in the example, the test imag was divided into 3 kinds of 
features: cropland, vegetable fields and housing estate by 
visual interpretation. 3 kinds of samples were selected from 
classified area,30 samples were selected for each kind. The 
size of each sample is 80 X 80 pixels. During the classifying 
process, whether the template size is too big or too samll, the 
result image will be affected negatively, so the size is always 
between 8X8 and 20 X 20. in this paper, 10X10 was selected 
as the template. 
The characteristics selected for classifying should 
differentiate different feature primely. In order to select good 
characters we have compared the test result, so in this paper 
contrast, entropy and variance of histogram characteristics 
were selected for classifying. 
The average of the 3 characteristics of cropland, vegetable 
fields and housing estate were calculated as the clustering 
center, the average is listed in the table belows. 
samples/ characteristics 
contrast 
entropy 
Variance 
Vegetable fields 
1.7290 
3.7256 
4.0796 
Housing estate 
5.0049 
6.3066 
9.8611 
cropland 
0.3317 
0.4124 
0.5715 
Table 1. average of characteristics of samples 
2) Initializing: m is always between 1.5 and 5, if m<2, the 
classification inclined to hard clustering, if m>3, the 
classification is too smooth, so in this paper, 2.5 is 
initialized to m. 
3) Initializing membership array £/ <0) . ¿/0 is calculated 
by formula (3), then membership array £/ (0) is 
recalculated by and formula (5). 
4) Image classification by the fuzzy classifier. Based on 
those idea, program was finished to realized 
classification based on Matlab. 
5) Result image after classification was achieved by the rule 
of maximum membership using fuzzy classification. In 
order to remove the noise exit on the classified image, 
median filter was adopted to the result image.The 
original image and the classified image after filtering is 
shows as figure 2 and figure 3. 
Figure 2. original image 
Figure 3. image classified by improved fuzzy c-mean 
classifier. 
Figure 4. image classified by ERDAS 
The test was processed based on unsupervised classificaiton, 
the classified image is only the dividing of three kinds of 
features. The three kinds of features was ascertained through 
priori knowledge after fuzzy clustering.In figure 3, he green is 
vegetable fields, the blue is housing estate, the red is cropland. 
We can see from image, different kinds of features were 
recognized correctly, and the boundary of them is very clear. 
4 ACCURACY EVALUATION FOR CLASSIFIED 
IAMGE 
Many standards were proposed to assess the accuracy for 
classified image, among them, confusion matrix is accepted 
widly, so confusion matrix was introduced in precision 
evaluation in the paper. 3 indexs were calculated, The overall 
accuracy represents the probability of consistence between 
classified image and real field of classified feature types. The 
producer’s accuracy represent the conditional probability of 
consistence between a random selected point from real field 
and correspondence one from classified image. The user’s 
accuracy represents the conditional probability of consistence 
between a random sample point selected from classified 
image and correspondence point in real field. We selected 
1,000 points randomly on the classified image, and compared 
with the priori knowledge, then the statistical table was 
achieved, it shows as tabel 2.