Self-Organizing Maps and Unsupervised Classification - Neural Networks: Methodology and Applications - page 426

Fig. 7.22. Expert-processed GAC image. The picture represented five zones whose
boundaries had been defined by the expert: the aerosols coming from the desert
( black ), the case-2 waters ( light grey ), the marine aerosols ( dark grey ), the clouds
( medium grey ) and the earth ( white )

7.5.1.3 The Role of Encoding

Numerical experiments were performed using different encodings of SeaWifs
spectra.
The first encoding uses directly the reflectance at the top of the at-
mosphere. In order to normalize the influence of the wavelengths, the re-
flectance spectrum values were normalized to the [

1 , +1] interval. The re-
flectance for wavelength λ k , is denoted ρ ( λ k ), the normalization has been
computed from the learning set for each wavelength, ( k =1 ,..., 8). Thus,
each observation was encoded by an 8-dimensional vector, each component of
that vector being the normalized reflectance for a given wavelength. Since the
reflectance values ρ ( λ k ) are numbers between 0 and 1, they were normalized
between

−

1.
The sampled LAC image that was used for training (actually one picture
line out of ten), after coding as described above, will be denoted by App cod1 .
A second coding was performed, in order to highlight the shape of the spectra
of interest. To that effect, the slopes of spectra were taken into account. They
were computed for each wavelength. Thus, the k th component of the new
coding vector was computed from the reflectance as

− 1 and 1, using to 2 ρ ( λ k ) −