'anbul 2004 International Archives of the Photogrammetry, Remote Se 
ae 
  
nsing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
nilar classi. 
reduces the 
It must be 
have been 
he segmen- 
natural boundaries (e.g. forest edges) were often imprecisely 
delineated by the used triangulation algorithm. Large homoge- 
neous areas were divided frequently. Simultaneously, effects of 
the above-mentioned raster-to-vector conversion could be 
found. Generally it can be pointed out, that the triangulation 
algorithm often leads to straight segment boundaries or sections 
  
: The Erdas 
6) leads to 
segmentation 
| land cover 
ontrast were 
of fields vs. 
merged into 
. inside only 
egment size 
uate and ho- 
ions in terms 
ndaries. The 
| new version 
> of the block 
set freely in 
ie image Size. 
too. But the 
  
sion ‘Image 
Figure 7. Segmentation result of the ‘Minimum Entropy 
Approach’. 
resp. typical segment shapes which are closer to a human 
interpretation. 
SPRING 4.0: The segmentation results of the region growing 
algorithm implemented in the image processing software 
SPRING are showing a good overall impression (figure 8). 
Homogeneous areas are delineated well but often over- 
segmented. Heavily textured areas as forests are mostly under- 
segmented. Sporadic segmentation mistakes occur. However, 
the ease of operation as well as the data handling of the 
software is insufficient. The implemented edge-based water- 
shed-algorithm was also tested, but was leading to worse results 
(strong over-segmentation) and was therefore not used for his 
evaluation. 
  
Figure 8. Segmentation result of SPRING 4.0. 
4.2 Comparison based on reference areas 
The overall results of all 20 reference areas are cumulated in 
table 2. As shown in this table the results of SPRING, 
eCogniton 2.1 and 3.0, InfoPACK and the ‚Minimum Entropy 
Approach’ are reaching the best average area conformity. 
Except for InfoPACK, the same result is shown in the case of 
the average conformity of perimeter and the Shape Index. The 
high conformity of the Shape Index in the case of the 
Minimum Entropy Aproach' is affected by the segment shapes 
resulting from the triangulation algorithm which is closer to 
human interpretation. 
Especially within the number of segments both versions of 
eCognition revealed their strengths. Both led to the slightest 
over-segmentation in this evaluation. In this regard the results 
of SPRING could be rated as good. The results of the Erdas 
Imagine extension ,Image Segmentation’ also reached a slight 
number of segments, but due to strong differences of the other 
values the result is indicating under-segmentation. 
  
  
  
  
  
  
  
  
  
Segmentation program eCognitio eCognitio Data CAESAR | InfoPACK Image Seg- | Minimum | SPRING 
n 2.1 n 3.0 Dissection 3.1 1.0 mentation | Entropy 4.0 
Tools (for Erdas | Approach 
Imagine) 
Number of reference 20 20 20 10! 20 20 11' 20 
areas 
Average difference of 2 a 3 2 
area [%] 12,5 15,9 2100,3 75.1 11,1 107,0 13,6 8,2 
Average difference of 
: n 3 10,0 10,8 
perimeter [94] 139 in? a S55 We m 
Average difference of 5 
= 3 3 : 7,1 10,0 11,7 
Shape Index [%] 15,7 02 ii Es 2 i 
Average number of 
= 5,9 9,0 6,2 
partial segments r9 Ls 1346 i: 7d 
fe —— 
Average quality, visual 
, Visug 9 0,2 0,8 0,9 
evaluated [0...2] La 09 i = os 
———— PH 
  
  
  
  
  
  
  
ba ; : 
differing number due to partial incomplete segmentation results 
  
  
  
n Entropy Ap 
boundaries o! 
More complet 
Table 2. Cumulated results of all 20 reference areas. 
1101