Inversion of NOAA-AVHRR satellite images by a classification process using neuronal networks

The classification of pixels making up a satellite image can be seen, not only in view to cluster discriminate these pixels, but exactly as an inversion procedure. The original technique we implemented is built with a combination of several self-organized neuronal networks; in order to be particularly robust. a sensitivity study, our procedure was applied to the extraction of cloud optical thickness and dropplets effective radius from AVHRR imagery. The method validation was conducted in well documented cases issued from the EUCREX campaign by comparisons against the CRTVL inversion code of Nakajima and King (1990) and also in-situ measurements.