anbul 2004 
RING 4.0 
National 
stitute for 
Space 
esearch 
  
"Ww.dpi. 
inpe.br/ 
spring 
Region 
'rowing/ 
Vatershed 
Remote 
sensing 
3ins et al. 
1996 
08/2003 
Vin, Unix 
tand-alone 
N 
30 min 
No 
  
2000 by | 
2000 | 
Em] 
"ORE 
Jo (external 
Ja 
et 
Freeware 
Lu à 
free 
  
ERE 
Pixel); ? when 
1ternal vector 
)3 (3500 €) 
Jg parameter 
c 
International Archives of the Photogrammetry, Remote Sensing 
  
  
Results for the software CAESAR were available only for the 
rural test area, The segmentations of the ,Minimum Entropy 
Approach' cover only one fourth of both areas (each 1000 by 
1000 Pixel) due to a lack of performance. Except for InfoPACK 
the segmentations were done in different levels using altered 
parameters affecting segment size. When an object was poorly 
segmented, coarser or finer segmentation levels could be used. 
32 Pre-processing of the segmentation results 
All segmentation results were converted into vector format 
(ArcView shape file) for the subsequent comparison of 
geometry. Only eCognition and the Erdas extension ‚Image 
Segmentation’ are able to generate a GIS-readable vector 
format. All other results were generated in raster format (TIFF) 
with a unique value for each segment. Geocoding was restored 
by adding a world file (TFW). Then a raster-to-vector 
conversion was carried out using Erdas Imagine. Only in the 
case of the ‚Minimum Entropy Approach’ this procedure results 
to some negative effects, because the implemented triangulation 
algorithm fractionalises the image without respect to raster 
boundaries. The preliminary segments are stored in a 
proprietary vector format, which cannot be saved. Rather the 
segmentation result was converted into a raster output, which 
admittedly leads to more partial segments and faulty 
segmentations (unclosed polygons etc.). These unavoidable 
effects have a negative influence to the quality assessment. 
33 Quality assessment 
Firstly, all results came under an overall visual survey. General 
criterions, like the delineation of varying land cover types (e.g. 
meadow/forest, agriculture/meadow, etc.), the segmentation of 
linear objects, the occurrence of faulty segmentations and a 
description of the overall segmentation quality were in the 
focus of this first step. 
Furthermore a detailed comparison based on visual delineated 
and clearly definable reference areas was carried out. Therefore 
20 different areas (varying in location, form, area, texture, 
contrast, land cover type etc.) were selected and each was 
visually and geometrically compared with the segmented pen- 
dants. The geometrical comparison is a combination of formal 
factors (area, perimeter, and Shape Index (area-perimeter-ratio)) 
and the number of segments resp. partial segments (in the case 
of over-segmentation). For all features the variances to the 
reference values were calculated. 
As partial segments all polygons with at least 50 94 area in the 
reference object were counted. The Shape. Index comes from 
landscape ecology and indicates the polygon form. It is 
calculated by the quotient of perimeter and four times the 
Square root of area. Additionally the quality of segmentation 
was visually rated (0 poor, 1 medium, 2 good). 
À good segmentation quality is reached, when the overall 
differences of all criteria between the segmentation results and 
the associated reference objects are as low as possible. 
Furthermore the objects of interest should not be over- 
Segmented too much. 
and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
4. RESULTS 
4.1 Overall visual survey 
eCognition: Despite their differences albeit using the same 
parameters the segmentation the results of eCognition 2.1 
(figure 1) and 3.0 (figure 2) are of good quality. Indeed they 
sometimes contain irregular or ragged delineated segments, 
especially at seam-forming boundaries and in woody areas. In 
areas of low contrast the occurrence of faulty segmentations is 
possible. Large homogenous image objects are divided 
arbitrarily sometimes. 
eCognition uses a new segmentation algorithm since release 3.0 
which enables a result not depending on image size. This is an 
important improvement because often parameters are tested on 
small subsets. Nevertheless the old algorithm of version 2.1 
could still be used alternatively in the current release 4.0. 
Altogether eCognition has a high potential due to its multi-scale 
segmentation and the fuzzy logic based image classification 
capabilities. Because of the various interfaces to other GIS and 
remote sensing software systems important user requirements 
are complied. 
  
  
Figure 2. Segmentation result of eCognition 3.0. 
Data Dissection Tools: The segmentations of the ,Data 
Dissection Tools' (figure 3) offer only partly satisfying results. 
The software tends to a strong over-segmentation of bright 
image areas, whereby a multitude of small segments occur. 
Homogeneous areas like fields, meadows or water bodies are 
segmented almost correct. Only very large areas are divided 
arbitrarily sometimes. Explicit mistakes of delineation appear in 
image areas of low contrast (e.g. woody areas). As in bright 
1099