The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
The working style of SONG technique requires a hypothetical 
neural network structure, which can be from e.g. existing maps 
or template description for any image object. 
The author of the current paper has published different studies 
on the self-organizing neuron graph technique [Barsi2003a, 
Barsi2003b]. 
2.3. Growing neuron gas (GNG) 
The third neural self-organizing method has been developed by 
B. Fritzke in 1997 [Fritzkel998], The basic idea is to create a 
neural network by continuously maintained network structure. 
The randomly selected data point is compared to all gas neurons 
- their distances are calculated by e.g. Euclidean norm a 
winner and a second winner is selected, then their connection is 
checked. If they didn’t have any, a new one is created. Then the 
quadratic distance of the data point and the winners are derived 
together with their weight update, which is a function of the 
computed distance. The ages of all connections are increased, 
followed by the remove of all overaged edges. If a neuron 
would lose all connection, it is also deleted. The overaging is 
controlled by a given age limit. The next phase is the update of 
the worst network segment: the neuron of the highest 
accumulated error is selected with its neighbor of highest error, 
and a new unit is established, of course with connections with 
the neighborhood [Fritzke 1998]. Thereafter an error vector is 
calculated. 
The network is growing in every controlled step, where the 
network error the highest. The above described operations are 
repeated till a stop condition is fulfilled (e.g. iteration step limit). 
The algorithm starts with two connected neurons, and neurons 
and their connections are added progressively. The technique 
acts practically without any prior hypothesis or knowledge on 
the analyzed phenomena. The algorithm can manage also 
disjunct structures. 
3. IMAGE ANALYSES BY SELF-ORGANIZATION 
TECHNIQUES 
3.1. Image data sets 
The above mentioned computational models have been tested 
on several data sets. Two of them were used for introducing 
growing neural networks; now only these latter two images are 
showed. Both data sets have high geometric resolution and are 
covering urban areas, where roads are significant man-made 
objects. The extraction of roads can be done either by their axe 
line (skeleton) or even by their bounding edges. The first data 
set has practically no elevation influence, because it’s of a flat 
terrain, while the second one covers a hilly region. 
The first one is a color aerial image taken by a Wild frame 
camera (Fig. 1). The image has a high geometric resolution of 
about 0.2 m per pixel, where the radiometric resolution takes 24 
bits. The image was taken on analogue film, which was scanned 
to be able to be processed by digital photogrammetric 
techniques. 
The second data set has been captures directly digitally: it is a 
cut-off from a Quickbird scene covering the Godollo hills 
(eastern of Budapest). The image (see a detail in Fig. 2) has the 
standard geometric resolution of 0.6 m (panchromatic band), 
and 2.4 m (multispectral band). The purchased bundle product 
has the three visible channels (RGB), the near infrared channel 
and the panchromatic one. Prior to the processing a resolution 
merge was performed by the method of the principal 
components. The resulting radiometric resolution is 24 bits. The 
image has been captured on 13/10/2005. 
Figure 1. Digital aerial image example 
Figure 2. Quickbird image example (Band 1) 
3.2. Executing the analyses 
The aerial image (Fig. 1) was preprocessed by band-wise 
intensity threshold and a growing neural gas of 200 neurons 
analyzed it in 80 000 epochs with a neuron insertion rate was 
100 (every 100 epochs had add a new neuron). The highest 
connection age factor was 50. The maximal connection age was 
90, the error decrease was 0.5, and the weight update coefficient 
took 0.05. The results chapter contains the fitted neural gas to 
the image (Fig. 3). 
The pre-hypothesis of processing of satellite images was proven 
by a Quickbird scenario (Fig. 2). The image has only a single 
spectral band (for this example the first, blue band was chosen, 
because it had the best quality. Another image data sources are 
- theoretically - usable, but their prior test is really suggested. 
Because of the initial random neuron positions, the whole 
iterative procedure can be different, when several repetitions are 
performing. Of course, if a stable state is the goal, many restarts 
are needed, followed by an evaluation, as the model errors are 
checked and the current version is qualified. Having many 
results, the best can easily be found automatically [Barsi2008]. 
The implementation was written in Mathworks Matlab Version 
7.5 under the operating system Windows XP.