(b) is the result of FCM clustering only employing spectral 
features of three bands while figure 8 are the ones that employ 
both spectral features and spatial autocorrelation features, 
  
(a) site 1 
(b) site 2 
Figure 6. Results of FCM clustering segmentation only employing spectral features 
  
i 
(a) d=1 pc (b) 4-2 
   
(a) d=1 (b) d=3 (c) d=5 
(c) d=3 (d) d=4 
Figure 7. Results of FCM clustering segmentation employing both spectral features and spatial autocorrelation features in site 1 
when window width (2d +1) is less than the maximum range 
    
Figure 8. Results of FCM clustering segmentation employing both spectral features and spatial autocorrelation features in site 2 
when window width (2d +1) is less than the maximum range 
Due to spectral variability, the segmentation results in figure 6, 
which only employs spectral features in FCM algorithm, look 
“broken” and include too many noise-like speckles which 
reduce the homogeneity of segments. 
However, FCM clustering segmentation incorporating spatial 
autocorrelation features can obtain more homogeneous 
segments or objects (figure 7 and figure 8). Since window 
parameter d has great effect on calculation of local spatial 
autocorrelation, it also affects the results of clustering 
segmentation. Figure 7 and figure 8 show that as d increases, 
noise-like speckles disappear gradually, and segments are 
becoming more and more homogeneous, but when window 
width (2d+1) approaches the maximum range of all the 
semivariograms characterizing the selected object classes, the 
edges of some small objects begin to become fuzzy and even 
disappear gradually. 
These facts show that the Getis statistic plays the role of a 
low-pass filter and spatial autocorrelation features can 
effectively suppress noise caused by spectral variability in FCM 
clustering segmentation. Therefore, this method can improve 
the quality of FCM clustering segmentation and obtain more 
homogeneous objects. However, there's one point which needs 
attention that improvement of segmentation quality does not 
necessarily mean the improvement of classification accuracy. 
5. CONCLUSION 
This paper focuses on the analysis and modelling of spatial 
autocorrelation features for improving the segmentation quality 
of high resolution satellite images. The semivariograms are used 
to model spatial variability of typical object classes while Getis 
statistic is used to calculate the degree of local spatial 
autocorrelation. Segmentation experiments are conducted via 
the Fuzzy C-Means (FCM) algorithm, which incorporate both 
spatial autocorrelation features and spectral features. The results 
show that spatial autocorrelation features play the role of a 
low-pass filter which can suppress noise caused by spectral 
variability and therefore improve the segmentation quality. 
For future research, we will focus on the determination of 
optimal neighborhood window width within the range of the 
semivariograms by quantitative evaluation on segmentation 
quality or classification accuracy. 
   
This researcl 
Basic Resea 
by National 
and by the 
Universities 
Akcay, H. 
geospatial 0 
IEEE Trans 
pp. 2097-21 
Balaguer, A 
Definition c 
features an 
classificatio! 
Bezdek, J. ( 
Function Al, 
Byun, Y., K 
segmentatio 
modified 
Internation 
Franklin, S 
Automated 
remote sen 
Geoscience. 
Johnson, B. 
evaluation 
Journal o, 
pp.473-483 
Meer, F. 
geostatistic 
pp. 5644-5( 
Myint, S. ' 
spatial m 
comparing 
Engineerin 
On J kK 
statistics: d 
Analysis, 2 
Wulder, M 
characteris 
Getis statis 
pp. 2223-2