Ghosh, Jayanta Kumar 
  
The linear water body at the top of the sub-scene is not present at all in the output of the minimum distance classifier 
(Plate 3). Whereas, the presence of the linear water body at the top of the image can be authenticated from FCC 
(Plate 1). The FCC also depicts that the considered linear water body is not as prominent as pure water bodies 
present in other part of the image. Thus, sub-pixel analysis by the proposed system provides some vital information 
regarding the information class, which is not possible through standard statistical methods. Other important factor is 
that the graded water bodies also comes into consideration for calculation of area under water in case of proposed 
image interpretation system but not in minimum distance to mean classifier and thus, improving the accuracy of area 
estimation by the proposed system. 
In classification of vegetation, there is a marginal difference (about 48 pixels i.e., about 0.296) between the number 
of pixels classified by the minimum distance to means classifier and that interpreted by the system as pure class but 
there is a significant difference (4,482 pixels i.e., about 2596) with the number interpreted as hard class. This 
discrepancy is transmitted to non-water non-vegetation category. Interestingly, the difference (4,943 pixels) 
between the number of non-vegetation under hard class interpreted by the system and that by minimum distance to 
means classifier is more or less same as the difference (4,842 pixels) between the number of vegetation under hard 
class and that by minimum distance to mean classifier. This clearly signifies that most of the sub-pixel vegetation of - 
hard class has been classified as non-vegetation by the minimum distance to means classifier. Also, the non- 
vegetation hard class pixels of the system are mostly interpreted by the minimum distance to means classifier also as 
non-vegetation. 
Thus, it can be concluded that enhancement in classified information is depicted in the classification of the 
proposed system of image interpretation by accounting sub-pixel analysis. The sub-pixel analysis by the 
proposed system provides information regarding the pure information classes, the pixels under graded 
category, the extent of gradation, detail information about areas under different categories. In case of interpretation 
of different vegetation types, it is found that there is a discrepancy in the numbers of pixels as interpreted by the 
system and that by minimum distance to means classifier. 
The spectral similarity between non-agricultural vegetation (specifically, scrub vegetation a component of non- 
agricultural vegetation) and tea may be the cause of large commission of non-agricultural vegetation to tea class in 
the classification of the minimum distance to means classifier. On the other hand, the inter mingling of the 
information classes causes coarser spatial texture at the boundary of tea gardens may result in some omission error 
in tea category to non-agricultural vegetation in the classification of the proposed interpretation system. These 
results in significant difference in the classification of the two approaches. However, by comparing the classification 
of the interpretation system (Plate 2) with the FCC of the sub-scene (Plate 1), it is found that the mapping of tea 
gardens of the sub-scene is done quite satisfactorily by the proposed system. 
S. DISCUSSION 
It is observed that the adopted model performs excellently in calculating mixed pixel contents. Thus, sub-pixel 
analysis of water and vegetation from satellite images can be done reliably by the proposed method. From the 
experimental result, it can be assured that the discrepancy between the expected and the estimated values of the 
partial membership in information classes will be maximum 15 percent i.e., estimated values are within 15 percent 
higher or lower than expected values. The mis-classification, if any, in finding the hard classes will arise in pixels 
having 40 to 60 percent component cover classes. The experimental results show that to estimate area under 
different land cover the proposed method will give good result. The model also performs well to classify agricultural 
vegetation from non-agricultural vegetation and that crop from tea. From the sample study, it can be concluded that 
the proposed method is a viable system for mapping of tea gardens from satellite images. 
6. CONCLUSION 
In this research work, a novel method has been studied for developing a humanist system for automated mapping 
of tea gardens from satellite image. Integration of domain knowledge and photo interpreters’ heuristics with the 
techniques of image processing, artificial intelligence and pattern recognition have been tried. The results show that 
the adopted method can be applied for mapping of tea gardens of the study area from the considered data. The 
system shows excellent performance in sub-pixel classification of land covers. It provides a huge enhancement of 
land cover information in comparison to the existing classification schemes. The sub-pixel interpretation of land 
covers also improves the accuracy of estimation of area under different land covers. However, the ultimate aim of 
developing a humanistic image interpretation system is to have a viable and versatile alternative to expert image 
  
  
466 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part B7. Amsterdam 2000.