Bayesian Ocean Colour Modelling

Abstract

Current atmospheric correction algorithms are based on physical models and pixel wise retrieval. The goal of this thesis is to build a Bayesian framework using a probabilistic approach to enable the use of priors for the joint retrieval of ocean and aerosol parameters on case I waters. Simulated data containing ocean and aerosol parameters as well as the corresponding top of atmosphere information and its components will be used to train neural networks able to output the top of atmosphere components given the ocean and aerosol parameters. A Bayesian framework will be built to enable the retrieval of the ocean and aerosol parameters from the top of atmosphere information using the neural networks previously trained and priors which will be designed according to biological and physical knowledge. The Bayesian framework will then be tested on small problems, parts of the global retrieval problem, and then on the global retrieval problem, on simulated data at first, and secondly on real data.