Improved SMAP Soil Moisture Retrieval Using a Deep Neural Network-Based Replacement of Radiative Transfer and Roughness Model

The L-band (1.4 GHz) brightness temperature (TB) is interpreted by the zeroth-order approximation of the radiation transfer model, known as the $\tau $ - $\omega $ model. The soil moisture active passive (SMAP) mission facilitates the operational retrieval of global soil moisture (SM) through the $\tau $ - $\omega $ model. The operational SMAP SM retrieval algorithms demonstrate favorable alignment with the International SM Network (ISMN) and the SMAP core validation sites (CVSs). Nevertheless, the performance of these retrieval algorithms is reduced in densely vegetated areas or on rough surfaces, due to the smaller sensitivity of L-band TB to SM and less optimized parameterization. Therefore, this study proposed a deep neural network (DNN) model that can replace the currently used algorithms. The complex dielectric constant was simulated using ISMN data with the dielectric model to directly train DNN and determine the relationship between measured TB and retrieved dielectric constant. This involved establishing a correlation between the measured V- and H-polarized TB and the parameters from the SMAP SCA, i.e., surface temperature, vegetation water content (VWC), b, $\omega $ , and h. The challenge posed by the scale mismatch between point-based and SM data was effectively managed using the triple collocation analysis (TCA). The accuracy of the proposed model was evaluated using the ISMN and SMAP CVS and compared with the existing SM retrievals such as SCA-V, DCA, SMAP-IB, and the recently developed MCCA. The developed SM in this study demonstrated enhanced agreement with ISMN, SMAP CVS in situ SM data, and the Tambopata site located in the Amazon compared with existing SM retrievals. Moreover, by inversely tracking the developed DNN, we propose a novel method for parameterizing the $\tau $ - $\omega $ model that potentially improves the parameterization of the existing SMAP-based SM retrieval algorithms.