International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
accuracy. Kappa statistics and its variance were also calculated 
to test the significance of difference in accuracy. The 
significance of difference test between the confusion matrices 
was done using the Z test with a — 0.05. 
In addition to the quantitative assessment, a qualitative 
assessment of the classified images was done by examining the 
classified maps visually and relating it to field knowledge. This 
is to find out if the map reflects reality. 
4. RESULTS AND DISCUSSIONS 
The ML classification was performed using different numbers 
of input classes and two different image resolutions 30 and 15 
meters. 
4.1 ML Results Using Landsat 30 m Resolution Data 
The ML classification of the original data was performed using 
different numbers of input classes. The first classification was 
done using six input classes (i.e. NLP, F1, F2, F3, F4 and NF). 
In the second classification no distinction was made between 
the forest classes, thus the input was NLP, F and NF. The third 
classification used only two classes, NLP and Other. The 
classification output was then compared to find out which 
number of input classes gave the highest class accuracy for 
NLP. The best classification output in terms of NLP class 
accuracy was selected for comparison with the SP classification 
output. 
Figure 5 presents the classification result using six input 
classes. The total detected NLP covers 3,946 ha of RKL1 which 
accounts for approximately 58.48% of the total area. The map 
shows a large amount of NLP detections along the main road in 
the upper right part of the image. The lower left part shows less 
NLP detections compared to the upper right part. Notice the 
road and the agriculture area in the upper left corner of the 
image. Figure 6 shows the same image in which the forest 
classes and the non forest classes were merged after 
classification. 
  
Maximum Likelihood Classification 
Output 
(30m resolution) 
       
  
Legend 
  
  
  
Figure 5. Classified map of the original image using ML 
Classifier (6 classes) 
Figure 7 presents the classification output using 3 input classes. 
This map shows less NLP detections compared to the first 
classification output. Again most of the NLP detections are 
found along the main road and smaller amount in the lower left 
part of the image. The total NLP detections amount to 2,193 ha 
of RKLI, approximately 32.5% of the area. This is less then the 
area found in the first classification. 
  
Maximum Likelihood Classification d 
Output " 
(30m resolution) 
  
   
   
    
Legend 
Wl New logged point 
[J Other 
  
  
  
Figure 6. NLP Detection Map derived from the 6 classes ML 
output map. 
The result of the third classification shows much more NLP 
detections compared to the previous classifications. The total 
area covered by NLP is 5,362 ha which corresponds to 79.46% 
of the total RKL1 area. Observe the detection of the road and 
agriculture area in the left part of the image. 
  
Maximum Likelihood Classification 
Output 
(30m resolution) 
  
  
   
   
      
Legend 
New logged point 
EB Forest 
[J Non forest 
  
0 = 4 km 
  
  
  
Figure 7. Classified map of the original image using ML 
Classifier (3 classes 
4.2 ML Results Using Landsat 15 m resolution data 
Figure 8 shows the output of the classification of the improved 
image. In general, the same trend in distribution of NLP 
detections can be observed as for the original image. However, 
the image shows a more distinct pattern. The area covered by 
NLP is 1,624 ha, which is about 24.07% of the total RKL I area. 
936 
Int 
  
4.3 
4.3. 
The 
mor 
clas 
qua 
the 
base 
gent 
Con 
thes 
Not 
is It 
clas 
muc 
vers 
wert 
othe 
NLE 
sligl 
61% 
furtl 
Figu 
with 
KA= 
4.3. 
Clas: 
accu