2004 
3 ha 
1 the 
; ML 
NLP 
total 
‚46% 
| and 
roved 
NLP 
vever, 
ed by 
area. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
Maximum Likelihood Classification EA 
Output va 
RA = Y 18, 
(15m resolution image) Ah 
     
  
Legend 
  
  
  
Figure 8. Classified map of the improved image using ML 
Classifier (3 classes). 
4.3 Accuracy Assessment of the ML Classification 
4.3.1 Landsat-7 ETM+ 30 m resolution data 
The quantitative accuracy assessment was performed to obtain 
more exact information on how accurate the image 
classification method can detect single tree felling. The 
quantitative accuracy assessment was carried out by calculating 
the overall accuracy, class mapping accuracy and kappa statistic 
based on the confusion matrix. The confusion matrix was 
generated after crossing the classified map with the test data set. 
Confusion matrices are presented a graphical representation of 
these accuracy measures for the three ML classifications. 
Notice the kappa value of the second classification (KA= 71). it 
is much higher compared to the first (KA- 53) and the third 
classification (K= 46). The overall accuracy (OA= 81%) is also 
much higher than the other two classifications (OA= 66% 
versus OA= 71%). This explains the more distinct classes that 
were observed in the output map (Figure 9) as compared to the 
other two outputs. However, the class mapping accuracy of 
NLP (CA nlp= 58%) which is the major issue in this research is 
slightly lower than in the first classification output (CA nlp= 
61%). For this reason, the first classification was selected for 
further analysis 
o i 
8 6 classes 
5 | @3 classes 
o 
S O2 classes 
o 
  
Figure 9. Comparison of Accuracies of ML Classified Maps 
with different number of classes. Note: OA= Overall Accuracy; 
KA= Kappa; CA nlp= Class Accuracy NLP. 
4.3.2 Landsat-7 ETM+ 15 m Resolution Data 
Classification of the improved image increased the different 
accuracy measures slightly compared to the second 
937 
classification of the original image. The kappa and the class 
accuracy of NLP were increased with 2% and the overall 
accuracy with 1%. The NLP class accuracy is 1% lower than 
the first classification of the original image. It would have been 
interesting to perform the classification using six input classes 
for a better comparison with the first classification of the 
original image. However, there was not enough time to carry it 
out. 
90.00 
80.00 
70.00 
60.00 
50.00 
40.00 
30.00 
20.00 
10.00 
0.00 
8054 8188 
    
Percentage 
FEE Rem tie ————Á 
  
  
  
OA KA CA nlp 
Figure 10. Comparison of Accuraties of ML Classified Maps 
(15 m versus 30 m). Note: OA- Overall Accuracy; KA- Kappa; 
CA nlp- Class Accuracy NLP. 
4.4 Sub-pixel Image Classification Results 
The image classification was performed using band 1-5 and 7 
of Landsat-7 ETM- (30 m resolution). The output of the SP 
classification shows eight different MOI fraction classes 
ranging from 0.2 to 1. There are no detections for MOI fractions 
less than 20%, because this is below the SP classifier threshold. 
Figure 11 shows the classified image after the merge. This map 
gives better view of the NLP detections compared to the 
original map. The map illustrated in Figure 11 shows the NLP 
detections in RKLI1. It shows a large amount of NLP detections. 
The area covered by these detections is 3019ha which equals to 
approximately 44.47% of the total area of RKLI. Notice the 
spatial distribution of the NLP in the map. It shows a large 
concentration of NLP along the main road seen here as a curved 
line feature. Moving in North West direction down the road the 
concentration of NLP decreases slightly at first but increases 
again up to where the road ends. From then on it decreases 
again in East West direction with some variation in intensity 
  
Subpixel Classification Output 
(2 classes) 
    
  
   
Legend 
CJ nip 
other 
   
  
  
  
Figure 11. NLP Detection Map derived from SP output map