Search for a command to run...
The aim of the current study was to select and test the appropriate model and input parameters for remote sensing retrieval of surface soil moisture (SSM) in the case of bare Chernozems on flat and sloping terrains in northern Bulgaria under different tillage systems. Normalized synthetic aperture radar (SAR) measurements from Sentinel-1 C-band dual-pol products (Gamma-Nought in VV, ratio) were utilized in two ways to delineate SSM from environmental factors that bias determination. The accuracy of the obtained SSM prediction was evaluated against ground-based volumetric water content (VWC) measured in the 0–3.8 cm soil layer at multiple points using a TDR meter. The TDR VWC data were preliminarily calibrated against gravimetric measurements in the 0–5 cm soil layer. The obtained data for soil water retention curves in all studied variants were used to determine the range of soil moisture variation. The measured ground-based data for surface roughness generally correlate with the co-pol Gamma-Nought in VV. The data modeled with the surface soil moisture script in Sentinel Hub (SSM-SH) was calibrated using the ground-based data. Incidence angle normalization of Sentinel-1 products improved the relationship between SAR observables and SSM, when expressed as the ratio of soil moisture to total porosity (rVWC). The modeling indicated the highest importance of the optical indices, together with the temporal differences of radar descriptors sensitive to variations in soil moisture over time. Although the applied Random Forest Regression (RFR) model achieved higher accuracy during training (nRMSE of 7.27%, R2 of 0.86), the Gaussian Process Regression (GPR) model provided better generalization performance on the independent validation dataset. The results proved the advantages of the joint utilization of temporal Sentinel-1 SAR measurements with Sentinel-2 optical acquisitions to determine SSM in different bare soil conditions for achieving high accuracy.