Use of a Neurovariational Inversion for Retrieving Oceanic and Atmospheric Constituents from Ocean Color Imagery: A Feasibility Study

Abstract

This paper presents a neurovariational method for inverting satellite ocean-color signals. The method is based on a combination of neural networks and classical variational inversion. The radiative transfer equations are modeled by neural networks whose inputs are the oceanic and atmospheric parameters, and outputs the top of the atmosphere reflectance at several wavelengths. The procedure consists in minimizing a quadratic cost function that is the distance between the satellite-observed reflectance and the computed neural-network reflectance, the control parameters being the oceanic and atmospheric parameters. First, a feasibility experiment using synthetic data is presented to show that chlorophyll-a can be retrieved with an error of 19.7% when the atmospheric parameters are known exactly. Then both atmospheric and oceanic parameters are relaxed. A first guess for the atmospheric parameters was provided by a direct inverse neural network whose inputs are at near-infrared wavelengths. Sensitivity experiments showed that these parameters can be retrieved with an adequate accuracy. An inversion of a composite SeaWiFS image is presented. Optical thickness and chlorophyll-a both give coherent spatial structures when a background term is added to the cost function. Finally, chlorophyll-a retrievals are compared with SeaWiFS product through in situ data. It shows a better estimation of the chlorophyll-a with the neurovariational inversion for the oligotrophic regions.