Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Improving urban classification through fuzzy supervised classificationand spectral mixture analysis
J. TANG*{, L. WANG{ and S. W. MYINT§
{Department of Geography and Environmental Systems, University of Maryland,
Baltimore County, 1000 Hilltop Circle, Baltimore, MD 21250, USA
{NCGIA, Department of Geography, University of Buffalo, State University of New
York, Buffalo, NY 14261
§Department of Geography, Arizona State University, Tempe, AZ 85287-0104, USA
(Received 22 July 2006; in final form 15 November 2006 )
In this study, a fuzzy-spectral mixture analysis (fuzzy-SMA) model was
developed to achieve land use/land cover fractions in urban areas with a
moderate resolution remote sensing image. Differed from traditional fuzzy
classification methods, in our fuzzy-SMA model, two compulsory statistical
measurements (i.e. fuzzy mean and fuzzy covariance) were derived from training
samples through spectral mixture analysis (SMA), and then subsequently applied
in the fuzzy supervised classification. Classification performances were evaluated
between the ‘estimated’ landscape class fractions from our method and the
‘actual’ fractions generated from IKONOS data through manual interpretation
with heads-up digitizing option. Among all the sub-pixel classification methods,
fuzzy-SMA performed the best with the smallest total_MAE (MAE, mean
absolute error) (0.18) and the largest Kappa (77.33%). The classification results
indicate that a combination of SMA and fuzzy logic theory is capable of
identifying urban landscapes at sub-pixel level.
1. Introduction
Understanding the extent, distribution and evolution of urban landscape is critical
in analysing and explaining its structure. Such information plays a vital role in
predicting future urban landscape pattern and guiding the direction of urban
planning for satisfying both environmental and socioeconomic sustainability.
Moderate resolution satellite images such as Thematic Mapper (TM) provide a
distinctive opportunity to capture urban landscape over a large area in a timely and
cost-effective manner (Yang and Lo 2002, Song 2005). Traditional per-pixel
methods, nevertheless, were inadequate in classifying urban landscape from
moderate satellite imagery due to the presence of mixed land cover categories in
each pixel, as a result of the rapid alternation of artificial and natural objects within
a small distance in a typical urban setting (Mesev et al. 2001, Small 2003, Lu and
Weng 2004). Specifically, treating each individual pixel to a pure class, as has been
assumed in such per-pixel classifiers, prevents the representation of varying land
cover and land use information below the inherited pixel resolution (Gong and
Howarth 1989, Foody and Arora 1996, Small 2004).
*Corresponding author. Email: [email protected]
International Journal of Remote Sensing
Vol. 28, No. 18, 20 September 2007, 4047–4063
International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online # 2007 Taylor & Francis
http://www.tandf.co.uk/journalsDOI: 10.1080/01431160701227687
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
Instead of assigning only one class per pixel, sub-pixel classification approachesquantify multiple class memberships for each pixel (Foody and Cox 1994, Foody 1996,Foody 1997, Myint 2006). Spectral mixture analysis (SMA) represents a widely usedmethod for tackling the decomposition of mixed pixels. The linear SMA has so far beenthe most popular approach among the SMA family methods given its simplemathematical form. It has been adopted to generate fractions of urban physicalconstituents according to the spectral variability inherited in either multi- orhyperspectral imagery (Settle and Drake 1993, Tompkins et al. 1997, Small 2001, Wuand Murray 2003, Lu and Weng 2004). To standardize the application of linear SMA,Ridd (1995) first proposed the vegetation-impervious surface-soil (VIS) model torepresent the biophysical composition of an urban environment as vegetation, imper-vious surface and soil. This model was later widely adopted as a scheme for choosingappropriate endmembers in the linear SMA model (Madhavan et al. 2001, Phinn et al.
2002). Some recent studies have further refined the VIS model, e.g. impervious surfacewas classified into low-albedo and high-albedo endmember (Small 2001, Wu and Murray2003) or shade endmember was added as a separate endmember (Lu and Weng 2004).
One critical problem encountered in the application of linear SMA is that thenumber of endmembers has to be kept at a fairly limited size since it is constrainedby the number of available bands offered by the remotely sensed imagery (Radeloffet al. 1999). Moreover, the impervious surface in the VIS model is not appropriate tobe treated as one endmember given that the presence of various components is notunusual (Wu and Murray 2003, Lu and Meng 2004). A number of approaches havebeen developed to address this question by either modification of endmembers oralternation from a simple linear SMA to a nonlinear function. Examples of thesemethods include iterative spectral unmixing (Meer 1999, Theseira et al. 2002);endmember bundles approach (Bosdogianni et al. 1997, Bateson et al. 2000); neuralnetwork (Foody 1997, Carpenter et al. 1999, Linderman et al. 2004); mixturediscriminant analysis (Ju et al. 2003); and Bayesian spectral unmixing (Song 2005).
