Reduction of Irrelevant Features in Oceanic Satellite Images by means of Bayesian Networks

Abstract

This paper describes the use of Bayesian networks for the reduction of irrelevant features [1,2] in the recognition of oceanic structures in satellite images. Bayesian networks are used to validate the symbolic knowledge -provided by neuro symbolic or HLKPs (High Level Knowledge Processors) nets- and the numeric knowledge. This provides an automatic interpretation of images. The main objective of this work is the construction of an automatic recognition system for processing AVHRR (Advanced Very High Resolution Radiometer) images from NOAA (National Oceanographic and Atmospheric Administration) satellites to detect and locate oceanic phenomena of interest like upwellings, eddies and island wakes. With this aim, this paper reports on a methodology of knowledge selection and validation. In knowledge selection, filter measures are used. For knowledge validation, Bayesian networks (Naïve Bayes, TAN and KDB) are evaluated.