The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3h. Beijing 2008 
618 
conditions. The results were illustrated in a chart with 
logarithmic scale abscissa. 
4. RESULTS AND DISCUSSION 
All analyses were performed using different images or image 
collages. Some examples are shown below (figure 4). 
Neighbourhood system extensions have barely influence on 
parameter extraction, because with it only new parameters are 
added. In enlarged neighbourhood systems (NBS) the existing 
parameters represent the same pixel interaction. 
In contrast to that the scaling of the picture has a strong 
influence on the model parameters. 
4.2 Factors that influence segmentation 
Figure 4. Synthesised textures (top). The upper left picture was 
used for this calculation; the upper right is the normalized 
image. Error in parameter extraction (bottom) with (+) and 
without ( ) parameter normalisation. 
Radiometric normalisation of the image is the first processing 
step. Thus, the segmentation is not always improved, but 
shading effects can be avoided effectively (see Figure 6). The 
normalisation to a uniform arithmetic mean and deviation leads 
to better results than a histogram equalization. 
Figure 6a. Segmentation result with shading effects: Original 
(left), shadow removal without normalisation (error = 16%) 
(right) 
4.1 Parameter Validation 
Extractions of the parameter vector from synthesised images do 
not always work perfectly. Although the extracted parameter 
vector has the right orientation it has the incorrect length, which 
has to be normalised to unit length to ensure reproducibility of 
the extraction. This effect can be shown with synthesised 
textures as seen in Figure 4. 
Using the normalised parameters, the extraction can be done 
with relatively small errors. The parameter extraction is very 
sensitive to sensor noise. Even if there is already an obvious 
degradation at a noise level of 5% a normalisation of the 
parameter vector has a positive effect (see Figure 5). 
Figure 5. Image segmentation error with different degrees of 
noise with (+) and without ( ) parameter normalisation. 
Figure 6b. Segmentation result with shading effects. Original 
(left), shadow removal with normalisation (error = 2%) (right) 
The normalisation of the extracted parameter vector also yields 
better segmentation results. The example has already indicated 
that the parameter normalisation is required for valid parameter 
extraction. Accordingly, this has a positive impact on the 
segmentation as well. 
Figure 7. Example of segmentation errors with and without 
parameter normalisation (left) and the original image (right) 
The texture window size and the scaling strongly affect the 
segmentation (see Figure 7). However, for each picture 
different values lead to an optimal result. In general the 
following trend can be derived: