ood 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Reference classes Reference classes 
Class Water | Field | Forest | Urban | Open | Total Class Water | Field | Forest | Urban | Open | Total 
Water 7 0 0 0 0 7 Water 7 0 0 0 0 7 
Field 2 52 6 2 3 65 Field 2 53 1 0 3 39 
Forest 0 1 9 0 0 10 Forest 0 0 14 1 0 15 
Urban 2 7 1 13 5 28 Urban 2 7 1 14 5 29 
Open 0 1 0 2 17 20 Open 0 1 0 2 17 20 
Total 11 61 16 17 25 130 Total 11 61 16 17 25 130 
  
  
  
  
  
  
  
  
  
  
Table 5. Confusion matrix of the segment-based Maximum 
Likelihood classification. 
4.2 Rule-based postclassification 
Table 6 presents the confusion matrix of the rule-based 
classification when also this classification was based on 
segments. When the rule-based classification was pixel-based, 
the confusion matrix presented in Table 7 was obtained. The 
result of the pixel-based classification is presented in Figure 5. 
It should be noted that segments had been interpreted in the 
Maximum Likelihood classification stage also in this case. 
  
  
  
  
  
  
  
  
Reference classes 
Class Water | Field | Forest | Urban | Open | Total 
Water 7 0 0 0 0 7 
Field 2 52 2 1 3 60 
Forest 0 1 13 1 0 15 
Urban 2 7 1 13 5 28 
Open 0 1 0 2 17 20 
Total 11 61 16 17 25 130 
  
  
  
  
  
  
  
  
  
Table 6. Confusion matrix of the segment-based 
postclassification. 
  
Forest 
  
Figure 5. Result of the pixel-based postclassification. 
Table 7. Confusion matrix of the pixel-based postclassification. 
The mean accuracy and the total accuracy of the classifications 
are presented in Table 8. In addition to the classifications 
discussed above, a pixel-based Maximum Likelihood 
classification was made to allow evaluation of usefulness of the 
segmentation as a preprocessing operation. 
  
  
  
  
  
  
  
  
  
Classifications 
Class 1 2 3 4 
Water 47% 78 9o 78 % 78 % 
Field 78 % 83 % 86 % 88 % 
Forest 69 % 69 % 84 % 90 % 
Urban 32% 58 % 58 % 61 % 
Open 72 % 76 % 76 % 76 % 
Total 67 % 75 % 78 % 81 % 
  
  
  
  
  
  
Table 8. The mean accuracy and total accuracy of different 
classifications. Column 1 is the pixel-based ML-classification, 
column 2 is the segment-based ML-classification, column 3 is 
the segment-based ML-classification followed by the segment- 
based postclassification and column 4 is the segment-based 
ML-classification followed by the pixel-based 
postclassification. 
5. DISCUSSION AND CONCLUSIONS 
The rule-based postclassification improved the interpretation 
results, especially when it was pixel-based. The biggest changes 
in the postclassification occurred in the hilly areas where some 
forests had been misclassified as field in the preclassification, 
and on the other hand in the flat area where some fields had 
been misclassified as forest. Some of the forest and field areas 
are very similar in the spectral space which causes the errors in 
the Maximum Likelihood classification. For interpretation of 
these classes, the old land use data and height data are very 
useful. It improved the result both in the pixel-based and in the 
segment-based postclassification. 
A very remarkable change in the pixel-based postclassification 
was that the numerous narrow canals of the old land use map, 
which were missing in the segment-based Maximum 
Likelihood classification, appeared in the results. This change 
cannot be seen in the error matrices because of the small 
number of reference points available for the accuracy 
evaluation. When the postclassification was segment-based, the 
canals were also missing in the final result because of errors in 
the segmentation stage. 
Mixels and errors in the segmentation cause a basic error which 
is impossible to remove in the interpretation stage. A big 
problem in the study area is that many features are too small to 
be reliably detected in Landsat TM images. For instance, the 
width of many canals is about one pixel or less and thus the 
999 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996