Enhancing Soil Moisture Statistical Retrieval from SMOS Using Partial Convolutions and Localization Strategies

Abstract Neural networks have been used for the retrieval of soil moisture (SM) from microwave observations over the last 20 years. The exploitation of the SM and Ocean Salinity (SMOS) observations has largely benefited from such statistical models. However, these retrievals are currently done at the pixel level, ignoring spatial context and using a constant incidence angle configuration approach. While pixels are seen with a varying number of angles from the SMOS instrument, only pixels monitored by a fixed preselected angle configuration are considered. These two limitations can have a negative impact on the quality of the retrievals. This paper introduces a new neural network (NN) architecture that combines two powerful innovations. First, the new model is image based: It ingests the entire SMOS orbit swath and thus leverages the strong spatial pattern present in the satellite observations. Second, a “partial convolutional layer” is tested. It allows being flexible, in the retrieval, on the angle configuration: More incidence angles can be exploited when they are available. Finally, a concept called “localization” is also exploited, helping the NN retrieval to adapt its behavior to specific local conditions. Experiments are conducted at the SMOS orbit scale over the contiguous United States (CONUS) region using 5 years of SMOS data (2016–19). A temporal correlation of 0.74 (unitless) with respect to ERA5 reanalysis and 0.62 with respect to in situ SM measurement network is obtained (to be compared to, respectively, 0.63 and 0.60 with the legacy pixel and fixed angle based approach). Furthermore, the use of partial convolutions results in enlarging the retrieval domain by +240% versus legacy retrieval and by +140% versus the operational SMOS level-2 product.