Joining a Discrete Radiative Transfer Model and a Kernel Retrieval Algorithm for Soil Moisture Estimation From SAR Data

This paper investigates the problem of retrieving soil moisture under crops using Synthetic Aperture Radar (SAR) data. First, we simulated the time series of L-band SAR signals over agricultural fields using a discrete radiative transfer model (RTM). Full growth cycles of winter wheat, maize, and sugar beet fields sampled during the AgriSAR2006 ( Agricultural bio/geophysical retrievals from frequent repeat pass SAR and optical imaging ) field campaign were considered. A generally good correspondence between the simulated crop backscattering coefficients and those measured by the airborne L-band E-SAR ( Experimental-SAR ) system was observed with an average root-mean-square error (RMSE) of 2.32 dB. The highest RMSE of 3.63 dB was obtained by the RTM simulations of HV polarized signals in the wheat field, whereas the smallest RMSE of 1.63 dB is achieved in RTM simulations of HV backscattering coefficients in the field of sugar beet. All discrepancies were critically discussed and interpreted. Then, soil moisture was estimated using a nonlinear inversion technique, support vector regression ( $\nu$ -SVR). The model was trained with the backscatter model simulations obtained by the RTM. For all fields considered, the RMSE of the predicted soil moisture was smaller than 5.5% Vol. and the corresponding correlation coefficient ( $r$ ) was equal to or higher than 0.71.