Statistical approaches to assimilate soil moisture information: Methodology, first assessment and future plans

Land surfaces are characterized by strong heterogeneities of soil texture, orography, land cover, soil moisture, snow and other variables. These are very challenging to represent accurately in radiative transfer models which have currently a still limited reliability over land. In this study, we compare two statistical modeling approaches: the traditional CDF-matching used routinely in NWP centers (used here as a normalization and as an inversion technique), and the Neural Network (NN) methods. NNs and CDF-matching are compared and combined. Two cases are considered: (1) the more traditional inversion scheme, and (2) the forward modelling that could be attractive for assimilation purposes. It is shown that in the context of ASCAT, the inversion approach seems better suited than the forward modelling but this could be different for another type of observations. It is also shown that it is possible to combine the global model obtained using the NN and the localized information of the LSM offered by the CDF-matching. A first assessment is performed over the USA using in situ soil measurements. Localization strategies for the NN models are introduced. Another necessity for the use of NN in an assimilation framework are estimations of NN uncertainties: this is unfortunately not available so far and we propose several schemes in order to obtain them. Finally, we will present future plans to develop a forward operator for low-frequency microwave channels (SMOS, AMSR-E, SMAP, CIMR) based on a statistical modeling of surface emissivities over continental, snow-ice and sea ice surfaces.