on one data source only. Exceptions to this were four rules 
based on both preclassification result and height data, which 
were needed to distinguish quarries from other open areas. 
Examples of rules based on the preclassification result are 
presented in Table 2. All the rules derived from the old land use 
data are presented in Table 3 and all the rules derived from the 
height data in Table 4. 
  
Condition 
If ML-class is fishpond 
If ML-class is fishpond 
If ML-class is fishpond 
If ML-class is fishpond 
If ML-class is water 
If ML-class is garden 
If ML-class is garden 
If ML-class is garden 
Action 
Confirm fishpond 0.50 
Confirm water 0.30 
Confirm garden 0.10 
Confirm open 0.05 
Confirm water 0.90 
Confirm garden 0.65 
Confirm forest 0.20 
Confirm rice 0.10 
  
  
  
  
  
Table 2. Examples of rules and believes determined from the 
preclassification result. 
  
Condition 
If land use is water 
If land use is cultivated land 
If land use is cultivated land 
If land use is cultivated land 
If land use is cultivated land 
If land use is forest 
If land use is forest 
If land use is forest 
If land use is urban 
Action 
Confirm water bodies 0.9999 
Confirm field 0.60 
Confirm water bodies 0.10 
Confirm urban area 0.15 
Confirm open area 0.15 
Confirm forest 0.50 
Confirm urban area 0.20 
Confirm open area 0.30 
Confirm urban area 0.9999 
  
  
  
  
  
Table 3. Rules and believes determined from the old land use 
data. 
  
Condition 
If height < 10 m 
If height > 10 m 
If height > 10 m 
If height < 25 m and 
ML-class is open 
Action 
Disconfirm forest 0.90 
Disconfirm water bodies 0.70 
Disconfirm field 0.80 
Confirm open 0.70 
  
If height < 25 m and Confirm quarry 0.05 
ML-class is open 
If height > 25 m and Confirm open 0.25 
ML-class is open 
If height > 25 m and Confirm quarry 0.50 
ML-class is open 
  
  
  
  
Table 4. Rules and believes determined from the height data. 
Combination of evidence. When all the belief values that the 
rules give for a particular segment or pixel have been assigned 
to classes, the evidence is combined using the Dempster-Shafer 
method for hierarchical cases. The method calculates a final 
belief value for each class in the hierarchy taking into account 
the belief values of the whole tree. In this study, the segment or 
pixel was classified as the terminal class having the highest 
final belief. 
4. RESULTS 
4.1 Maximum Likelihood classification 
Result of the Maximum Likelihood classification for a subarea 
of size 11.7 km x 13.75 km is presented in Figure 4. The 
confusion matrix of the Maximum Likelihood classification can 
be seen in Table 5. The classification was segment-based 
(Figure 3). 
E ó d Ea 1 
CADRE Bp 5 
SRE Wa Hg GER 
JEBEL Bl I 
| "e t 
Rh 
Sepik 
[ela] DY 
  
  
  
  
  
Urban and open areas 
T Water 
Figure 4. Result of the segment-based Maximum Likelihood 
classification. 
  
Forest 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996