Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 423

topological map. The classification methodology that is presented below
in the section Classification and PRSOM allow that identification.
4. In order to demonstrate the various capabilities of self-organizing maps,
the presentation of the next application is organized as follows:
• Description of the application field, of the problems of interest, and of

the available data;

• Description of experiments that allow the understanding of the influ-

ence of data encoding on the partition and on the topological order
that is obtained;

• Description of experiments that allow the assessment of the impact of

expert knowledge

7.5.1.1 The Color of the Ocean

The biological activity of the ocean is crucial for the natural economy of
the earth, for it is strongly connected to the fishing resource, and is part
of the biochemical cycles with strong climate consequences. During the past
few years, several multi-spectral sensors dedicated to measure the color of
the ocean were launched on different satellites ((MOS, POLDER-1, OCTS,
SEAWIFS, MODIS) or will be in the future (MERIS, POLDER-2, GLI). They
are expected to estimate the chlorophyll contents of the upper layer of the
ocean, and to assess its space and time variability.
Two steps are needed to recover the pigment fields from the ocean color
satellite data. The first step deals with atmospheric correction (indeed, the at-
mosphere contributes more than 80% of the measured signal); the second step
is dedicated to ocean interaction (see Fig. 7.20). The atmospheric correction
algorithm currently computes sea-level reflectance by canceling atmospheric
effects (aerosols and air molecules). The second step aims at inverting that
reflectance to provide the chlorophyll pigment concentration. This is a tricky
issue, because one has to take into account both uncertainties that arise from
atmospheric corrections, and uncertainties that are intrinsically related to the
variation of biological population. Thus, the knowledge of the atmospheric
aerosol composition and of the water constituent concentration is crucial. The
following experiments aim at assessing various classes of aerosols and waters
from the top of atmosphere spectra, which are provided by the sensors.

7.5.1.2 The Data

The data that are used hereafter were provided by US radiometer SeaWifs,
which was located on the SeaStar satellite. That sensor has eight spectral
bands in the visible and the near infrared spectrum (see Table 7.1).
For each measurement location at the ocean surface, the observation vector
is 8-dimensional. Its components are the eight radiances that are measured
on the top of atmosphere. The following results are a representative example