Fuzzy classification is another robust method which provides class membershipsat the pixel level. After Zadeh (1965) introduced the fuzzy set concept in 1965 todescribe imprecision, fuzzy set concept was then widely applied to sub-pixel imageanalysis. Fuzzy C-means (FCM) algorithm (Bezdek et al. 1984, Foody 1992), fuzzy-maximum likelihood classification (Wang 1990), and neuro-fuzzy (Carpenter et al.
1999), are three prevalent approaches in this category. Based on an iterativeoptimization of clustering, FCM limits its application when the a priori knowledgeof number of cluster and local optima is not enough (Bandyopadhyay 2005) beforeclassification. Fuzzy-maximum likelihood classification and neuro-fuzzy are twosupervised soft classifications by softening the output of traditional maximumlikelihood classification and neural network classification. Typically, these twomethods need to identify the training samples to characterize the classes in theimage. However, inability to account for mixed pixels when collecting informationfrom the training samples, as a general practice when the aforementioned methodswere applied, has circumvent the implementation of a ‘fully’ fuzzy classification(Zhang and Foody 2001). Therefore, it is critical to choose proper endmembers andsolve the mixed pixel problem for the fuzzy training data in order to perform anaccurate supervised soft classification in urban area.
In this study, we propose a new fuzzy-SMA method to improve the estimation of
urban land use/land cover fractions from Landsat ETM + imagery. The method is
based on both fuzzy supervised classification and linear SMA. Particularly,
endmember signatures of training samples in the fuzzy classifier are not treated as
4048 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
constants, but as ratios derived from the linear SMA. In order to examine the
effectiveness of this method, we finally compare the classification results from fuzzy-
SMA with those from partial-fuzzy and linear SMA.
2. Study site and data preparation
The study site was selected in the eastern metropolitan area in Houston, Texas
(figure 1), covering an area of 1200 km2. Surrounded by the Buffalo Bayou, this
region is one of the fastest growing centres of petroleum manufacturing,
warehousing and transportation in Houston. The typical land use/land cover classes
include: residential area, commercial/industrial area, transportation, woodland,grassland and barren/soil. Considering the practical value for later applications, we
adopt a class system with both the land use and land cover classes. Based on the field
experience and spectral attribute in the image, the residential area, transportation
and commercial/industrial area can be referred to as the impervious surface with
low, medium and high albedo, respectively.
A subset image from Landsat 7 ETM + (path 25, row 39) acquired on 2 January
2003 is employed in this study. For the sake of reducing atmospheric distortion, the
digital numbers (DN) at each band were first converted to the normalized at-sensorreflectance using the method that was introduced in Markham and Barker (1987).
The minimum noise fraction (MNF) transformation was then applied to remove the
correlation existing between the six bands (the thermal band was excluded since it
has a different spatial resolution). Bundled IKONOS images, comprising a 1-metre
panchromatic and a 4-metre multispectral image, that was acquired on 2 January
2002, was adopted as reference for choosing the training samples for different
endmembers as well as the test samples of known fractions from Landsat ETM + .
3. Methodology
A systematic framework of our model is presented in figure 2. Three sets of
proportion maps were derived using linear SMA, partial-fuzzy and fuzzy-SMA
Figure 1. Study site in Houston, Texas from a Landsat ETM + imagery in January 2003.
Improving urban classification 4049
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
classifications, respectively. The major difference between the partial-fuzzy and
fuzzy-SMA lies in the training spectrum among these classes. In the fuzzy-SMA, we
applied the linear SMA to obtain the fuzzy mean and fuzzy covariance matrix for
the training samples instead of treating them as constants. Based on these statistical
measures, we derived the fractions of urban images using the fuzzy set theory. The
following sections will describe the fuzzy logic and SMA in details.
3.1 Fully-fuzzy supervised classifier and its training data
Conventional discriminant analysis is a widely used classifier in the classification of
remotely sensed data. This method allocates each pixel to the class which has the
highest a posteriori probability of membership (Foody 1996, Richards and Jia
1999). Without considering the prior probability, the a posteriori probability, also
named as discriminant function, is expressed as:
gi xð Þ~{ln Mij j{ x{mið ÞtMi{1 x{mið Þ ð1ÞWhere x is the b-dimension pixel vector for b bands; mi and Mi are the mean vector
and variance–covariance matrix of training samples of class i, respectively (Richards
and Jia 1999).
