Pakzad, Kian 
  
5.2 Initial Segmentation 
The assumption that a biotope mapping exists does not apply for every area. For some areas an automatic initial 
segmentation is necessary instead. A view on the aerial images of moorland shows that there are many spectrally 
inhomogeneous areas. Because some moorland classes consist of inhomogeneous areas a simple multispectral 
classification is not suitable for segmentation. The segmentation process is depicted in Fig.2. It requires CIR-aerial 
images. In the case of grayscale images a biotope mapping or a manually created segmentation is necessary. 
The most important information for the segment 
borders in moorland are the streets and the main DM du 3 
ditches. This information can usually be Streets Sp ed 7 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Ditches Sedments 
extracted from the GIS database, e.g. from aad 
ATKIS Basis DLM. Based on this information an with high N ER n 
initial segmentation is performed. The CIR Aerial Compute 1 | .NiDhwlus with high Segmented 
knowledge exploited for the segmentation is the image s I Exact / So rep 
same used in section 5.3. The next step is to Suse" aen zo 
refine the initial segmentation: Inside every NVDI-density 
  
segment the normalized difference vegetation 
index (NDVI) is computed and the regions with 
high density of vegetation index and also with 
very low density were extracted. For every extracted subsegment the form parameter compactness is computed. Only 
subsegments with high compactness and a minimum area size are accepted as valid regions for the final segmented 
image. Fig.5 shows the result of this procedure for two parts of the test area. 
Figure 2. Initial segmentation of moorland 
5.3 Interpretation 
This section describes the monotemporal interpretation of moorland from aerial images. As described in section 2 the 
system we used for this work needs the prior 
knowledge in the explicit representation form of 
semantic nets. Therefore, the prior knowledge about 
the relevant area is formulated in a concept net. Fig.3 
shows a simplified version of this net. 
  
  
moorland 
  
  
  
|—part-of 9» 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Scene moorsegment 
Layer 
be w? oe bs Based on section 4 we distinguish four states of 
frost agriculturally area of re-/ area of peat moorland classes: forest, agriculturally used area, 
used area degeneration extraction . . 
= x FT re, area of re-/degeneration and area of peat extraction. 
Sor ge $e tof eof 2 ; 
high medium ET p The states area of degeneration and area of 
vegetation i i . . . > ge . 
po a tracks Vegetation regeneration are combined, because their distinction 
  
  
  
  
  
  
  
density 
rf 
  
  
    
     
  
  
1n aerial images is very difficult. As shown in Fig.3 we 
NE distinguish two layers of abstraction in the concept 
wy || net: a scene layer and an aerial image layer. In the 
scene layer the different states are described with their 
[petes | obligatory parts. E.g. the state area of peat extraction 
|| is characterized by harvester tracks and low vegetation 
density. The state area of re-/degeneration is also 
characterized by harvester tracks in one part, but also 
Figure 3. Concept net for the interpretation of moorland by mid vegetation density. The nodes in the second 
layer, the aerial image layer, describe the depiction of 
the scene layer nodes in CIR aerial images and their 
properties. The nodes describe the structures and colors to be looked for, if a state is to be assigned to a segment. Thus 
both color and textural information are used. That’s why the concept net is suitable for both color and grayscale images 
(see below). At the bottom of Fig.3 segment analysis operators are shown. Every node at the bottom of the aerial image 
layer has access to a special operator. 
  
      
  
    
   
mid NDVI- 
density 
    
  
  
E low dismembered 
|] homogeneity structure 
Segment analysis operators 
  
The segment analysis operators have to verify the meaning of the node for a particular segment. This means that they 
look inside a given segment and estimate whether the hypothesis of the node is correct or not. In this way the operators 
transform the explicitly formulated hypothesis into image processing operations. Only on this level the interpretation 
has direct access on the raster images. E.g. the operator connected with the node dismembered structure analyses the 
structure of the edges in a particular segment. Short and curved lines lead to better assessment than long and straight 
lines. 
  
1106 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
mem M. uu fmm mle of ud based 
Cy 
Q " gh (C