Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 422

• The first section describes a satellite remote sensing application. It is a

field of growing importance, and a lot of statistical problems must be
solved by physicists and research engineers who are in charge of designing
models. Considering that a very large amount of data is now available,
this field is particularly suited to neuronal modeling. That application will
fully illustrate the methodology that was described in previous sections.
It uses the probabilistic model of self-organizing maps (PRSOM).

• The second section gives a brief account of one of the most popular applica-

tions that was developed at UTH: the WEBSOM system. That is devoted
to information research on the Web. The earlier version was implemented
in 1995. The salient feature of that application is the high dimensionality
of the data. Dimensioning the topological map with a very large set of neu-
rons and tuning the algorithm (regarding computing time and convergence
accuracy) were the basic issues that were successfully faced at UTH. The
development of WEBSOM spurred research oriented towards shortening
the training phase for the design, and towards shortening the document
research time during the exploitation phase.

7.5.1 A Satellite Remote Sensing Application

A lot of data is generated by the observation of earth with on-board sensors,
and handed to geophysicists. All the neural methods that are presented in this
topic are helpful to process those data because they solve multidimensional
statistical problems. Among those methods, unsupervised training is espe-
cially useful, because it allows extracting information even when expert infor-
mation is scarce. Gathering expert information often requires costly analyses
(ground mission, sophisticated biological and chemical analyses). That ex-
plains why expert-appraised data is scarce as compared to the amount of
available satellite data.
Self-organizing bring valuable contributions to satellite data analysis, be-
cause estimating the observation probability density, and designing represen-
tative data partitioning, can be performed in a relatively straightforward way.
Such information provides new insights into the physical phenomena of inter-
est:

1. PRSOM estimate the variance and local uncertainty of the observations.
2. The partitions that are obtained are useful to the expert of the various
application fields (physicists, chemists ... ) because they may serve as an
accurate summary of the observation set. Investigating such a summary
may be crucial for understanding the phenomena of interest.
3. In all the fields that are concerned by experiments, heavy and expen-
sive experimental campaigns are carried out regularly. With respect to
the amount of satellite data, the expert-assessed observations are scarce,
but they contain extremely valuable information. A few expert-labeled
observations allow the identification of subsets of the partition from the