, clustering Segmentation 
cult to obtain satisfactory 
1al autocorrelation feature, 
orporated into clustering 
> segmentation quality, 
nts are conducted via the 
which incorporates both 
| spectral features. The 
ocorrelation features can 
quality of high resolution 
CUSSIONS 
object classes (figure 3) 
re 1(a)): vegetation, water 
ects like ships are not 
ivariograms of theirs are 
(figure 4). 
    
vater (c) road 
classes in site 1 
ect classes correspond to 
ey have different spatial 
and vegetation are quite 
and their ranges are all 
pection. Comparing with 
of the road is relatively 
nt and unstable when lag 
ads have more complex 
semivariogram is about 
information of the three 
[. 
<< road 
"water 
+ vegetation 
| 14 16 18 20 2 2 
gh) 
band 2 
t bands 
Vegetation | Bare land 
— 
2~4 
2~4 
sses in site 2 
  
  
  
ection of neighborhood 
    
windows of spatial autocorrelation analysis, that is, window 
width (2d+1) is less than or equal to the maximum range of all 
the semivariograms characterizing all selected object classes. 
For detailed spatial variation structures or patterns, we have to 
study semivariance function models, which are not concerned in 
this paper. 
With the same procedures, we select five typical object classes 
in the image of site 2: building 1, building 2, shadow, vegetation 
and bare land, and by their semivariograms, we obtain their 
corresponding range information listed in table 2. 
42 Spatial autocorrelation analysis 
In this section, spatial autocorrelation analysis is made using 
Getis statistic in three bands of the original image of site 1, 
nu oF 
05 0. 23 02 0 01.0 02 03 
(e) band 3, d=1 (f) band 3, d=2 
  
  
-02 0.1 0 0.1 n2 
(j) band 2, 4-2 
-02 -016 -01 005 D 05 0.1 05 02 923 
(1) band 2, d-1 
respectively. In experiments, we take different values for d, 
which do not make the window width (2d-1) exceed the 
maximum range of the typical object classes. By table 1, the 
maximum range is within 8-9 pixels, so the proper value for d 
is 1, 2, 3 and 4. For each spectral band, its spatial 
autocorrelation bands calculated by different parameter ds are 
visualized as images (figure 5). 
Figure 5 show that for each spectral band of original image, 
typical object classes (vegetation, water and road) in their 
spatial autocorrelation images are visible, and by the color bar, 
red regions and blue regions in autocorrelation images represent 
higher and lower autocorrelation degree, which correspond to 
“hot spots” and “cold spots” in original image, respectively. The 
autocorrelation images include most structure information of the 
object classes and filter out some small detailed information. 
ET 01 8 01! 02 0 = i 02 0 02 04 
(c) band 4, d=3 (d) band 4, d=4 
    
     
       
    
42 0 0.2 04 06 
(h) band 3, d+4 
  
$t n2 43 04 RE 06 04 08 
(g) band 3, d-3 
  
      
  
£4 03 02 Of 90 01 02 03 04 05 08 0.4 02 0 02 
(k) band 2, d=3 (1) band 2, d=4 
Figure 5. Spatial autocorrelation images of different spectral bands with different window parameter ds 
43 FCM clustering segmentation incorporating spatial 
autocorrelation features 
In this section, two segmentation experiments are conducted via 
the Fuzzy C-Means (FCM) algorithm, which employ the images 
of site 1 and site 2, respectively. The feature vectors for 
clustering segmentation consist of spectral features (three bands) 
and spatial autocorrelation features (also three bands). For 
comparison, the segmentation experiments only employing 
spectral features are also conducted. 
In the first experiment, the image of site 1 is used for clustering 
segmentation. As the experiment focuses on the segmentation of 
three typical object classes (vegetation, water and road), cluster 
number is set as C=4. When iterating 30 times, segmentation 
results are shown in figure 6 (a) and figure 7, where figure 6(a) 
is the result of FCM clustering only employing spectral features 
of three bands while figure 7 are the ones that employ both 
spectral features and spatial autocorrelation features (three 
spectral bands and their spatial autocorrelation bands). 
The second experiment uses the image of site 2 and for the 
segmentation of five typical object classes, cluster number is set 
as C=6. When iterating 50 times, segmentation results of FCM 
clustering are shown in figure 6 (b) and figure 8, where figure 6