valid constraints and subtracted otherwise. Optional con- 
straints are represented by having only the supporting 
weight greater than zero so that evidence is added if the 
constraint is valid, otherwise do nothing because the fail- 
ure of an optional constraint does not necessarily refuse 
a category. Contradictory constraints are represented by 
having only the opposing weight greater than zero so 
that evidence is subtracted if the constraints is invalid, 
otherwise do nothing because the success of a contra- 
dictory constraint does not neccessarily support a cate- 
gory. 
We have established the knowledge sub-bases for 8 cat- 
egoriesforest, grass, new/old buildings, clear/turbid water, 
soil and cropland in SPOT/TM image, as well as a pro- 
gramme for increasing, deleting and modifying the know- 
ledge base interactively. 
CONTROL STRATEGY AND DECISION MAKING 
In knowledge base, each category has its independent 
sub-base represented by a set of rules and weights. If on- 
ly one constraint in a subbase is valid for a 
segmentation region, we cannot conclude the region is 
the category corresponding to the sub-base, because the 
knowledge sub-base of that category consists of several 
rules and weights, and they are a whole. Neither is the 
idea is not preferable that only when all the constraints 
in a sub-base are valid for a segmentation region can 
we conclude the region is that category. Because the 
base has been established in ideal conditions, there are al- 
ways some differences between ideal conditions and reali- 
ties. For the above reasons, we present control strategies 
as follows: 
1. Compute the supporting and opposing evidence 
amounts of candidate category using all sub-base for 
each segmentation region. 
2. Compute category scores with a function of the rel- 
ative proportion of supporting and opposing evidence. 
3. Make decisions based on the scores achieved. 
Equation (3) and (4) are used to compute category 
scores. Equation (3) is used if the supporting evidence is 
greater than the opposing evidence, otherwise (4) is used. 
score = 100(1 —Eopp/Esup) (3) 
score = 100(1—Esup/Eopp) (4) 
where Esup is the amount of supporting evidence and 
Eopp is the amount of opposing evidence. 
As a result, a category score of a region is com- 
puted for each sub-base. All the category scores of a re- 
gion are ordered based on their values. If the maximum 
score is smaller than a certain threshold(e.g.,20), and if 
the region doesn't belong in a corresponding category, 
then it is represented by category 0 (i.e., to refuse 
recognition). If the maximum score is bigger than the 
threshold and the difference between the first and se- 
89 
cond maximum is smaller than a certain threshold(e.g., 
5), the region is considered unidentifiable. Then, man 
interference is necessary. Otherwise, the segmentation re- 
gion is considered as the category corresponding to the 
maximum Score. 
EXPERIMENTAL RESULTS AND CONCLUSIONS 
The method presented in this paper has been evalu- 
ated by the 
S12-by-512 pixel SPOT image. The main topographical 
testing its classification accuracy for 
factors in the image are settlement place, water, soil and 
forest. Firstly, the image is divided into 21 classifications 
as segmented image by K-means algorithm. Secondly, sev- 
en categories such as new/old settlement places, forest, 
clear/turbid water and cropland are extracted based on 
spectral and textural knowledge. Finally, new/old settle- 
ment place and clear/turbid water are merged into settle- 
ment place and water respectively. À part of experi- 
mental results is shown in Fig.2. The classification accu- 
racies of the presented method and the K-means 
  
   
  
  
  
  
  
  
  
  
  
  
  
lgorith are shown in table 1. 
m = ~ 2 K means | knowledge | 
| based technique | 
| classif cation >. | 
forest confusion with crops 94 | 
new settlement | 64 86 | 
| old settlement | 73 | 88 | 
| clear water | 83 | 90 | 
turbid water 81 | 87 | 
| cropland(crops) | confusion with forest | 84 | 
| ther | 7 | 82 | 
| 
  
Table 1. Accuracy comparison between knowlege-based 
technique presented in this paper and k-means 
Table 1 shows that K-means algorithm has a low 
classification accuracy. The one reason is that the sand 
content of water changes greatly in the region covered 
by the image. Of the 7 categories in the image water 
even occupies 3, the rest 4 categories are found con- 
fused so heavily that they are unidentifiable. The solu- 
tion of the problem is to increase the amount of 
classifications. But the amount of work is too great for 
man to interpret and merge the classifications, For this 
reason, such a large number(e.g.21) of classifications are 
segmented by knowledge-based classification method in 
advance. After that, with a little man interference, the 
knowlege base is used to discriminate the above 
classifications (e. g. , segmentation regions) automatically. 
Because water and forest have typical reflectance curves, 
and the textural properties of settlement place are very 
different from those of the others, they can be 
discriminated with a high accuracy by a knowledge-based