Ghosh, Jayanta Kumar 
carried out, the linguistic variables of all the pixels of the preprocessed image are calculated and are stored in files in 
terms of their fuzzy labels. Separate files are maintained for different fuzzy labels of any linguistic variable. The 
inference is started from the top stage of the decision tree classifier and continued till the final stage 
Table 1 Domain knowledge in terms of Linguistic variables, Fuzzy Labels and 
Threshold parameters at different stages of the interpretation. 
STAGES| Cover Class |Linguistic Variables| Fuzzy Labels Parameters 
  
  
  
  
I Water (i) Band ratio Low & High | 0.4, 0.9, 1.40 
vs (Band 4/ Band 3) 
Non-water (ii)Tone(Band 4) | Dark & Light 9, 14, 19 
II Vegetation (i) NDVI High & Low | 0.35, 0.50, 0.65 
MS (ii) Saturation High & Low | 0.3, 0.425, 0.55 
Non-vegetation (ECC) 
III Agriculture Contrast Fine & Coarse 1.0* 
vs 
Non-agriculture 
  
  
  
  
  
  
  
IV Crop (i)Tone(Band 3) | Light & Dark 8* 
vs Gi)Tone(Band 4) | Light & Dark 45* 
Tea 
  
* Crisp categorisation. 
simulating the use of "elimination keys" in visual image interpretation. Interpretation of the full image at each 
stage of the classifier is first completed before proceeding to the next step. The interpretation of the succeeding 
stage involves the output of the preceding stages. This process is similar to sieving operation where further analysis 
of the data is carried out only for those that satisfy the requirements of the previous stage. The interpretation process 
lies in addressing the appropriate rules (depending on the stage of the classifier) and to infer the possible 
information classes of the sample by using fuzzy rules of inference i.e., the Conditional conjunction rule and the 
Conjunction composition rule and to obtain a crisp output defuzzification is carried out by using the Disjunction 
composition rule. 
Subsequently, at the first stage, water cover is interpreted against non-water land cover. At Stage II, the non-water 
category found at Stage I is interpreted to separate vegetative areas from non-vegetative areas. At Stage II, the 
vegetation category found at Stage II is further interpreted to separate out agricultural type vegetation from non- 
agricultural type vegetation. At Stage IV i.e., final stage, the agricultural type vegetation is interpreted to separate 
out crop from tea, leading to mapping of tea gardens from satellite images. Thus, each stage of interpretation, the 
system provides outputs that are useful sources of land cover information. 
4. PERFORMANCE OF THE SYSTEM 
The evaluation of the system for each stage of interpretation is experimented for actual data of the study area as well 
as that of the system for a synthetic image of reference data over and above the sample study of a sub-scene of the 
area. 
4.1 Experimental Results 
The sub-pixel membership values of water cover interpreted by the proposed interpretation system at the first stage 
are compared to those of the expected values. An excellent correlation [0.9504] is found between the two while the 
accuracy for hard classification is found to be 96 percent (Ghosh, 2000a). The correlation between the estimated 
values of the different types of vegetative covers, interpreted at Stage II, in association with different types of non- 
vegetation compared to those of the expected values and are found to be more than 0.95 and hard classification 
shows more than 95 percent accuracy on an average (Ghosh, 2000b). The overall accuracy of spatial analysis of 
vegetation types into agricultural vegetation and non-agricultural vegetation is 86.30 percent. The overall accuracy 
analysis to classify agricultural vegetation into tea and crop is 94.87 percent (Ghosh, 1996b). 
4.2 Performance Evaluation 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 463