Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 430

in a coarser way. The visualization technique enables to select some patterns
and then to specifically study the set of observations that are allocated to one
of these patterns or to locate them on the SeaWifs image.
A first investigation is useful to check the quality of the process (Fig. 7.25):
it is possible to identify wrong spectra with respect to the measurement
process. Actually, on this map, patterns that are associated to neurons
17,28,35 and 39 have a vanishing wavelength reflectance. If one gathers the
concerned information, it appears that they all present the same fault. It is
then possible to infer that in these cases a channel was defective and that
some neurons have specialized in detecting this fault. Figure 7.26 displays the
spectral patterns that are associated to neurons 17 and 35. and their variance.
One may perform a similar analysis for any of the 100 neurons of the
map. Figure 7.27 shows the spectrum that is associated to neuron 51, which
is located in a high-density zone. Then the set of allocated radiance spectra
is displayed as well as the associated geographical zone on SeaWifs image.
When it is compared with SeaWifs image of Fig. 7.21 one may notice that
neuron 51 controls a light colored zone located on the sea and for which there
is apparently neither desert aerosols nor clouds. When the ordering of spectra
that is proposed on Fig. 7.25 is inspected, one notices that the proposed
coding is governed by the spectral intensities. Thus the ordering favors the
emergence of underlying physical properties. The same experiments have been
performed using another encoding process that takes into account both the
intensity and the shape of the spectra (App cod2 ). Figure 7.28 shows the new
ordered patterns that were obtained. (On that figure, the patterns have been
decoded in order to show their original spectrum profile). The organization
of the neurons is now performed with respect to their intensities and their
shapes.
7.5.2 Classification and PRSOM
The first experiment group allowed us to assess the quality of the vector
quantizations, which were obtained using PRSOM. We shall use now these
quantizations to achieve classification tasks.
A first possibility was displayed in previous section. Recall it amounts to
study separately the physical property of each neuron pattern of the map.
This study has to be performed by an expert, who is able to recognize the
aerosol class and thus to label the patterns from their spectral properties. If all
the neurons are identified, then the partition that is obtained through the self-
organizing map enables to use it straight to classify the whole image SeaWifs.
Moreover, if the learning set is representative of the physical problem, it may
be used to label other SeaWifs images that share the same physical properties.
If this identification process is not possible, i.e. if the expert is not able to
label accurately every neuron of the map, it is possible to cluster the neurons
according to an unsupervised way. One can proceed as it was demonstrated in
the previous section “Classification and topological map,” by aggregating the