■ 
190 
extraction and the classification, first must analyze the 
information characteristic earnestly.The grey level curve of 
different ground features on different wave bands is as 
following figure 2. 
Figure 2. The grey level curve of different ground features on 
different wave bands 
The texture also is the remote sensing image important 
information, which not only has reflected the image grey level 
statistics information, but also has reflected the ground feature 
itself structure characteristic and the spatial arrangement 
relations, so it also is one of the visual interpretation important 
symbols[8]. Many research indicated that, it can enhance the 
interpretation precision to overlay the texture information on 
the primary image spectrum information. In the remote sensing 
image, the texture mostly is random, which obediences 
statistical distribution, it often described by Grey level Co- 
occurrent Matrix. Grey level Co-occurrent Matrix is a matrix 
which is constituted by level-two union conditional probability 
density between the image grey levels, it has reflected the 
correlation between two random points grey level in the image. 
In this foundation, we had defined eight texture characteristic 
statistical value: Mean, Variance, Homogeneity, Contrast, 
Dissimilarity, Entropy, Second Moment,all the value described 
the image texture characteristic from many aspects. 1[9J Using 
Grey level Co-occurrent Matrix to carry on the texture analysis 
involves three Important parameters: motion window size, 
motion step and motion direction. The appropriate window size 
is especially important to the texture analysis.In general, the 
window size is decided by the primary image texture structure, 
the small window denotes the slight texture characteristic while 
the big denotes the rough. The choice of the motion step is also 
decided by the image texture granularity, the short step is 
suitable for the slight texture while the long is suitable for the 
rough.Many researches pointed out taking the step as one is 
quite effective to all the different texture. In studies,we usually 
take four main direction values of 0 ,45 ,90 ,135 
direction or the mean of these four direction values. 11101 However, 
Franklin and the Pebble pointed out the texture characteristic of 
single direction was better than the mean of these four 
directions. [11] Through analyzing and comparing different 
window sizes and steps, the study selected the window whose 
size was 5x5 and the step was one. 
In addition, because of the region natural condition control and 
the human factor intervention, the spatial distribution of the 
ground features often have some region differentiation 
rules.Therefore, this study also took the geographic coordinates 
and the terrain factor as the forecast variables. 
In summary, the study used 18 following testing variables: two 
geographic coordinates (X, Y), six wave band grey levies of 
TM image(l,2,3,4,5,7 bands)(the 6th wave band exception), 
seven texture characteristics (Mean, Variance, Homogeneity, 
Contrast, Dissimilarity, Entropy, Second Moment (SM)and 
three terrain factors (DEM, Slope, Aspect) . Among these, the 
slope was separate variable whose value is 0,1,...,7, and the 
other were continual variables.The goal variablesof this study 
was: Water body, Paddy field, Arid land, Building area, 
Road,Vegetation and Subsidenceland. 
4.3 Classifying based on CART 
In the study ,we first chose 3,200 sample collections for these 
above-mentioned testing variables and target variables, then 
make use of CART to analyze and study the samples,finally,we 
structured a decision tree who had 38 leave nodes. Its studying 
accuracy is 92.8% and verification accuracy is 90.3%. The 
structure of the decision tree can be expressed into the If- 
Then[CF] form very conveniently. The testing route from the 
tree's father node to every leaf node corresponded to a rule , so 
there were 38 rules in all. For examples: 
If (19.688< Mean<=19.958 && 10.000<TM6<=45.000 && 
13<TM5<=71.000) Then (class = 1) CF=0.987 
If (42.333<Mean<=44.500 && 40.000<TM6<=44.00 && 
42.000<TM3<~47.000 && 4578510<X<4578522) Then (class 
= 7) CF=0.854 
In the If-Then[CF] form, CF indicates the confidence measure 
of the rule,and its value region of CF is [0 , 1]. If the value is 0, 
then the possibility of the present pixel belongs to the given 
class out is also zero; while the value is 1, then the confidence 
value is invariable. Making use of the above-mentioned rules 
and the following simple matching strategies, we gained the 
classified image such as figure 3(a). 
1. When only satisfied some one rule, then took the 
output category of the rule as the classified category. 
2. When simultaneously satisfied multi-rules, then took 
the output category of the rule whose confidence 
value was more as the classified category. 
3. When didn’t satisfied all the rules, then it was not 
classified. 
In order to compare with other classification methods, the study 
also had used supervised classification to classify and test, the 
classified image such as figure 3(b). 
Water body 
Paddy field 
Arid land 
Building area 
Road 
Vegetation 
Subsidenceland 
Figure 3(a). The output image classified by CART