3. Istanbul 2004 
VE 
ssified roof pix 
o 
o 
ed iterations 
nother satellite 
ipore in August 
ule-based RGB 
oof. In order to 
1 points were 
ructures having 
. 
  
e) in IKONOS 
on graph 
d: the ordering 
had 500-3000. 
0001 (tuning). 
to 10 with the 
ordering; the 
tuning with a 
). 
   
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
4.3. Detecting roads 
The presented self-organization technique can be applied also in 
a running-window environment. In this case the necessary 
structuring (moving) element was a simple cross having 5 
processing elements (neurons) and 4 connections. The test 
image was a 0.4 m ground resolution black-and-white 
orthophoto about the Frankfurt am Main region. The image cut- 
off was captured over a sparse village part near to the Frankfurt 
international airport. 
The image preprocessing was solved by thresholding the 
intensities, so the further steps used binary data. The adequate 
parameterized technique identified the junctions as it is shown 
in Figure 5. 
  
Figure 5. Recognizing a four-arm junctions by the cross kernel 
The road detection approach has a new type of SONG 
application: the input image has been partitioned, then the 
previously defined and in the whole detection procedure fixed 
  
Fi 
369 
structuring element is evaluated by the SONG technique. The 
developed SONG algorithm was limited for the ordering phase; 
the tuning was not so important; the processing speed could be 
accelerated in this way. This means that the structuring element 
must be created and described before the run. The partitioning is 
carried out by creating regular square raster on the image. The 
applied structuring element was the same as in Figure 5, a 5 
neuron cross with 4 connections. The adjacency matrix and the 
rule to build the initial neuron weights were constructed in the 
starting step. 
The main question in the practical implementation was (1) to 
find the right size of the image partitioning and (2) to get the 
right parameter set for this special case. Both questions were 
answered after executing an experiment series with various size 
partitions and different SONG control parameters. The best 
result after visual evaluation was a partition size of 20 pixel by 
20 pixel (8 x 8 m). 
The control parameters have not too large variety, because the 
given neuron graph was not large; the highest neighborhood 
could be 2. The free parameters were the starting learning rate 
and the number of epochs. From earlier experience, these two 
controls have strong interaction, so during the tests both are 
varied: learning rates were between 1.0 and 0.5 and the number 
of epochs between 20 and 500. 
The SOM and therefore (because of the inheritance) SONG 
have an observation: after a "critical" number of epochs, the 
system gets a stabile status, 1e. the neuron weights don't change 
anymore. This critical epoch number was searched, and found at 
50 for the 20 x 20 partitions. Then the effective starting learning 
rate was found at 1.0. 
The essence of the algorithm for this running window style 
computing model is the following: 
gure 6. Road detection by running window SONG algorithm