Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 428

Fig. 7.23. PRSOM (10 × 10) map obtained using App cod1 as the training set. The
map representation displays the 2D topological ordering. Each square contains a
number that identifies the neuron, and the figure above each square is the number
of pixels of the training set that are assigned to that neuron

same map, and, for each neuron, the associated variance. It is clear that
scattered values of the reflectance are represented at the upper right corner
of the map, whereas the lower left corner contains representative of similar
spectra. Physical considerations provide an interpretation of the various zones
of the map:

• The spectra are more stable if the sky is clear and if the signal permits an

analysis of the ocean

• High variability zones may stem from atmospheric influence that is due to

aerosols or to cloud reflection

The hundred patterns associated to the neurons are characterized by spectra
that are of the same kind as the observations (vectors of

8 ). Figure 7.25 is
presenting for the first coding, the set of patterns and their topological con-
nections through the map. It is the same map than in Figs. 7.23 and 7.24 but
each neuron is now attached to its encoded pattern. Each encoded pattern
is an average spectrum over the allocated reflectance observations. The pat-
terns are organized according an order, which is visualized through the map.
The patterns that are associated to weak variance local densities have similar
shapes: the observation space is sampled there in a very fine way. Patterns
that belong to the high variance regions are sampling the observation space

R