f the 
lysis 
other. 
1 the 
ratio 
d the 
The 
ering 
ed in 
imber 
neans 
wn à 
were 
ning 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
  
  
  
  
  
  
t 
SE 
£7 s 
Figure 10: Competitive learning method (V class regions 
highlighted in green) 
  
"M 
A 
E A 
  
  
  
  
  
  
  
  
  
  
  
  
K-means 
= B-class T-class Back- Null 
& e ound 
zg groun 
s. 2 B-class 178851 0 2140 0 
” = | T-class 0 152557 7060 9945 
ce 
= | Background 25091 0 609439 0 
B | Null 769 45 6856 9248 
dà Total 204711 152602 625495 19193 
  
  
  
Table 1: Confusion matrix between K-means clustering and 
Competitive learning method 
Notice that the confusion matrix is almost diagonal which of 
course could be expected. It shows that both clustering 
algorithm recovered the three classes B-class, T-class and 
Background to a high degree of agreement. Null is used for 
rejection indicating assignment to none of the three classes. 
In percentage values the degree of agreement between the 
clusters of both clustering algorithms is summarized in Table 3: 
  
  
  
  
  
  
  
Total common area 
B-class 180991 98,8 % 
T-class 169562 90,0 % 
Background 634530 96,1 % 
Null 16918 54,7 96 
Total 1002001 84,3 96 
  
  
  
  
Table 2: Common clustering areas of K-means clustering and 
Competitive learning networks method 
That B-class and T-class can be easily identified with regions 
covered by buildings and trees was already discussed above. 
  
  
  
  
  
  
Class K-means Competitive Fuzzy c-means 
(Count, 946) learning (Count, 46) 
(Count, 96) 
B-class 180991, 204711, 144063, 
18.196 20.496 14.4% 
T-class 169562, 152602, 196599, 
17.0% 15.2% 19.6% 
  
  
  
Background 634530, 625495, 657313, 
63.3 % 62.4% 65.6 % 
Null 16918, 19193, 4026, 
1.7% 1.91% 0.40% 
Table 3: Clustering areas for all three clustering methods 
Taking all three clustering areas simultaneously into account is 
shown in Table 3. Already by comparing the counts in each 
class a striking difference to the Fuzzy c-means result has to be 
observed. For the two classes of major interest in this study, the 
B-class and T-class, the differences are quite significant. Visual 
interpretation indicates that the B-class of K-means and 
competitive learning include building areas but also regions 
related to roads which supports the smaller number of counts of 
the fuzzy C-means method to be more precise. Similarly the 
higher number of counts for the T-class indication (3D) 
vegetation regions (trees, bushes) obtained with the fuzzy C- 
means method is supported by visual interpretation. Without 
ground truth we do not intend to draw further conclusion at this 
stage of our investigations. 
5. SUMMARY 
On a first view all three clustering algorithms provide 
reasonable classes which point back to vegetation, buildings 
and background. Comparison between the three clustering 
algorithms indicates a higher consistence of the results of K- 
means and Competitive learning networks. Fuzzy C-means 
deviates stronger but without comprehensive ground truth a 
absolute quality assessment is not feasible. The importance of 
the two input channels NDDI band and TopHat filtered last 
pulse range image for separating vegetation region with 3D 
extend and building regions from background has been shown 
clearly by the experiments. 
REFERENCES 
Arefi, H., Hahn, M., Lindenberger, J., 2003. LIDAR data 
classification with remote sensing tools. Joint ISPRS 
Commission IV Workshop "Challenges in Geospatial 
Analysis, Integration and Visualization II", Stuttgart, 
September 8- 9. 
Axelsson, P., 1999. Processing of laser scanner data — 
algorithms and applications. ISPRS Journal of 
Photogrammetry and Remote Sensing, 54(2-3): 138-147. 
Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective 
Function. Plenum Press. 
Bezdek, J.C., 1987. Some non-standard clustering algorithms. 
In: Legendre, P. & Legendre, L. Developments in 
Numerical Ecology. NATO ASI Series, Vol. G14. Springer- 
Verlag. 
Hung, Chih-Cheng, 1993. Competitive Learning Networks for 
Unsupervised Training, International Journal of Remote 
Sensing, Vol. 14, No. 12, pp. 2411-2415. 
Maas, H.G., 1999. The potential of height texture measures for 
the segmentation of airborne laserscanner data. Proceedings 
of the Fourth International Airborne Remote Sensing 
Conference, Ottawa, Canada. pp. 154-161. 
TopScan, 2004. Airborne LIDAR Mapping Systems. 
http://www.topscan.de/en/luf/messsyst.html (accessed 10 
Feb. 2004)