To implement a fully-fuzzy supervised classification, we derived the fuzzy mean
and fuzzy variance matrix (Wang 1990) from the training samples in addition to
‘soften’ the output of conventional ‘hard’ classifiers by the fuzzy classifier. The fuzzy
mean mi* and fuzzy variance matrix Mi can be calculated using equations (2) and (3):
m�i ~
PS
i~1
f xið Þxi
PS
i~1
f xið Þð2Þ
Figure 2. Flowchart for the fuzzy-spectral mixture analysis (fuzzy-SMA) model. Note:IKONOS image is used to select pure endmembers/training samples and assess theclassification accuracy of the resultant proportion maps from fuzzy-SMA model.
4050 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
and
Mi~
PS
i~1
f xið Þ xi{m�i� �
xi{m�i� �T
PS
i~1
f xið Þð3Þ
where xi is the pixel spectral vector of the training samples of class i; S is the total
number of training samples; and f(xi) is the membership function of class i.
Training samples for six classes, including residential area, commercial/industrial
area, transportation, woodland, grassland and barren/soil, were selected from the
Landsat ETM + image by referring to the IKONOS image. Water was masked out
before the analysis since it only occupies a small area in the study site. Samples for
each class come from thirty training plots and each training plot covering 90690 m2
on the ground. Due to the mixture of membership in each training plot, it is
important to identify the proportions of endmembers in these training plots before
an accurate fuzzy mean vector and fuzzy variance matrix can be derived. For
example, Wang (1990) estimated the endmembers’ fractions from training plots with
manual interpretations from the aerial photography. However, the inaccurate
estimation of fuzzy mean vector and fuzzy variance matrix through manual
digitization often leads to poor classification results. In this research, we applied the
SMA to identify the fraction membership of the training samples.
3.2 Spectral mixture analysis
Spectral mixture analysis (SMA) aims to map the fractions of landscape classes
within the mixed pixels using an inverse least square devolution method and the
spectra of pure endmembers (Shimabukuro and Smith 1991). In this study, we
employed a modified constrained least square method for the linear spectral
unmixing analysis using the following expression:
Rn~XE
e~1
rn, efezen withXE
e~1
fe~1 and 0ƒfeƒ1 ð4Þ
where Rn is the normalized spectral reflectance for each band n; fe is the fraction of
endmember e; E is the total number of endmembers; rn,e denotes the normalized
spectral reflectance of endmember e within a pure pixel on band n; and en is the
residual.
The choice of spectral reflectance for the endmembers is critical for the spectral
mixture analysis. An optical approach for selecting endmembers is to use
laboratory-based spectra from the field. However, substantial problems exist in
accounting for the atmospheric conditions when upscaling the field spectral value to
the level that can be employed with the satellite sensor data (Settle and Drake 1993,
Wu and Murray 2003). Alternatively, the endmembers’ spectra can be obtained by
locating pure pixels on the same image (Lu and Weng 2004). In our study, the
endmember for each class is carefully selected from the representative homogenous
pixels, i.e. having a full membership to the specified class and zero membership to
the other classes. To ensure the purity of the chosen pixels from the ETM + image,
higher spatial resolution IKONOS image was employed as a reference image. The
clear delineation of feature spaces among the six classes in the first three principle
Improving urban classification 4051
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
components (figure 3) suggests that the reflectance spectra of ETM + image could be
represented by a six-endmember linear mixing model.
3.3 Accuracy assessment
The accuracy assessment is an essential step for the landscape mapping from
remotely sensed data (Foody 2002). An independent set of test samples was chosen
for this evaluation. Considering the statistical requirement and the size of our study
area, we randomly selected 200 sample plots with each at the size of 90690 m2 from
the IKONOS image using ERDAS Imagine accuracy assessment module (figure 4).
For each sample unit, its actual land use types were acquired through digitizing the
IKONOS image (figure 4(b)). The actual fractions of each landscape class in the
sample units were obtained through dividing the class area by the total area of one
sample plot.
As stated by Foody (1996), conventional confusion matrix is inappropriate for the
fuzzy classification evaluation since multiple memberships exist with each pixel. In
this research, we adopted two accuracy assessment methods to evaluate the accuracy
of urban composition for each land class. The first method is the mean absolute
error (MAE) (Pontius et al. 2004, Willmott and Matsuura 2006), defined as follows:
MAEi1P, 2P� �
~
PN
n~1
1Pi{2Pi
��
��
Nð5Þ
Total MAE~
PJ
i~1
MAEi
2ð6Þ
Figure 3. Feature space representation of the first three principle components. Red:residential; magenta: commercial/industrial; yellow: transportation; green: woodland; cyan:grassland; black: barren/soil
4052 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
Where 1Pi is the estimated land use composition fraction from the classification and2Pi is the ‘actual’ land use composition fraction digitized from IKONOS for class i in
the test sample n; N is the total number of test samples and J represents the number
of categories.
