canals disappear in the region-based segmentation process, 
although the segmentation results are otherwise quite good. To 
improve the interpretation results, images with higher spatial 
resolution are needed. 
The advantage of the segmentation in this study was the 
increase in the Maximum Likelihood classification accuracy 
compared to a pixel-based classification (Table 8). A segment- 
based classification gives a result with homogeneous regions 
like in a map. If segmentation is not used, good generalisation 
procedures are needed. 
In this study, only three information sources were available in 
the rule-based postclassification, and the belief values, which 
were based on relationships between reference points and 
classification attributes, were the same for all the test area. The 
changes in the postclassification compared to the Maximum 
Likelihood results were quite straightforward. For instance, 
every segment or pixel which was cultivated land on the old 
map, forest in the Maximum Likelihood classification and the 
height of which was between 0 and 10 metres was 
postclassified as rice. A more ideal solution would be belief 
values which change continuously depending on characteristics 
of each segment or pixel. For instance, the probability that a 
field changes to urban area is much bigger near the town centre 
than far from it and the probability that forests exist increases as 
the height increases. However, to reliably exploit this kind of 
information, more reference points than what was available for 
this study would be needed. 
The future research will focus on use of SAR images and SPOT 
panchromatic images together with Landsat TM images and use 
of spatial information, such as neighbourhood relationships of 
segments, in the interpretation process. The interpretation 
method has been designed and implemented so that it allows 
addition of these new data sources. The most difficult problem 
is to find good rules which give realistic and objective belief 
values based on the different data sources. 
REFERENCES 
Ahokas, E., Jaakkola, J. and Sotkas, P., 1990. Interpretability of 
SPOT data for general mapping. European Organization for 
Experimental Photogrammetric Research, Official Publication 
No. 24. 
Beaulieu, J.-M. and Goldberg, M., 1989. Hierarchy in Picture 
Segmentation: A Stepwise Optimization Approach. IEEE 
Transactions on pattern analysis and machine intelligence, 
11(2), pp. 150-163. 
Duda, R. O. and Hart, P. E., 1973. Pattern Classification and 
Scene Analysis. John Wiley & Sons, New York. 
Gordon, J. and Shortliffe, E. H., 1985. A Method for Managing 
Evidential Reasoning in a Hierarchical Hypothesis Space. 
Artificial Intelligence, 26(3), pp. 323-357. 
Shafer, G. and Logan, R., 1987. Implementing Dempster's Rule 
for Hierarchical Evidence. Artificial Intelligence, 33(3), pp. 
271-298. 
Wilkinson, G. G. and Mégier, J., 1990. Evidential reasoning in 
a pixel classification hierarchy — a potential method for 
integrating image classifiers and expert system rules based on 
geographic context. International Journal of Remote Sensing, 
11(10), pp. 1963-1968. 
1000 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996