Ghosh, Jayanta Kumar 
  
To carry out the performance evaluation of the developed system, 256 reference samples of nine broad categories of 
the land covers present in the study scene are collected from VDU near the training fields and supported by ground 
reconnaissance. They are arranged in rows and columns to form an image of 16 X 16 in a way as the study scene 
appears. Now the interpretation of the image is carried out by the developed system and classification by the 
minimum distance to means classifier after using the same set of training samples for both. The output of the 
proposed system of interpretation and that of statistical classification algorithms are compared with respect to the 
reference data and resulted in the error matrices (Ghosh, 1996b). The commission of tea to non-agricultural 
vegetation is more than expected in the output of the developed system for image interpretation. This is because of 
artificial arrangement of samples in the test image. The non-uniformity in the transition from tea to other classes in 
the test image results from coarse spatial contrast and thus, commission to non-agricultural vegetation. It is found 
that the overall classification accuracy of the proposed system is about 88 percent and that statistical minimum 
distance to means classifier is about 77 percent. Thus, the proposed system gives much better result than minimum 
distance to means algorithm. It is expected that the commission of tea to non-agricultural vegetation category will be 
much less in case original scene and thus will result in further improvement in overall accuracy by the system 
4.3. Sample study 
A sub-scene consists of 165 rows and having 165 pixels in each row is taken to evaluate the performance of the 
proposed system on satellite image. The FCC of IRS LISS II image of the sub-scene is shown in Plate 1. The 
mapping of tea garden from satellite image by the proposed image interpretation system is shown in Plate 2. The 
output of the sub-scene, classified by the minimum distance to means classification algorithm, is shown in Plate 3. 
The results of histogram analysis of the land cover classes are given in Table 4. However, a comparison between the 
two methods has been carried out. It is found that the numbers of pixels in the pure class of water, interpreted by 
proposed system of image interpretation is consistent with the number of pixels under water by the minimum 
distance to means classifier. The proposed system also provides additional information regarding the number of 
pixels under graded category. It can be noted that the sub-pixels (having graded membership more than 0.5 and less 
than 1.0) are interpreted by the system as water which has been mostly classified as non-water by minimum distance 
classifier. This is a significant enhancement regarding the information content in the output of the proposed system 
of image interpretation. The effect of this enhancement of information content can be visualised more prominently 
by considering the Stage I output of the proposed system of image interpretation and that from minimum distance to 
mean classifier (Ghosh, 20002). 
Table 4 Comparison of the Results of outputs of the sub-scene 
  
  
  
  
  
  
  
  
  
  
  
  
  
Land Proposed System of Image Minimum Distance to Means 
Cover Interpretation Classifier 
| Types | Number of Pixels | Area | Number of Area | 
(Sq. Km) Pixels (Sq. Km) 
| | Pure | Hard class | | | | 
| Wa | 24 | 37 | 0468 | 236 | 031 |] 
ENV 26.665 | 2684 | Ba 77T 726987 17350 «id 
|. vede I xol 3-9 [ i1 [| 3I | 
| wv sana | 205 D as T 7936 7 pq | 
INA cds aui TSP solo I 150574] 
| Crop | 1546 7 pocos qq IS) 
|[- 1e ]..12413..] (15920 >| ise T, Bet 
  
Legend: WAT - Water; NV - non-water non-vegetation; NA - non-agricultural 
Vegetation; CD - classified data; RD - reference data. 
  
464 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.