A D-vine copula quantile regression approach for soil moisture retrieval from dual polarimetric SAR Sentinel-1 over vegetated terrains

Soil moisture retrieval from Synthetic Aperture Radar (SAR) over vegetated terrains requires an isolation of soil and canopy signals from observed backscatter (sigma degrees). This study develops a probabilistic soil moisture retrieval method from dual polarimetric C-band SAR Sentinel-1 (S-1) with uncertainty quantification at distinct vegetation covers (VCs). Both sigma(VV)degrees and sigma(VH)degrees were used to represent ground and volume scattering due to the high respective sensitivity to soil moisture and vegetation dynamics. A novel D-vine copula quantile regression (DVQR) was adopted to provide the soil moisture estimates based on modelling trivariate dependence of sigma(VV)degrees-sigma(VH)degrees -soil moisture anomalies (VV-VH-Mv), with a support from the innovative cosmic-ray soil moisture as ground-truth data. The feasibility of DVQR was underlined for: (1) multivariate nonlinear dependence structure modelling and (2) soil moisture retrieval with associated uncertainty. An inter-dependence analysis, which assesses the correlations among three major variables, indicated that the dependence between each pair of variables decreased as canopy density increases from herbs to forests, mainly due to the ci degrees attenuated by vegetation effect. The dependence structures simulated from the D-vine copula revealed highly nonlinear and asymmetric shapes with tail dependences occurred in most VCs, which can be well captured by different associated Archimedean copulas. Soil moisture anomaly (Mv) estimated using the DVQR and Multiple linear quantile regression (MLQR) were compared against ground-truth data for both in-sample and out-of-sample predictions. Superior performances of the DVQR in most VCs, with 10% and 16% improved in RMSE at grasslands and broadleaf forests, respectively, demonstrated the robustness of this method for S-1 soil moisture retrieval due to the highly nonlinear dependence structures captured by the D-vine models. Over VCs, better performances were obtained at low-canopy herbaceous regions (grasslands and croplands); whereas extremely dry conditions and complex structures in shrublands and dense forests resulted in inferior performances. A sensitivity analysis was then conducted to evaluate the change in My estimation accuracy given distinct VV and VH quantile levels. Result underlines that VV is the primary factor controlling the retrieval accuracy, but the increase in VH level also contributes to higher errors in soil moisture estimation, especially under wet conditions.