Angular Second Moment (ASM): 
YLpüjÿ 
i i 
(3) 
Standard Deviation (SD): 
Variance = £ Z (*>’)/»(*. J) ~ P x M y ^ 
SD=SQRT(Variance) (5) 
Contrast (CON): 
'ZÎ.Q-iïpff.j) <6) 
> i 
Dissimilarity (DIS): 
YL\i-j\Püj) 
(7) 
Entropy (ENT): 
“Z Z ^’7) l°ë,(p(hj)) 
i j 
(8) 
Correlation (COR): 
Z Z (ü)p(^j)-p x p 
y 
(9) 
2.2 Parameters Set and Selection 
In order to properly use the co-occurrence method for image 
texture analysis, several variables should be considered: the 
number of quantization levels, the number and type of 
measurements, the window size to analyze, the pixel pair 
sampling distance and orientations. 
For quantization levels, Soh and Tsatsoulis (1999) indicate that 
a complete grey-level representation of the image is not 
necessary for texture mapping, an eight-level quantization 
representation is undesirable for textural representation, and 
setting quantization levels to 64 is efficient and sufficient. 
Clausi (2002) recommends setting quantization level to 24. 
However, larger values of quantization level greater than 64 are 
deemed excessive. Here in our research, we use 64 as the 
quantization level. 
Because there is no obvious structure characteristics of SAR 
imagery, and it presents isotropic behaviour, directional 
invariant texture measures which are the averages among 
texture measures for 4 directions (0°, 45°, 90°, and 135°) are 
advocated. 
Window size and pixel pair sampling distance are key two 
parameters that affect the extracted texture feature from GLCM. 
It reaches different conclusions from different research. These 
two factors will be studied detailedly in this research. 
remote sensing imagery. When we apply the GLCP method to 
describe texture, window size to process determines the ability 
to capture texture features at different spatial extents. In general, 
smaller window size could be easily influenced by SAR noises 
while it can describe small texture feature; lager window size 
could describe the whole scenery better and not easily 
influenced by SAR noises but cannot describe small texture 
features. This comes our idea to integrate multi-scale texture 
features in order to accurately capture the textural 
characteristics of a given surface and reach a better 
classification performance. 
In the multi-scale analysis, the selection of scale is a key issue. 
This paper presents a method that using semi-variogram model 
to estimate the scale of the ground objects. Semi-variogram is a 
feature which is based on the spatial autocorrelation. Let the 
grey levels comprising a given digital image be represented as 
G(x,y). Then, the semi-variogram for these grey levels is 
defined as: 
rW=^i,Mx,y)-G(x,y)] 2 OO) 
where, h is the Euclidean distance (lag distance) between the 
pixel value G at row x, column y , and the pixel value at row x’, 
column y\ Figure 1 is an example of a semi-variogram model. 
Figure 1. An example of a semi-variogram model 
This method is based on the assumption that a regional variable 
becomes random at a great distance, i.e. the correlation of a 
regional variable with other pixels will normally decrease with 
increasing distance. Thus we use it here for estimate the spatial 
structure of the image. 
Non-parametric classifier, Fuzzy K-means, is used as the testing 
classifier, and non-parametric classifier, K-Nearest 
Neighbourhood is used as the final classifier. 
3. EXPERIMENTAL RESULTS AND DISCUSSION 
Image dataset used here is Radarsatl fine beam mode with the 
spatial resolution 6.25m (C-band with HH polarization). Land 
use/land cover types concerned are vegetation, residential area, 
and water body. 
3.1 Texture Feature Selection 
The last consideration is the number and type of measurements 
taken from the co-occurrence matrices. In our research, 7 
texture measures mentioned above will be processed and 
analyzed. 
2.3 Multi-scale Texture Analysis 
Different ground object has different scale, and this scale can 
also be represented by the texture primitive element in the 
Feature selection is to take a set of candidate features and select 
a subset that performs the best under some classification system. 
This procedure can reduce not only the cost of recognition by 
reducing the number of features that need to be collected, but in 
some cases it can also provide better classification accuracy. 
Firstly, the texture features can be classified into three groups 
according to the structure they reveal and their inter-feature 
correlations. The first group contains homogeneity, angular 
second moment, and entropy, which are the homogeneity 
statistics and measu; 
the GLCM. The s< 
contrast, and dissi 
smoothness of the 
highly correlated. 1 
statistics, which is £ 
with any of the othei 
HOMO 
CON 
HOMO 
1.000000 
-0.475 
CON 
-0.475280 
1.000 
DIS 1 
-0.683933 
0.960 
SD 
-0.528559 
0 946 
ENT 
-0.977448 
0455 
ASM 
0 851364 
-0 311 
COR 
0.216479 
-0 205 
Table 1. Correia 
Secondly, the selecti 
group is based on 
texture features. Wi 
unlike other featur 
deviation texture im; 
a certain statistical 
pair sampling distan 
Figure 2. Texture 1