The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
Beginning at the top of the pyramid each node is labelled as root 
or to which root it belongs. When the label propagation has 
reached the bottom of the pyramid, each pixel of the original im 
age data is labelled by a number indicating to which root, i.e. to 
which segment it belongs. 
4 MERGING 
The described approach results in a very fine segmentation, which 
is very sensitive to changes in the image. This ensures that all 
real-world boundaries are contained in the segmentation, but re 
sults in an image, which is oversegmented to a certain degree. 
However, the obtained segments cover only strong homogeneous 
regions and have therefore good spectral properties. A simple 
merging algorithm is sufficient to reduce the number of segments 
and to obtain a good segmentation result. Beginning with the 
largest segment each of its neighbours is investigated. If the sim 
ilarity s(Ri, R2) between segment Ri and its currently investi 
gated neighbour R2 is greater than a certain threshold both seg 
ments are merged. This is done until all similiar neighbouring 
segments are merged. Afterwards the next region is investigated 
in the same way. 
s(Ri, Ri) 
6(X) 
\k £ ä(x) (22) 
f 1, ifdw(X,Yi) < 7 d w (X,Y a ) 
1 0, else ' ’ 
where Yi is the sample covariance matrix of region R t and 
dvv(X, Yi) the distance measure defined by (8). The constant 
7 = (1 + e), where e is a small number, relaxes this constraint a 
little bit. In our experiments 7 was set to 7 = 1.01. 6 
5 RESULTS 
In Figure 2 the segmentation result of a 727 x 1047 image of fully- 
polarimetric SAR data is shown. 
Figure 2: left: original; right: segmentation 
Figure 3 shows a magnification of a part of Figure 2. The homo 
geneous regions have been extracted successfully. Regions with 
different properties have been separated and are considered as in 
dependent segments. 
Figure 3: The figure shows the segmentation in more detail. 
At the top left side the original image data is shown and at the top 
right side the segmentation result. 
The bottom line shows the segment borders without (left) and 
with (right) merging. 
6 CONCLUSIONS 
The proposed algorithm has several adavantages with respect to 
other segmentation approaches. There is no need for any kind of 
handmade initialisation. Furthermore, not only a segmentation is 
obtained, but a hierachical structure of the image, which contains 
more information than a simple collection of disjoint regions.