The generalized cross-tabulation matrix as introduced by Pontius and Cheuk
(2006) was also adopted to further analyse the disagreement between the estimated
results and reference maps. Specifically, a composite operator was used to calculate
(a) (b)
Figure 4. The sample fraction validation using IKONOS, (a) sample distribution in thestudy area; (b) Thematic Mapper (TM) and the corresponding IKONOS sample unit, themagenta line is used to digitize the residential area in this unit.
Figure 5. Classified images generated by benchmark Maximum Likelihood Classificationmethod.
Improving urban classification 4053
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
the entries in the cross-tabulation matrix through:
Pi, j~min 1Pi,2Pi
� �if i~j
~ 1Pi{min 1Pi,2Pi
� �� �|
2Pj{min 1Pj,2Pj
� �
PJ
j~1
2Pj{min 1Pj, 2Pj
� �� �
2
6664
3
7775
if i=jð7Þ
where Pi,j is the estimated entry in the cross-tabulation matrix for classes i and j.
Once all entries were filled, overall accuracy and Kappa values were calculated in
the same manner as that applied in the traditional hard classification. Further
details about the accuracy assessment in soft classification can be found in Pontius
and Cheuk (2006).
4. Results and discussion
Figure 5 is the classified image generated by the supervised MLC method. It reveals
the central-zonal spatial pattern throughout the whole study area. Most of the
industrial or commercial area is located in the central business district (CBD) or
distributed along the major road or the Buffalo Bayou. High-density residential
areas are found around the CBD, especially around the Highway I45 and the
Buffalo Bayou in the southern study area. The woodland is mainly seen in the north-
eastern area of Houston. Grassland is mainly distributed in the southern part,
mostly around the Houston International Airport. Although the supervised
classification result describes the overall urban land use pattern, it could not
provide a realistic or accurate measurement on urban pattern since it neglected the
relative abundance of surface materials within a pixel.
Conversely, a richer amount of land use information has been derived by the ‘soft’
classification. Figure 6 illustrates six fraction images unmixed from the ETM +image using the linear SMA model. We used the grey level to designate the
percentage of a specific class that a pixel contains. Therefore, a brighter pixel means
a higher fraction. It is noticeable that the fraction images of all classes have a similar
grey level. These results may be attributed to the spectral similarity among the six
classes. The most obvious misinterpretation is the confusion between transportation
and commercial/industrial area in the central part of the study area. Therefore, the
linear SMA would result in substantial misinterpretation due to the spectral
similarity among classes.
The fraction images derived from the partial-fuzzy method (figure 7) present a
more obvious pattern than the linear SMA. Particularly, fractions of the
commercial/industrial area and the residential area have been refined. The bright
pixels in the two-fraction images correspond well to the known impervious area
within the original ETM + image. Moreover, the woodland fraction image concurs
with the field woodland distribution, i.e. almost zero in the CBD and ranging
between 60–80% in the residential areas.
The fraction images from the fuzzy-SMA model achieved a remarkable
improvement compared to both linear SMA and partial-fuzzy methods. In figure 8,
the brightest spot, as an indicator of high proportion for the commercial/industrial
fraction image, is located in the central area of the study area. The residential
fraction is near zero in the high-density commercial or industrial area, while it
4054 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
gradually increases to 10–50% in the transition zone between the industrial area and
residential area, and approaches 90% in the high-density residential area. The
transportation fraction image is also consistent with its actual distribution, showing
an interlaced pattern. Most grassland is located in the south-eastern corner and
peaks at almost 100% in the park area. The woodland has the least fraction in the
central area of Houston, ranging between 10 and 30%. The brightest spot of the
woodland is found in the known woodland area, the north-eastern area of Houston.
The fact that fuzzy-SMA performed better than the linear SMA can be attributed to
the incorporation of fuzzy mean and fuzzy variance, with which the spectral
overlaps between the endmembers has been significantly reduced.
Table 1 shows the accuracy indices (MAE) and the cross-tabulation matrix for the
fuzzy-SMA, partial-fuzzy, linear SMA and MLC. In table 1, a smaller index denotes
a better classification result. From this table we can easily find that the fuzzy-SMA
performed better than the other two methods since it has the smallest MAE (0.35). A
detailed examination with the MAE shows that fuzzy-SMA performed well with the
impervious surface, e.g. the commercial or industrial area (0.06), residential area
(0.05) and transportation area (0.09). The best fraction is associated with the barren/
soil category, having the smallest MAE (0.01). A possible explanation could be that
most of the training samples for barren/soil are mixed with vegetation, especially for
the ‘fragmentized’ grassland. The fuzzy-SMA accounts for the fractions in dealing
with the training samples, thus reducing the signature confusion between the
vegetation and soil.
