sensor Feature Reference area (km2) Methodology Areal extent (km2)
1 LISS3 Lake 32.48 Mahalanobis 24.31
Minimum distance 23.40
Maximum likelihood 25.12
Parallelepiped 26.24
Feature space 26.87
CNN approach 28.71
2 LISS3 Parks 2.13 Mahalanobis 0.51
Minimum distance 0.72
Maximum likelihood 1.53
Parallelepiped 1.14
Feature space 1.46
CNN approach 2.01
3 LISS4 Lake 32.81 Mahalanobis 25.42
Minimum distance 24.31
Maximum likelihood 27.37
Parallelepiped 28.58
Feature space 26.82
CNN 29.01
4 LISS4 Parks 2.37 Mahalanobis 0.82
Minimum distance 0.89
Maximum likelihood 1.45
Parallelepiped 1.37
Feature space 1.56
CNN 2.21
5 LANDSAT Open cast coal mines 36.48 Mahalanobis 22.14
Minimum distance 48.40
Maximum likelihood 31.12
Parallelepiped 28.74
Feature Space 30.91
CNN approach 36.13
6 LANDSAT Water body 7.88 Mahalanobis 3.23
Minimum distance 4.81
Maximum likelihood 5.51
Parallelepiped 5.08
Feature space 5.32
CNN Approach 7.38
(a) Road Network (b) River
Figure 2 Features extracted using CNN method.
An evolutionary computing frame work toward object extraction from satellite images 167
(a) Original image (b) Road network (c) Road Network
Figure 3 Road network extracted from PAN sensor image using CNN method.
(a) Original Image (b) Initial extraction (c) Refined Detection
Figure 4 Illustration of manual intervention requirement.
168 P.V. Arun, S.K. Katiyar
provides a basic framework and further investigations areneeded to optimize it.
5. Conclusion
We have discussed the use of evolutionary computing tech-
niques for effective object detection. Feature shape model-ing and context knowledge representation have beenaccurately accomplished using the approach. Frame work
has been critically evaluated and its integration to existingapproaches has also been investigated. Complexity of theapproach has been considerably reduced using Coreset
based approximation. Disambiguation of features, enhanceddetection, self-learning, minimal human interpretation, andreliability are features of the system. Further investigationsare needed on the improvement of proposed framework,
especially on parallelizing and optimizing different opera-tions for complexity reduction.
References
Agarwal, P.K., Aronov, B., Sharir, M., 2001. Exact and approxima-
tion algorithms for minimum-width cylindrical shells. Discrete
Computational Geometry 26 (3), 307–320.
Back, T., and Ron, B., 2005. Using genetic algorithms to evolve
behavior in cellular automata, Proceedings of the 4th international
conference on Unconventional Computation, 3699 (3), 1–10.
Badoiu, M., Har-Peled, S., Indyk, P., 2002. Approximate clustering
via Coreset. Proceedings of 34th Annual ACM Symposium, Theory
of Computation, 250–257.
Carlos, A., Aguirre-Salado, Eduardo, J., Trevino-Garza, Oscar, A.,
Aguirre-Calderon, Javier Jimenez-Perez, Marco, A., Gonzalez-Tagle,
Liliana Miranda-Aragon, J., Rene Valdez-Lazalde, Alejandro, I.,
Aguirre-Salado, Guillermo Sanchez-Dıaz., 2012. Forest cover map-
ping in North-Central Mexico: a comparison of digital image
processing methods. GIScience & Remote Sensing 49 (6), 895–914.
Chang, T., Kuo, C.J., 2006. Texture analysis and classification with
tree-structured wavelet transform. IEEE Transactions on Image
Processing 2 (4), 429–441.
Chi, K.C., 2004. Surface registration using a dynamic genetic
algorithm. Journal of Pattern Recognition 37 (1), 105–117.
Congalton, R.G., 1991. A review of assessing the accuracy of
classifications of remotely sensed data. Remote Sensing of Envi-
ronment 37 (1), 35–46.
Daniel, A.L., Gang Hong, Yun Zhang, 2006. Performance assess-
ment of automated feature extraction tools on high resolution
imagery, MAPPS/ASPRS Conference, American Society of
Photogrammetry and Remote Sensing, San Antonio, Texas,
November 2006.
Haralick, R.M., Watson, L.T., Laffey, T.J., 1983. The topographic
primal sketch. International Journal of Robotics Research 2 (1),
50–72.
Harvey, M., 2008. User centric evaluation of semi automated feature
extraction from remote sensing images. IGARSS 23, 234–240.
Hosseini, R.S., Homayoun, S., 2009. A SVMS-based hyper spectral
data classification algorithm in a similarity space. Workshop on
Hyper spectral Image and Signal Processing: Evolution in Remote
Sensing – WHISPERS 09 21 (2), 1–4.
