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

Soil moisture retrievals based on rigorous physicalbackscattering models require a comprehensive description ofthe vegetation structure and biophysical parameters, includingthe density of the scatters, height, and vegetation water content.Semi-physical models such as the water cloud model (WCM), arealso extensively used and rely on estimates of vegetation watercontent or biomass derived from optical vegetation indices suchas LAI and NDVI. However, such indices only contain parts ofthe true variability of vegetation structure and how it changesacross various land cover types. In this study, we introduceRTNet (Radiative Transfer Neural Network), which combines aparameterized first-order Radiative Transfer (RT) model withfour scattering components (surface, volume, double-bounce, andtriple-bounce scattering components) and deep residual neuralnetworks for the soil moisture retrieval. The input featuresconsist of the HV backscattering coefficient, vegetation watercontent, and several other information categories such as soiltexture and weather data. The RTNet is optimized to minimizethe difference between the estimated and measured HH totalbackscattering. After imposing a physical constraint to theRTNet outputs, they are then applied to the ensemble randomforest machine learning regressor to retrieve the volumetric soilmoisture. The proposed framework is validated using theSMAPVEX12 L-band UAVSAR data, aggregated to a resolutionof 100 meters, which is finer than the NISAR Level 3 soilmoisture product (200 meters resolution). The estimated HHtotal backscattering coefficients show a high agreement with theUAVSAR measured HH backscattering with a root mean squareerror (RMSE) of approximately 3 dB across the entire image innon-forested regions. The retrieved volumetric soil moisture alsoshows a very high agreement with the in-situ soil moistureachieving the RMSE of 5.65 % and R2 of 0.7.