il 2004 
image 
oes in 
| size 
alysis 
pixel 
  
ences 
dc 
k 
rences 
x 
Nat) NN M NNI RN Ine 
> 
2 
  
can be 
cted in 
S is to 
id dark 
and the 
ie edge 
if. The 
rofiles, 
mined. 
tained. 
re built 
erneath 
S, itis 
| object 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
  
space (Parker, 1997). But in general, the grey value profile will 
not be sharp due to the imaging sensors. In Figure 5a, the edge 
lies in left-right direction, the edge in 5b lies in diagonal and the 
edge in 5c lies in upper-down direction in the image. All of 
them are the edges between roof (white) and grass (dark) and 
the grey value profiles are not sharp like in the object space. For 
analysis, the EDGE module of BLUH program system of 
Hannover University has been used. It is realized that the 
profiles are not so smooth. However, the profile of diagonal 
edge is more smooth than a and c. Figure 5d is an edge of 
swimming pool and the dark side is the water and the profile of 
diagonal edge.is the sharpest among the samples. The grey 
value profile gives the effective pixel size from the differences. 
The width of point spread function at 50% height can be used as 
effective pixel size (Topan et al., 2004). For KVR-1000 image 
used in this paper, the effective pixel size is nearly 2.7 m. This 
means that, during the digitizing process, the operator senses 
approximately 2 pixels instead of I pixel. This situation will 
effect the digitization negatively. 
  
  
  
  
  
Figure 6. Created segments 
Starting point of object oriented approach using the commercial 
software eCognition v3.0 is to create segments which are basis 
for building objects. The output of the segmentation step is 
shown in Figure 6. The boundaries show equi-characteristic 
cluster of pixels. The characteristics are defined by parameters 
before processing. 
  
  
147 
Figure 7. Mixed segments 
A lot of problems occurred in the course of processing. One 
problem experienced is spreading of grey values over 
neighboring pixels due to buildings having the same 
characteristics would have been classified into the same class. 
But this is not the case because some buildings are shadowed by 
the adjacent buildings. The similar reflectance properties of 
different neighboring objects give rise to missegmentation of 
these different classes. This situation is shown in Figure 7. 
  
Figure 8. Objects in buildings class 
Due to above stated problems and selecting small scale 
parameter the real world cannot be extracted exactly. On the 
contrary setting large scale parameter values leads to clutter of 
buildings. Several experiments are carried out with different 
parameters settings but the expected results are not satisfactory. 
The most acceptable output is obtained using mean grey value 
criterion. Grey values falling below 185 are disregarded and 
some extra.setting values are used in the framework of this 
study. The classification results obtained hereafter are shown in 
Figure 9.. 
  
Figure 9. Overlap of manual and automatic digitized objects