International Archives of the Photogrammetry, Remote Sensing 
image areas sometimes single pixels are segmented 
(comparable to the salt-and-pepper effect). The near infrared 
was only marginally used in consequence of an unfavourable 
especially the 
weighting, which impacts separation of 
deciduous vs. coniferous forest. 
  
Figure 3. Segmentation result of the ‘Data Dissection Tools’. 
CEASAR: The program CEASAR 3.1 which was developed 
for radar data leads to results that cannot be used for further 
processing (figure 4). The produced segments are compact and 
of a similar size. This effect occurs even though using different 
segmentation parameters which yield only to a varying average 
segment size. Thus, small structures and in particular linear 
elements are often segmented faulty and an over-segmentation 
is the consequence. Boundaries of low contrast are represented 
badly, sometimes boundaries of sufficient contrast too (e.g. 
forest vs. meadow). 
  
Figure 4. Segmentation result of CAESAR 3.1. 
InfoPACK: The result of InfoPACK 1.0 (figure 5), the further 
development of CAESAR, shows a good delineation for most of 
the objects, but tends strongly to over-segmentation. Homo- 
geneous areas are thereof less affected and are adequately 
represented. In particular especially forests and built-up areas 
were much partitioned. At land cover transitions often inter- 
fering seam-forming segments were created. Generally low 
contrasted boundaries were segmented correctly. Compact and 
nearly similar sized segments as in CAESAR exist no longer. 
For processing scenes of any size the software uses an imple- 
mented tiling algorithm. Indeed this leads to additional segment 
boundaries at the tile transitions. Furthermore margin effects 
can yield to different results on both sides of the tile boundary. 
As eCognition the software contains additional classification 
and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
tools. Thus, a classification based on merging of similar classi. 
fied and neighbouring segments is possible and this reduces the 
number of elements to be classified significantly. It must be 
pointed out, that InfoPACK as well as CEASAR have been 
developed to analyse very noisy radar data. Hence, the segmen- 
tation of optical data could be suboptimal. 
  
Figure 5. Segmentation result of InfoPACK 1.0. 
Erdas Imagine extension *Image Segmentation': The Erdas 
Imagine extension ‘Image Segmentation’ (figure 6) leads to 
over- and under-segmentation within the same segmentation 
result. Well-contrasted boundaries between main land cover 
classes were correctly represented. Areas of low contrast were 
often not segmented. In particular the delineation of fields vs. 
meadow was problematic. Forested areas were merged into 
large conglomerates, with small island segments inside only 
slightly greater than the parameter minimal segment size 
chosen. Linear elements were segmented inadequate and ho- 
mogeneous image objects were divided frequently. 
Furthermore, the result contains faulty segmentations in terms 
of non-explainable horizontal or vertical boundaries. The 
degree of this effect has been slightly reduced by a new version 
from September 2002. It was mainly a consequence of the block 
size used by the software, which can now be set freely in 
accordance to the available system resources or the image size. 
Thus, the computing time has been reduced too. But the 
segmentation quality remained nearly unimproved. 
  
Figure 6. Segmentation result of the Erdas extension ‘Image 
Segmentation’. 
‘Minimum Entropy Approach’: The ‘Minimum Entropy AF 
proach’ (figure 7) was well reproducing straight boundaries 0 
man-made features (e.g. field boundaries, roads). More complet 
1100 
Inte 
natu 
delit 
neoi 
the 
foun 
algo 
  
Tesp. 
inter 
SPRI 
algor 
SPRI 
Home 
segm 
segm. 
the e 
softw 
shed- 
(stron 
evalu: 
Segn 
Num 
areas 
  
Aver, 
area | 
Aver: 
perim 
Aver: 
Shape 
Aver: 
partia 
| 
Avera 
evalu: 
  
M 
! differ