Jenssen, R., Hild, K.E., Erdogmus, D., Principe, J.C., Eltoft, T., 2003.
Clustering Using Renyi’s Entropy. Proc. International Joint
Conference on Neural Networks. Portland, USA, pp. 523–528.
Kumar, S., Hebert, M., 2003. Discriminative random fields: a
discriminative framework for contextual interaction in classifica-
tion. Proceedings of Ninth IEEE International Conference on
Computer Vision 2 (4), 1150–1157.
Lari, Z., Ebadi, H., 2011. Automated building extraction from high-
resolution satellite imagery using spectral and structural informa-
tion based on artificial neural networks. International Conference
on Object Extraction 21 (2), 23–27.
Lee, C., Schmidt, M., Greiner, R., 2005. Support Vector Random
Fields for Spatial Classification’’. 9th European Conference on
An evolutionary computing frame work toward object extraction from satellite images 169
Principles and Practice of Knowledge Discovery in Databases
(PKDD). Portugal, USA, p. 196.
Lindi, J.Q., 2004. A review of techniques for extracting linear features
from imagery. Photogrammetric Engineering & Remote Sensing 70
(12), 1383–1392.
Marneffe, M., MacCartney, B., Manning, C., 2006. Generating typed
dependency parses from phrase structure parses. Proceedings of 5th
International Conference on Language Resources and Evaluation
(LREC 2006), 24–26th May, Genoa, Italy, pp. 449–454.
Mitchell, M., Crutchfield, J.P., Das, R., 1996. Evolving cellular
automata with genetic algorithms: a review of recent work. First
International Conference on Evolutionary Computation and Its
Applications (EvCA’96) 1 (1), 120–130.
Mladinich, Carol S., 2010. An evaluation of object-oriented image
analysis techniques to identify motorized vehicle effects in semi-arid
to arid ecosystems of the American West. GIScience & Remote
Sensing 47 (1), 53–77.
Orovas, C., Austin, J., 1998. A cellular system for pattern recognition
using associative neural networks. IEEE International Workshop
on Cellular Neural Networks and their Application 2 (5), 11.
Porway, J., Wang, K., Yao, B., Zhu, S.C, 2008. A hierarchical and
contextual model for aerial image understanding. IEEE Computer
Society Conference on Computer Vision and Pattern Recognition
49 (7), 155–160.
Qi, Pengcheng, Zhao, Chuanyan, Feng, Zhaodong, 2010. GIS- and
machine learning––based modeling of the potential distribution of
broadleaved deciduous forest in the Chinese Loess Plateau.
GIScience & Remote Sensing 47 (1), 99–114.
Schnitzspan, P., Mario, F., Bernt, S., 2008. Hierarchical support vector
random fields: joint training to combine local and global features.
Computer Vision – ECCV 5303 (5), 527–540.
Sikdar, B.K., Paul, K., Biswas, G.P., Yang, C., Bopanna, V.,
Mukherjee, S., Chaudhuri, P.P., 2000. Theory and application of
GF(2P) cellular automata as on-chip test pattern generator.
Proceedings of 13th International Conference on VLSI Design
India 556–561.
Srivastava, A. N., 2004. Mixture density mercer kernels: a method to
learn kernels directly from data. SIAM Data Mining Conference,
Lake Buena Vista, Florida, USA: Florida University. 22–24 April,
369–378.
Stine, Roy S., Chaudhuri, Debasish, Ray, Peter, Pathak, Prasad, Hall-
Brown, Mary, 2010. Comparison of digital image processing
techniques for classifying Arctic Tundra. GIScience & Remote
Sensing 47 (1), 78–98.
Sunil, R.R., Dennis, D.T., Eric, K., Charles, G.O., 2004. Comparing
Spectral and Object Based Approaches for Classification and
Transportation Feature Extraction from High Resolution Multi-
spectral Imagery. ASPRS Annual Conference Proceedings, Denver:
Colorado.
Vapnik, V., 1998. Statistical Learning Theory. Wiley Publishers Inc.,
New York, pp. 230–240.
Wang, H., Yang, Y., Ma, S., Guo, C., 2009. Automatic object
extraction based on fuzzy mask. Intelligent Systems and Applica-
tions, ISA International Workshop On 23 (5), 1–4.
Yuan, J., 2009. Automatic road extraction from satellite imagery using
LEGION networks. Proceedings of International Joint Conference
on Neural Networks 45 (2), 120.
Zhang, Daxiang., Zhang, Chuanrong., Cromley, Robert., Travis,
David., Daniel Civco, 2012. An object-based method for contrail
detection in AVHRR satellite images. GIScience & Remote Sensing
49 (3), 412–427.