ght: 
els. 
factory 
  
  
ation of 
st pulse 
it is not 
Sander Oude Elberink 
4 CLASSIFICATION 
  
    
  
  
Tor 5 - 
41 Results asl A j 
v r i à + ] 
; : : ed 3 Ant j i / | 
The classification has been done by I of j ( : 
performing an unsupervised, K- : j | | 
um : E 5 Ha ! ; | 4 
means classification. In the first step, 3 ] 
houses. sheds and trees will be zb ! | | ; 
classified using the contrast texture ; i | 
alic / OL a Jos ; Ld 
measure and the normalised DSM at E cmd CO WS x E: : 
s : 2 | ocotion 
non-tree pixels. The profiles of these es 
two bands are shown in figure 10b. Fig. 10a: Left: Height image, with profile #1; right: spatial profile. 
  
  
  
  
  
Thereafter the reflectance image has T r a Tm A | 
been used to classify pixels at v f ] e ; \ i 
eround level. Figure 11 a-d show the 3 L | \ 1% ul / \ x i 
two step classification of a small part s P ; 3 / \ | | ] 
of the Optech data set, where figure 3 | | 3 á P ~ | fo 
11 e-f show the classification of the ef | | i er | | | | 7 
FLI-MAP data set. Ss bE à TIR ed oL? " E pe 
ocation acotion 
42 Accuracy assessment Fig. 10b: Left: Profile in anisotropic contrast texture measure. Right: spatial 
profile in DSM at low contrast values. 
Accuracy is determined empirically, 
by selecting a sample of pixels from the classification result and checking their labels against classes determined from 
test areas. Houses and sheds can be determined separately with an accuracy of about 90 96, but if those two classes are 
combined to a new class "buildings", an accuracy of 98 9o is obtained. This is obvious because the distinction between 
houses and sheds only depends on their height; this distinction is not as clear as the distinction between the complete 
   
   
© 
     
        
  
    
b) K-means result : s result 
height objects ground level objects 
Le TT 3 
        
T 
L oA 
e) Fli-Map height image f) K-means result g) Fli-Map reflectance — h) K-means result 
height objects ground level objects 
Fig. 11: Original data and classification results. 
  
building and its surrounding. Trees can be detected with an accuracy of 97 % in the very dense laser scanner data with 
the contrast texture measure, and an accuracy of 98 % is obtained in the 1 meter resolution Optech data sets with 
simultaneous first and last pulse registration. The accuracy of the ground level objects directly depends on the quality of 
the reflectance image. In these two data sets ground level objects could be detected with an accuracy of no more than 70 
%. Better results can be achieved by using additional data, like multispectral data or reliable GIS data. 
43 Interpretation of the results 
The determination of a normalised DSM is an important step to narrow the gray value cluster of buildings and trees, and 
to make a proper discrimination between high and low objects. These results show the reliability of the classification of 
trees on the hand of first and last pulse data, compared to first pulse only data with a higher resolution. After the 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 683