Building a Learning Database for the Neural Network Retrieval of Sea Surface Salinity from SMOS Brightness Temperatures

TL;DR

This paper improves neural network methods for retrieving sea surface salinity from SMOS brightness temperatures by optimizing the learning database to reduce regional biases and enhance accuracy.

Contribution

It demonstrates that carefully distributing geophysical parameters in the training data significantly reduces regional biases in neural network salinity retrievals.

Findings

Regional biases are minimized between 40°S and 40°N.

Standard deviation of retrievals remains between 0.6 and 1 psu.

Bias reduction is achieved through database equalization and new training techniques.

Abstract

This article deals with an important aspect of the neural network retrieval of sea surface salinity (SSS) from SMOS brightness temperatures (TBs). The neural network retrieval method is an empirical approach that offers the possibility of being independent from any theoretical emissivity model, during the in-flight phase. A Previous study [1] has proven that this approach is applicable to all pixels on ocean, by designing a set of neural networks with different inputs. The present study focuses on the choice of the learning database and demonstrates that a judicious distribution of the geophysical parameters allows to markedly reduce the systematic regional biases of the retrieved SSS, which are due to the high noise on the TBs. An equalization of the distribution of the geophysical parameters, followed by a new technique for boosting the learning process, makes the regional biases almost disappear for latitudes between 40{\deg}S and 40{\deg}N, while the global standard deviation remains between 0.6 psu (at the center of the of the swath) and 1 psu (at the edges).