41 
method to 
the ability 
In general, 
AR noises 
ndow size 
not easily 
all texture 
lie texture 
textural 
a better 
key issue, 
am model 
ogram is a 
n. Let the 
esented as 
levels is 
0) 
:tween the 
at row x’, 
n model. 
del 
il variable 
ition of a 
rease with 
the spatial 
the testing 
Kt-Nearest 
»SION 
; with the 
on). Land 
ntial area, 
and select 
on system, 
piition by 
:ed, but in 
iracy. 
ee groups 
er-feature 
, angular 
nogeneity 
statistics and measure the uniformity of the non-zero entries in 
the GLCM. The second group contains standard deviation, 
contrast, and dissimilarity, which measure the degree of 
smoothness of the texture. Within each group, features are 
highly correlated. The third group contains only correlation 
statistics, which is an independent measure and not correlated 
with any of the other textual statistics (Table 1). 
HOMO 
CON 
DIS 
SD 
ENT 
ASM 
COR 
HOMO 
1,000000 
-0.475280 
-0.683933 
-0.528559 
-0.977448 
0 851364 
0.216479 
CON 
-0.475280 
1 000000 
0.960253 
0.946297 
0455767 
-0.311692 
-0.205872 
DIS 
-0.683933 
0.960253 
1.000000 
0.9334 30 
0 669171 
-0.478668 
-0.241290 
SD 
-0.528559 
0.946297 
0.933430 
1.000000 
0.534742 
-0.400020 
0.062577 
ENT 
-0.977448 
0455767 
0.669171 
0.534742 
1 000000 
-0 8 27797 
-0.131947 
ASM 
0.851364 
-0.311692 
-0.478668 
-0.400020 
-0 827797 
1 000000 
0.091338 
COR 
0.216479 
-0.205872 
-0.241290 
0 062577 
-0.131947 
0.091338 
1.000000 
Table 1. Correlation matrix of the seven texture features 
Secondly, the selection of the most valuable feature among each 
group is based on the consideration of the stability of the 
texture features. Within the first group and the second group, 
unlike other features, entropy feature image and standard 
deviation texture image always presenting a stable texture under 
a certain statistical window size with the change of the pixel 
pair sampling distance (Figure 2- 5). 
Textural Feature - Stand Dev. 
3 
•a 
> 
c 
2 
WinSize5 
WinSize7 
WmSize9 
—«w— WinSize 13 
—WinSizel7 
—WinSize 21 
—«— WinSize 2 5 
WinSize29 
WinSize35 
WinSize51 
Figure 2. Texture feature stability description 0 of Standard 
Deviation 
Textural Feature - Stand Dev 
- WinSize 5 
WinSize7 
WinSize9 
WinSize 13 
-WmSizel7 
- WinSize 21 
- WinSize 2 5 
- WinSize 2 9 
WinSize 35 
WinSize 51 
Figure 3. Texture feature stability description (D of Standard 
Deviation 
Textural Feature - Stand Dev. 
♦ WinSize 5 
WinSize7 
WinSize 9 
WinSize 13 
—WinSize 17 
-•-WinSize 21 
WinSize25 
WinSize29 
WinSize 35 
WmSize51 
Figure 4. Texture feature stability description (3)of Standard 
Deviation 
4 
<D 
0 
2 
1 
in 
0 
Textural Feature - Stand Dev. 
—WinSize 5 
—WinSize7 
WinSize 9 
WinSize 13 
—•—WinSize 17 
—*— WinSize 21 
WinSize25 
11 13 15 17 19 
5 
- WinSize 2 9 
WinSize35 
WinSize 51 
Figure 5. Texture feature stability description ® of Standard 
Deviation 
After the two-step analysis, correlation, entropy, and standard 
deviation are selected as the most significant features for the 
further analysis. 
3.2 Scale of Ground Objects and Window Size to Analyze 
Scale of ground objects are estimated according to the semi- 
variogram model mentioned above. Figure 6 is an example of 
the scale estimate of a parcel of farmland, which gives the result 
of 5. 
Figure 6. An example of scale estimation of farmland 
Because of the diversity of the ground objects, the majority 
scale of the three land cover types of our research are estimated, 
which are residential 9, vegetation 5, and water body 14. 
Further research on the relationship between object scale and 
processing window size reveals that when the processing 
window size is rather small, the classification accuracy is very 
low; the classification accuracy increases when the window size 
increases, and when it reaches to the size of the objects, there is 
a obvious increase in the classification accuracy; after that,