Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 435

neurons into classes and trying to label the classes that were obtained through
hierarchical classification. In order to illustrate as best as possible the quality
of results when PRSOM and BUHC (Bottom-Up Hierarchical Classification)
are sequentially processed, two experiments of different complexity are pre-
sented:

• The first experiment is relative to the determination of a mask that is able

to detect the thick clouds and to discriminate them from other spectra. It is
known that clouds are strongly reflecting the signal: the top of atmosphere
signal that are registered by the satellite sensors are presenting stronger
and more variable intensities than when sea or aerosols are concerned. The
discrimination between thick clouds and other constituents amounts to
build a binary classifier. Since clearly distinct properties in the observation
set physically characterize this problem, the two classes that are searched
have to be fully separated.

• In the second experiment one tries to recover the five classes, which have

been identified by the expert; these classes have been determined by com-
paring the data with aerosol physical models. Actually, the number of
classes is higher and the expert has possibly introduced a lot of mistakes,
so this problem is far more complex.

The two vector quantizations that have been obtained using PRSOM will
be used to recognize the expert-identified classes. Class determination will be
performed through bottom-up hierarchical classification using the Ward index
that has been previously defined in the paragraph Looking for a partition
that suits the classes of interest .
In the first experiment, bottom-up hierarchical classification is performed
on the PRSOM 10 10 map obtained from App cod1 . Since the searched clas-

sification is a binary classification to select thick clouds, clustering has been
pursued up to the obtention of two classes. Figures 7.29 and 7.30 show the clas-
sifications that were obtained on the topological map and on the image. The
visualization of the map enables to observe the neurons of each class. Clearly,
the associated zones are well connected on the map. This classification has
been compared with the expert-based classification by computing the confu-
sion matrix. Here, SeaWifs provided the expert-based classification since the
cloud mask is available. The confusion matrix is represented on Table 7.2; it

×

Table 7.2. Confusion matrix that compares the SeaWifs labeled classification and
the classification that was obtained from PRSOM + BUHC. PRSOM was obtained
from Appcod1 and BUHC uses Ward index

PRSOM + BUHC

Clouds Apparent sea

SeaWifs-labeled clouds 0.91 0.09