26
J.L. Giraudel . S. Lek
Then, each hexagon is coloured in different levels of grey according to the value
of the environmental factor. In Fig. 2.7, the soil texture has been displayed for the
upland forest data and a decreasing gradient of this factor can be seen from the left
to the right part of the map.
S 8
o
2
3
4
5
Figure 2.7. Soil texture for the upland forest dataset.
2.3.4
Clustering with a SOM
Visual inspection of the SOM allows some groups to be seen immediately: the
SVs in the same hexagon are reputed to be in the same cluster. By this way, 8
clusters are defined for the upland forest dataset (Figs. 2.5, 5): Al (SVs 1, 2), Au
(SV 3), Am (SV 4), Aly (SV 5), Ay (SV 6), A Vl (SV 7), AyU (SV 8), and A VlU (SVs 9
and 10). It can be noticed that these clusters are the same with the Euclidean
distance and the Whittaker's relative transformation.
By combining some hexagons, it is possible to form bigger clusters. But, a
deficiency of the initial SOM algorithm was the difficulty in detecting the cluster
boundaries on the map for units in different hexagons. A few enhancement
techniques have been proposed to tackle this problem, for instance, hierarchical
feature maps (Miikkulainen 1990) or adaptive coordinates (Merkl and Rauber
1997). However, these two methods are not considered further in this paper, the