technique. 
Our programmes were designed in C language, and 
ran on MC-6700. It took 12 minutes to segment the 
512-by-512 pixel image into 21 classifications using 
K-means algorithm. Based on that, it took 6 minutes to 
merge 21 classifications into 7 categories using the spec- 
tral and textural knowledge base. 
The results above support the following conclusions: 
1. It is feasible to segment an image first using an 
unsupervised classification algorithm, then to discriminate 
the category of every segmentation region using a know- 
ledge base. It is of fast computation speed and high 
classification accuracy. 
2. The method presented in this paper not only can 
segment an image into regions of same properties, but al- 
so can discriminate the categories of the regions 
automatically. 
REFERENCES 
Ferrante, RD. M.J. Carlutto, J. Pomaraede, and P.W. 
  
(a) Part of a SPOT image, band 2 
  
(d) 7 classifications such as forest, 
new/old settlement place, clear/turbid 
water, cropland extracted 
by the knowledge based technique. 
  
(b) Textural energy image. 
  
(e) The image classified into 7 
Bain, 1984. Multispectral image analysis system. In Proc. 
1st. Conf. Artificial Intelligence Applications (Denver, 
Co), pp. 357-363. 
Landgrebe, D.A.,1981. Analysis technology for land re- 
mote sensing. Proc. IEEE, Vol. 69, No. 5, pp.628-642. 
Mehldau, G., and R. A. Schowengerdt, 1990. A 
C-extension for rule-based image classification systems. 
Photogrammetric Engineering and Remote Sensing, Vol. 
56, No. 6, pp.887-892. 
Wechsler, H., and T. Citron, 1980. Feature extraction 
for texture classification. Pattern Recognition, IZ, pp. 
301-311. 
Wharton, S. W., 1987. A spectral-knowledge-based ap- 
proach for urban land-cover discrimination. In proc. 
IEEE, Vol. GE-25, No. 3, pp.272-282. 
Zhang Xiangqian, Li Xiaofeng and Yu Xuchu, 1991. 
A rule-based method for contrast enhancement. Proc. of 
The Chinese Joint Conference on Pattern Recognition 
and Artificial Intelligence, Harbin, China, pp.69-72. 
  
(c) The image classified into 21 
classifications by k — means. 
  
(f) Merging the 7 classifitions of Fig. 
classifications by k — means. 2(d) into 4 categories such as 
drttlement place, forest, water area 
and cropland, 
Fig.2 A Part of experimental results.