SMOS is the first dedicated soil moisture instrument in space, using a novel retrieval approach based on interferometric radiometry at L-band microwave frequencies.

The nominal SMOS soil moisture retrieval algorithm is based on physical models, aiming at minimising the difference between the actual measured brightness temperature and its modelled counterpart by adjusting surface parameters (such as soil moisture and vegetation water content) in a physical radiative transfer model. Such an approach relies on good knowledge of the surface characteristics (land use, soil texture etc.) as well as good estimates of some variables (e.g. surface temperature). The approach is robust and can be improved as modelling knowledge increases. At the same time, however, the method is computationally intensive and relies on many auxiliary data files. Further complications arise due to the presence of RFI (Radio Frequency Interference).

In this context, the present project proposes to test alternative, statistical methods for SMOS soil moisture retrieval, using neural network-based approaches (NN) applied to data gathered since the launch of the mission as a training set. Intercomparisons of the different NN-based approaches will be performed, together with testing against other datasets, such as ASCAT-based soil moisture and in-situ measurements. Additionally, a novel joint SMOS/ASCAT dataset will be developed. The work might also result in better detection of RFI in the SMOS brighness temperatures.