Neural networks to retrieve sea surface salinity from SMOS brightness temperatures

One of the critical issues for the SMOS mission is the inversion algorithm which will be used to retrieve the sea surface salinity (SSS) from the SMOS brightness temperatures (TBs). Most of the scientists have chosen an inversion processing based on a forward model, developed with theoretical or semi empirical models. This forward model allows simulating brightness temperatures, given a triplet of geophysical parameters (sea surface salinity, sea surface temperature and wind speed). The iterative method tries to adjust the cost function between the SMOS measured brightness temperatures and the simulated ones. The development of this physical inversion method is necessary because it allows to take into account pixel by pixel the surface physics and especially to improve our knowledge in surface emissivity modeling. Nevertheless, one of the main drawbacks of such a method is the use of a forward model. We know that this model does not reproduce perfectly all the physics, especially the influence of the sea surface roughness (foam effect, swell...). All the studies performed with simulated data give errors which meet the GODAE requirements, but all the scientific community acknowledge that, once SMOS inflight, the retrieval algorithm based on a forward model will provide inaccurate salinities. The improvement of the salinity retrieval can be performed only by improvement of the physics modelling in the forward model. No strategy has been defined yet to perform this task, which needs a large amount of work and can not be achieved in reasonable delays in order to release SMOS products to the users, at the end of the commissioning phase. In this context, we propose to develop in parallel an empirical inversion algorithm, that will provide realistic SSS at the beginning of the mission. For that purpose, we propose an algorithm using neural network methods, built with SMOS measurements (TBs) co-located with in-situ salinities, during the commissioning phase. For pre-launch studies, the feasibility of neural network inversion is demonstrated with simulated data sets. This simulation phase is very important because it allows fixing the architecture of the network and helps to identify the critical issues to build a reliable algorithm. Once SMOS in-flight, the work will consist in two steps. First, the constitution of the learning database, using a suitable editing to make a representative dataset. In a second step, the coefficients of the algorithm will be updated, keeping the architecture and strategy defined during the pre-launch study. Another advantage of the neural formulation is to allow, on the same learning database (SMOS measurements colocated with in-situ salinities), the formulation of the “exact” forward model, to go from salinity to brightness temperatures. Unlike physical forward model, it takes into account all phenomena, inaccurately modeled (foam), or not modeled at all (swell) at that time, or even not identified (wiggles). The use of this neural forward model is of great interest to help the scientists to improve their understanding of the modelling errors in the forward physical model. The neural forward model tuned on SMOS measurements colocated with in-situ salinities will provide the real relationship to be related to the physics (assuming that all calibration errors on SMOS brightness temperatures are well corrected for). For example, it will provide the real dependence of the brightness temperatures on the wind speed. Thus, comparing this empirical model to the theoretical one will certainly help us to understand the defaults of the theory, in a straightforward way. In this context, the objective of this feasibility study was to develop a neural inversion for the subsatellite track, and to evaluate its performances in terms of sensitivity to the input brightness temperatures and to the auxiliary parameters (noise and bias).