Ghosh, Jayanta Kumar 
  
The system is developed for interpreting geocoded satellite data. By this, it pre-supposes that the data to be 
interpreted is free from instrumentation error and geometric distortion. Thus the preprocessing module of the system 
is developed to take care errors due to atmospheric effect by dark object subtraction technique (Chavez, 1988). 
2. 4 Characteristic Functions 
In order to address the linguistic variables in terms of their fuzzy labels, the system using characteristic functions 
fuzzifies the input features. For Stage I and Stage II, standard S-function (Zadeh, 1975) or an inverse of it is used to 
define the characteristic functions. The parameterizations of these are carried out by Modified Fuzzy Threshold 
Technique (Ghosh, 1996b) where entropy, as a measure of uncertainty, is used as a criteria to find the threshold 
parameters. For Stage III and Stage IV, the threshold parameters are found by amplitude thresholding. 
2. 5 Classifier 
The basic philosophy adopted in designing the classifier is to divide the variability factors present in a scene into 
categories, which are related to the information desired and those that are not. The expert analyst's heuristics and 
domain knowledge about the land covers’ implicit hierarchical structure has been primarily utilized for framing the 
classifier by Manual Design Procedure (Swain & Hauska, 1977). An expert analyst generally adopts a multi-stage 
classification scheme for interpretation of a satellite image involving multi-feature at each of classification. Thus, a 
binary decision tree classifier is designed for mapping of tea gardens from satellite images as shown in figure 2. 
  
INFORMATION 
CLASSES OF THE SCENE 
| 
| | 
  
  
  
  
  
  
  
  
  
  
WATER | | NON-WATER x] 
| | 
| VEGETATION | | NON-VEGETATION 
| AGRICULTURE | | NON-AGRICULTURE | 
  
Figure 2 Binary Decision Tree Classifier 
2. 6 Knowledge Base 
The knowledge base consists of facts and principles accumulated from the spectral and spatial domain and from 
training data. It also contains most importantly the heuristics of satellite image interpreter. From the available 
features, subsets of features compatible to the stages of the decision tree classifier are decided by using hierarchical 
search. The domain knowledge in terms of linguistic variables and their fuzzy labels and parameters of the 
characteristic functions at different stages of the decision tree classifier used in the system is given in a summarized 
form in Table 1. 
2. 7 Inference mechanism 
The inference mechanism of the system involves composition rules of atomic fuzzy propositions. Suppose F and G 
is two fuzzy subsets of the universes U and V respectively. Let us consider, two atomic fuzzy propositions X is F 
and Y is G which can also be represented by the possibility distributions x, and x, respectively, where 7, —F and m, 
=G. Then according to Zadeh (1979, 1981), the composition of the propositions X is F and Y is G can be viewed as 
a particularization of the possibility distribution x, by the possibility distribution x,. In order to evaluate the 
particularization due to compound fuzzy proposition, the translation rules of fuzzy calculus developed by Zadeh 
(1979, 1981) have been used in the proposed system. 
3. WORKING OF THE SYSTEM 
The multi-stage classifier involving multi-features is used for inferencing in the system for image interpretation 
emulating an expert image analyst. The working of the proposed image interpretation system thus proceeds from top 
to bottom of the classifier i.e., from general to a particular type of land cover class. Before any classification is 
  
462 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.