Figure 6. Fraction images generated by the linear spectral mixture analysis (linear SMA)model.
Improving urban classification 4055
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
Table 2 shows the cross-tabulation matrix for the MLC, linear SMA, partial-fuzzy
and fuzzy-SMA. The diagonal numbers in the matrix were underlined to highlight
the agreement between the classified results and empirical map. Evidently, fuzzy-
SMA showed an obvious improvement in the results with the highest overall
accuracy (82.11%) and Kappa (77.33%). In particular, fuzzy-SMA has the highest
agreement in the commercial/industrial area (11.63%), transportation (26.83%) and
woodland (16.40%).
Further analysis on the cross-tabulation matrix suggests that there are still some
drawbacks with the sub-pixel classification methods. In the cross-tabulation matrix,
the improvement from the MLC to the linear-SMA can mainly be found in the
transportation (from 6.02 to 12.12%), grassland (from 6.73 to12.36%) and barren/
soil (from 0.57 to 1.71%). These classes usually have a distinct spectral
characteristic. Other classes, such as the residential and commercial/industrial, due
to the spectral confusion, generated even worse results with linear-SMA.
A further comparison was carried out between the actual class fractions and
estimated fractions for the three models (figures 9–12). The distance that the test
samples deviate from the line y5x was then used to assess the closeness between the
actual proportions and estimated proportions of each class. The total distance (SD)
for each class was also shown in each plot. The closer the test samples to the line
y5x, the better the estimated result will be. In the one-to-one relationships observed
in the MLC model (figure 9), a number of fractions of landscape classes are
estimated as either 0 or 1, especially for the commercial/industrial and transporta-
tion, which indicates that the conventional ‘hard’ MLC approaches are ineffective
Figure 7. Fraction images generated by the partial-fuzzy method.
4056 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
for mapping urban internal structures, especially when using the coarse resolution
image.
In contrast to the MLC method, the linear SMA has largely alleviated the extreme
situation of having either ‘0%’ or ‘100%’ fractions. A much clearer correlation can
be discerned between the estimated and actual fractions (figure 10) though most
observed points do not distribute along the y5x line. As previously discussed, such a
situation may be a result of the spectral similarity among urban landscape classes.
Although many researchers have found that SMA is an effective approach in
characterizing urban landscape (Lu and Weng 2004, Wu 2004), few endmembers
(e.g. three or four surface covers) were usually the cases in their studies. When more
Figure 8. Fraction images generated by the fuzzy-spectral mixture analysis (fuzzy-SMA)model.
Table 1. Comparisons of mean absolute error (MAE) for fuzzy-spectral mixture analysis(fuzzy-SMA), linear SMA and Maximum Likelihood Classification methods.
Accuracyassessment
Residen-tial
Commercial/industrial
Transport-ation
Grass-land
Wood-land
Barren/soil
Total
MAE MLC 0.48 0.08 0.25 0.17 0.12 0.04 0.58Linear
SMA0.15 0.17 0.20 0.17 0.13 0.14 0.48
Partial-Fuzzy
0.13 0.10 0.28 0.17 0.14 0.46 0.64
Fuzzy-SMA
0.05 0.06 0.09 0.06 0.08 0.01 0.18
Improving urban classification 4057
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
Table 2. Cross-tabulation matrix (%) for the Maximum Likelihood Classification, linear spectral mixture analysis (linear SMA), partial-fuzzy and fuzzy-SMA methods. Maximum Likelihood Classification is given in roman, the linear SMA is given in bold, the partial-fuzzy method is given in italics and the
fuzzy-SMA is given in bold italics. The underlined value is the agreement between the classified results and empirical map.
