International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
E Ed 
PA d s 
E CE 
D 
Ej 
d 
Figure 7: Fuzzy c-means clu re 
highlighted in yellow) 
3.3 Competitive learning networks 
The clustering results found by the competitive leaning 
networks algorithm are shown in Figures (9) and (10). 
On a first view all three clustering algorithms provide 
reasonable classes which point back to vegetation, buildings 
and background. A comparison will be carried out in the next 
section. At this point we want to emphasis that with the two 
input channels NDDI band and TopHat filtered last pulse range 
image sufficient unique feature information is provided to 
clustering to separate the vegetation with 3D extend and 
building regions from background. 
4. ANALYSIS OF CLUSTERING RESULTS 
The confusion matrix is often used to discuss the results of 
image classification. Given some ground truth the relation 
between the "true" classes and the classification result can be 
quantified. With the clusters the same principle can be applied. 
Mostly a much bigger number of clusters is then related to the 
    
42 
- i 
DEZ 
Figure 9: Competitive learnin 
highlighted in yellow) 
  
given ground truth classes to examine the quality of the 
clustering algorithm. If no ground truth is available the analysis 
may focus on comparing clustering results against each other. 
This kind of relative quality analysis is carried out in the 
following. 
Input to the clustering processes has been the NDDI ratio 
between first and last pulse range images (NDDI band) and the 
TopHat filtered last pulse range image (TopHat band). The 
three processes K-means clustering, fuzzy C-means clustering 
and competitive learning networks are employed as discussed in 
the section before. 
The following confusion matrix (Table 1) contains the number 
of pixels assigned to each cluster in the results of K-means 
clustering and competitive learning networks. Reading down a 
column shows how pixels in one class of K-means were 
assigned in the clustering results of competitive learning 
networks. 
Internatioi 
  
Figur 
  
  
SYIOMJOU 
Sururro] sannadwo) 
  
  
  
Table 
Notice tl 
course i 
algorithr 
Backgro 
rejection 
In perce 
clusters 
Table 
That B- 
covered 
  
  
| Class 
B-clas 
T-clas