SMOS Near-Real-Time Soil Moisture processor. Part 1: Neural network evaluation and algorithm description

The best approach to retrieve soil moisture in Near-Real-Time (NRT) using neural networks (NNs) has been discussed using SMOS CATDS Level 3 brightness temperatures and Level 3 soil moisture (SM). NN retrievals have been first evaluated comparing the output SM to the L3 SM. The NN output has also been evaluated against in situ measurements over the SCAN network, the USDA ARS watersheds and OzNet. The recommended input configuration for the NRT SM processor is using SMOS T_b’s from 30^◦ to 45^◦ incidence angles in 5^◦ bins for both H and V polarizations, and a corresponding set of normalized indexes computed taking into account the brightness temperatures local extreme values and the associated L3 SM values. Finally, the input data should add the 0-7 cm soil temperature forecast by ECMWF. The recommended NN architecture is two layers with a hidden layer containing 5 non-linear neurons and an output layer with one linear neuron. This configuration is the best trade-off of retrieval performance and swath width (914 km). The recommended NN configuration for the NRT SM product has been specifically evaluated against the reference L3 SM data and against a large number of in situ measurements from the International Soil Moisture Network. Average statistics are somewhat better than those of the reference L3 SM data for most of the sites. In summary, the recommended NN configuration performs as well or better than the reference SM dataset but the retrieval can be done in Near-Real-Time after a global training phase. Finally, the recommended processor architecture is discussed.