The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
4. RESULTS 
The growing neural gas has been evaluated on aerial image and 
a satellite image. The former one could find most of the streets, 
but two false alarms are also produced: there are two unreal 
blobs (disjunct graphs), which aren’t part of the street network. 
Other image parts (like the main roads) could be found in 
excellent quality. This result has proven the ability of this 
cutting-edge technology not only on artificial image patterns, 
but also real photographed images. 
As one can notice, the detected road network needs some 
manual correction, but the general quality is good. The founding 
is strongly correlated with the existing data points; that’s why 
several road segments are mapped so dense. Its counterpart can 
also be seen: in the upper right comer of the image the arc 
hasn’t been detected exactly, because there was a lack of 
adequate data (the drawback effect of using simple image 
processing as prior step!). 
Figure 3. Growing Neural Gas has produced this road network 
based on simple image operations 
The last test contained the digital satellite imagery of Quickbird. 
This example was the worst of all tested one. The reason could 
be the low contrast, so the exact data pixels weren’t found by 
the previously executed processing. The road network is only 
partly reconstructed, although the two connecting roads are 
motorway and a rural main road. 
Figure 4. High-resolution satellite image analysis by growing 
neural gas. Godollo Hills, Hungary 
One possible reason can also be the relatively high number of 
data points: the image cutoff was 5884 x 5241 pixels, where the 
data set was established by simple thresholding. (This operation 
was kept during the whole testing, because it is very simple, and 
the most relevant advantages of the connecting methods can be 
seen so.) 
5. CONCLUSION 
The self-organizing techniques can also be implemented by 
neural networks. The neural self-organizing methods have three 
main representatives: 
• SOM, where a general winner-takes-all rule controls 
the neurons’ weight (in our geometric case, the 
position) 
• SONG, where the SOM technique has been 
generalized by the combination of SOM 
computational model with undirected cycle free 
graphs and the topology is fixed 
• GNG, where the network topology is being created 
during the training. 
The SOM methodology has long “history” in data analysis and 
computer science, also in data mining, which can inspire the 
image analysis efforts. The parameterization of SOM networks 
can be quickly set, the computation is really fast. These 
advantages forecast further photogrammetric and remote 
sensing applications. 
SONG is a special SOM implementation, where a prior 
knowledge or at least hypothesis is needed. This can be taken 
from a manual evaluation of the first image in a series, and then 
the next images can be analyzed and evaluated automatically. 
Initial hypothesis can be taken from practical point of view: e.g. 
searching for image object having fixed topology can 
successfully supported by the SONG algorithm. Because it is 
the generalization of the SOM technique, its software 
representation can manage also the SOM usage. 
GNG has the most freedom in topology, because this self 
organizing method varies not only the neuron positions, but also 
their relations, their connection. This flexibility means that the 
gas will produce in all cases an output, but it shall be tested and 
evaluated critically. The last Quickbird application has the 
message that pure novel neural methods cannot serve with high 
quality maps. Of course, if enough suitable data points are 
available (like in Fig. 3), this technique can figure out the 
possible data structure (in this case the road network) based on 
few parameters (like number of neurons and epochs, connection 
age). 
Future experiments are planned with the combination of 
sophisticated image analyzing techniques, like image 
enhancements and filtering, special edge detection operators, 
image classifications [Duda2001, Gonzalez2004] and then the 
grouping can be replaced by the presented self-organization 
technology. 
All of the described neural computational models were applied 
in two dimensions, but the distance calculation phase can be 
much general: it can be interpreted and therefore used in n- 
dimensional space, e.g. in considering multispectral or even 
hyperspectral image information.