Towards High-resolution Self-organizing Maps of Geographic Features - Geographic Visualization: Concepts, Tools and Applications - page 163

Figure 8.2 Mapping of 51 features onto low-resolution SOM (nine neurons), including disam-
biguated geometry through random placement inside winning neuron polygons
labelled at each neuron location and any further input vectors at that location remain basi-
cally invisible. One solution is to map features randomly near the respective best-matching
neuron (Skupin, 2002), as seen in Figure 8.2. Fifty-one geographic objects are here mapped
onto the same nine-neuron SOM. However, this is a purely graphic solution and, sim-
ilar to the mixed-pixel problem known in raster GIS, the model provides no means to
actually distinguish n -dimensional differences among vectors assigned to the same neu-
ron. In more general terms, one can say that a low-resolution SOM only allows visualizing
global or macro-structures existing in n -dimensional data, while finer structures remain
hidden.
Akin to resolution effects in raster GIS, the approach proposed here is to provide a larger
number of map units (i.e. neurons). For example, a 20-by-20 neuron SOM provides 400
different map units onto which the 51 geographic objects can be mapped (Figure 8.3). It is
important to note that such a SOM at this point stops functioning as a clustering method,
Figure 8.3 Mapping of 51 features onto higher-resolution SOM (400 neurons), with much reduced
need for disambiguation of geometry