Empirical map
Classifiedresult
Residential Commercial/industrial
Transportation Grassland Woodland Barren/soil Total
Residential 17.77 3.73 20.30 8.77 15.90 0.70 67.1711.84 1.71 3.86 1.83 1.51 0.11 20.86
9.15 0.25 2.28 1.37 0.66 0.01 13.7216.17 0.07 0.37 1.35 1.09 0.09 19.14
Commercial/industrial
0.13 6.20 1.48 0.38 0.29 0.54 9.022.05 5.26 3.30 1.71 3.69 0.03 16.040.01 3.51 0.89 0.35 0.09 0.00 4.850.45 11.63 1.78 2.19 0.92 0.26 17.23
Transportation 0.00 1.53 6.02 0.64 0.13 0.17 8.490.23 0.70 12.12 0.72 1.33 0.13 15.230.01 0.37 1.39 0.10 0.02 0.00 1.890.98 0.26 26.83 2.81 1.98 0.47 33.33
Grassland 0.03 0.01 0.34 6.73 1.07 0.15 8.331.20 0.98 2.32 10.36 2.18 0.04 17.080.07 0.03 0.44 4.04 0.22 0.00 4.800.10 0.00 0.03 10.07 0.71 0.14 11.05
Woodland 0.00 0.00 0.00 0.00 3.48 0.00 3.480.73 1.61 2.32 0.64 9.35 0.13 14.782.05 0.40 4.52 4.31 14.68 0.11 26.070.31 0.01 0.40 0.85 16.40 0.10 18.07
Barren/soil 0.05 0.63 1.27 0.77 0.24 0.57 3.531.96 1.82 5.52 1.96 3.07 1.71 16.046.77 7.53 19.85 7.08 5.41 2.04 48.680.00 0.06 0.00 0.07 0.03 1.02 1.18
Total 17.98 12.01 29.41 17.29 21.11 2.13 10018.01 12.08 29.44 17.22 21.13 2.15 10018.06 12.09 29.37 17.25 21.08 2.16 10018.01 12.03 29.41 17.34 21.03 2.08 100
Overall accuracy: MLC (40.77); linear-SMA (50.65); partial-fuzzy (34.81); fuzzy-SMA (82.11)Kappa: MLC (27.85); linear-SMA (40.82); partial-fuzzy (26.76); fuzzy-SMA (77.33)
40
58
J.
Ta
ng
eta
l.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
endmembers are designated within the urban area category, the method will easily
reach its limit due to the spectral closeness and overlap. The correlation between the
estimated fraction using the partial-fuzzy and actual fraction from the IKONOS is
shown in figure 11. Although the partial-fuzzy method has advantages over the
MLC method, most classes still have zero fractions in several test plots. Therefore,
we can conclude that from the MLC to the partial-fuzzy method, there is little
improvement for the urban landscape classification.
Figure 9. Relationships for the 30 test samples between the estimated fraction usingMaximum Likelihood Classification and the ‘actual’ fraction digitized from the IKONOSimage.
Figure 10. Relationships for the 30 test samples between the estimated fraction using linearspectral mixture analysis (linear SMA) and the ‘actual’ fraction digitized from the IKONOSimage.
Improving urban classification 4059
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
Compared to the MLC, linear SMA and partial-fuzzy methods, fuzzy-SMA has
achieved a promising result with the smallest overall distance to the y5x line
(figure 12). The best classification result using fuzzy-SMA was observed in the
barren/soil and residential area, with distances of 1.27% and 4.66%, respectively.
The test samples were much closer to the y5x line in the fuzzy-SMA, indicating that
the fuzzy-SMA method gives the most accurate estimation of the landscape
composition in this study area.
Figure 11. Relationships for the 30 test samples between the estimated fraction usingpartial-fuzzy method and the ‘actual’ fraction digitized from the IKONOS image.
Figure 12. Relationships for the 30 test samples between the estimated fraction using fuzzy-spectral mixture analysis (fuzzy-SMA) and the ‘actual’ fraction digitized from the IKONOSimage.
4060 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
5. Conclusions
Using moderate resolution satellite images, traditional ‘hard’ classification is usually
inappropriate for urban landscape analysis since multiple class memberships exist
within one pixel. In this paper, we developed a fuzzy-SMA method to estimate the
subpixel fractions for urban landscape and the performance was compared to linear
SMA, partial-fuzzy and MLC methods. The same sets of training and test samples
were used in these methods based on IKONOS data. Previous studies proved that
the SMA and partial-fuzzy methods were effective in the urban classification,
especially when addressing fuzziness in the training and testing samples. However,
the straightforward application of linear SMA was unsatisfactory when the number
of endmembers goes over four using TM images. Conversely, we propose to carry
out SMA in the first stage and serve its result in the calculation of fuzzy mean and
fuzzy covariance, followed up with a fuzzy classification to derive the final fractions.
It turns out that this fuzzy-SMA performs consistently better than the MLC, linear
SMA and partial-fuzzy methods.
For most subpixel classifications, endmember selection is still a critical step
towards the final success. Various unmixture methods are vulnerable with an
inappropriate identification of pure endmembers. This is particularly true without
having a higher spatial resolution reference image. Moreover, most SMA studies
only choose three or four endmembers, i.e. VIS models developed by Ridd (1995)
and Wu and Murray’s SMA models (2003). This circumvents a finer characteriza-
tion of the urban landscape since it usually has various surface types.
