Optimizing the Radiative Transfer Model Using Deep Neural Networks for NISAR Soil Moisture Retrieval

Soil moisture retrievals based on rigorous physical backscattering models require a comprehensive description of the vegetation structure and biophysical parameters, including the density of the scatters, height, and vegetation water content. Semi-physical models such as the water cloud model (WCM), are also extensively used and rely on estimates of vegetation water content or biomass derived from optical vegetation indices such as LAI and NDVI. However, such indices only contain parts of the true variability of vegetation structure and how it changes across various land cover types. In this study, we introduce RTNet (Radiative Transfer Neural Network), which combines a parameterized first-order Radiative Transfer (RT) model with four scattering components (surface, volume, double-bounce, and triple-bounce scattering components) and deep residual neural networks for the soil moisture retrieval. The input features consist of the HV backscattering coefficient, vegetation water content, and several other information categories such as soil texture and weather data. The RTNet is optimized to minimize the difference between the estimated and measured HH total backscattering. After imposing a physical constraint to the RTNet outputs, they are then applied to the ensemble random forest machine learning regressor to retrieve the volumetric soil moisture. The proposed framework is validated using the SMAPVEX12 L-band UAVSAR data, aggregated to a resolution of 100 meters, which is finer than the NISAR Level 3 soil moisture product (200 meters resolution). The estimated HH total backscattering coefficients show a high agreement with the UAVSAR measured HH backscattering with a root mean square error (RMSE) of approximately 3 dB across the entire image in non-forested regions. The retrieved volumetric soil moisture also shows a very high agreement with the in-situ soil moisture achieving the RMSE of 5.65% and R2 of 0.7.