Shade, caused by tall buildings or trees, is an inevitable part of urban area,
especially in the developed areas. In this research, we classified it into barren and soil
instead of treating it as a separate endmember. This simplification adversely
influences modelling results, e.g. barren and soil area is overestimated and
commercial/industrial area is underestimated. In future research, we will incorporate
the shade as a single endmember to reduce the error within other classes.
Although in this study, the fuzzy-SMA method achieved considerable improve-
ments in the urban landscape classification, the exact location of each endmember
within one pixel is still unresolved. To locate each endmember, we could incorporate
the context information from its surrounding pure pixels. How to combine this
knowledge in a fuzzy classification to develop a more sophisticated soft method is
still an interesting research topic.
ReferencesBANDYOPADHYAY, S., 2005, Satellite image classification using genetically guided fuzzy
clustering with spatial information. International Journal of Remote Sensing, 26, pp.
579–593.
BATESON, C.A., ASNER, G.P. and WESSMAN, C.A., 2000, Endmember bundles: a new
approach to incorporating endmember. IEEE Transactions on Geoscience and Remote
Sensing, 38, pp. 1083–1094.
BEZDEK, J.C., EHRLICH, R. and FULL, W., 1984, FCM: the fuzzy c-means clustering
algorithm. Computer and Geosciences, 10, pp. 191–203.
BOSDOGIANNI, P., PETROU, M. and KITTLER, J., 1997, Mixture models with higher order
moments. IEEE Transactions on Geoscience and Remote Sensing, 35, pp. 341–353.
CARPENTER, G.A., GOPAL, S., MACoMBER, S., MARTENS, S. and WOODCOCK, C.E., 1999, A
neural network method for mixture estimation for vegetation mapping. Remote
Sensing of Environment, 70, pp. 138–152.
Improving urban classification 4061
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
FOODY, G.M., 1992, A fuzzy sets approach to the representation of vegetation continua from
remotely sensed data: an example from lowland health. Photogrammetric Engineering
and Remote Sensing, 58, pp. 221–225.
FOODY, G.M. and COX, D.P., 1994, Sub-pixel land cover composition estimation using a
linear mixture model and fuzzy membership functions. International Journal of
Remote Sensing, 15, pp. 619–631.
FOODY, G.M., 1996, Approaches for the production and evaluation of fuzzy land cover
classifications from remotely-sensed data. International Journal of Remote Sensing, 17,
pp. 1317–1340.
FOODY, G.M. and ARORA, M.K., 1996, Incorporating mixed pixels in the training, allocation
and testing stages of supervised classifications. Pattern Recognition Letters, 17, pp.
1389–1398.
FOODY, G.M., 1997, Fully fuzzy supervised classification of land cover from remotely sensed
imagery with an artificial neural network. Neural computing and Applications, 5, pp.
238–247.
FOODY, G.M., 2002, Status of land cover classification accuracy assessment. Remote Sensing
of Environment, 80, pp. 185–201.
GONG, P. and HOWARTH, P.J., 1989, Performance analyses of probabilistic relaxation
methods for land-cover classification. Remote Sensing of Environment, 30, pp. 33–42.
JU, J., KOLACZYK, E.D. and GOPAL, S., 2003, Gaussian mixture discriminant analysis and
sub-pixel land cover characterization in remote sensing. Remote Sensing of
Environment, 84, pp. 550–560.
LINDERMAN, M., LIU, J., QI, J., AN, L., OUYANG, Z., YANG, J. and TAN, Y., 2004, Using
artificial neural networks to map the spatial distribution of understorey bamboo
from remote sensing data. International Journal of Remote Sensing, 25, pp. 1685–
1700.
LU, D. and WENG, Q., 2004, Spectral mixture analysis of the urban landscape in Indianapolis
with Landsat ETM + imagery. Photogrammetric Engineering and Remote Sensing, 70,
pp. 1053–1062.
MADHAVAN, B.B., KUBO, S., KURISAKI, N. and SIVAKUMAR, T.V.L.N., 2001, Appraising the
anatomy and spatial growth of the Bangkok metropolitan area using a vegetation-
impervious-soil model through remote sensing. International Journal of Remote
Sensing, 22, pp. 789–806.
MARKHAM, B.L. and BARKER, J.L., 1987, Thematic mapper bandpass solar exoatmospheric
irradiances. International Journal of Remote Sensing, 8, pp. 517–523.
MEER, F.V.D., 1999, Iterative spectral unmixing. International Journal of Remote Sensing, 20,
pp. 3431–3436.
MESEV, V., GORTE, B. and LONGLEY, P.A., 2001, Modified maximum-likelihood classification
algorithms and their application to urban remote sensing. In Remote Sensing and
Urban Analysis, J. Donnay, M.J. Barnsley and P.A. Longley (Eds), pp. 71–88 (New
York: Taylor and Francis).
MYINT, S.W., 2006, Urban vegetation mapping using sub-pixel analysis and expert system
rules: A critical approach. International Journal of Remote Sensing, 27, pp. 2645–2665.
PHINN, S., STANFORD, M., SCARTH, P., MURRAY, A.T. and SHYY, P.T., 2002, Monitoring the
composition of urban environments based on the vegetation-impervious surface-soil
(VIS) model by sub-pixel analysis techniques. International Journal of Remote Sensing,
23, pp. 4131–4153.
PONTIUS JR, R.G., SHUSAS, E. and MCEACHERN, M., 2004, Detecting important categorical
land change while accounting for persistence. Agriculture, Ecosystems and
Environment, 101, pp. 251–268.
PONTIUS JR, R.G. and CHEUK, M.L., 2006, A generalized cross-tabulation matrix to compare
soft-classified maps at multiple resolutions. International Journal of Geographical
Information Science, 20, pp. 1–30.
4062 J. Tang et al.
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008
RADELOFF, V.C., MLADENOFF, D.J. and BOYCE, M.S., 1999, Detecting jack pine budworm
defoliation using spectral mixture analysis: separating effects from determinants.
Remote Sensing of Environment, 69, pp. 159–169.
RICHARDS, J.A., and JIA, X. (Eds), 1999, Remote Sensing Digital Image Analysis: An
Introduction, pp. 182–183 (Berlin, Germany: Springer-Verlag).
RIDD, M.K., 1995, Exploring a V-I-S (Vegetation–Impervious–Surface–Soil) model for urban
ecosystem analysis through remote sensing: comparative anatomy for cities.
International Journal of Remote Sensing, 16, pp. 2165–2186.
SETTLE, J.J. and DRAKE, N.A., 1993, Linear mixing and the estimation of ground cover
proportions. International Journal of Remote Sensing, 14, pp. 1159–1177.
SHIMABUKURO, Y.E. and SMITH, J.A., 1991, The lease squares mixing models to generate
fraction images derived from remote sensing multispectral data. IEEE Transactions on
Geoscience and Remote Sensing, 29, pp. 16–20.
SMALL, C., 2001, Estimation of urban vegetation abundance by spectral mixture analysis.
International Journal of Remote Sensing, 22, pp. 1305–1334.
SMALL, C., 2003, High spatial resolution spectral mixture analysis of urban reflectance.
Remote Sensing of Environment, 88, pp. 170–186.
SMALL, C., 2004, The Landsat ETM + spectral mixing space. Remote Sensing of Environment,
93, pp. 1–17.
SONG, C., 2005, Spectral mixture analysis for subpixel vegetation fractions in the urban
environment: how to incorporate endmember variability? Remote Sensing of
Environment, 95, pp. 248–263.
THESEIRA, M.A., THOMAS, G. and SANNIER, C.A.D., 2002, An evaluation of spectral mixture
modeling applied to a semi-arid environment. International Journal of Remote
Sensing, 23, pp. 687–700.
TOMPKINS, S., MUSTARD, J.F., PIETERS, C.M. and FORSYTH, D.W., 1997, Optimization of
endmembers for spectral mixture analysis. Remote Sensing of Environment, 59, pp.
472–489.
WANG, F., 1990, Improving remote sensing image analysis through fuzzy information
representation. Photogrammetric Engineering and Remote Sensing, 56, pp. 1163–1169.
WILLMOTT, C.J. and MATSUURA, K., 2006, On the use of dimensioned measures of error to
evaluate the performance of spatial interpolators. International Journal of
Geographical Information Science, 20, pp. 89–102.
WU, C. and MURRAY, A.T., 2003, Estimating impervious surface distribution by spectral
mixture analysis. Remote Sensing of Environment, 84, pp. 493–505.
WU, C.A., 2004, Normalized spectral mixture analysis for monitoring urban composition
using ETM + imagery. Remote Sensing of Environment, 93, pp. 480–492.
YANG, X. and LO, C.P., 2002, Using a time series of satellite imagery to detect land use and
land cover change in the Atlanta, Georgia metropolitan area. International Journal of
Remote Sensing, 23, pp. 1775–1798.
ZADEH, L.A., 1965, Fuzzy sets. Information and Control, 8, pp. 338–353.
ZHANG, J. and FOODY, G.M., 2001, Fully-fuzzy supervised classification of sub-urban land
cover from remotely sensed imagery: statistical and artificial neural network
approaches. International Journal of Remote Sensing, 22, pp. 615–628.
Improving urban classification 4063
Downloaded By: [University at Buffalo (SUNY)] At: 20:50 26 August 2008