Synergistic Use of Sentinel-1, SMAP, and Ancillary Data for High-Resolution Soil Moisture Mapping

Accurate, field-scale soil moisture estimates are critical for hydrological modeling and agricultural monitoring, yet they remain difficult to obtain from any single satellite system. L-band radiometers such as SMAP provide reliable large-scale soil moisture retrievals, but their coarse spatial resolution limits local applicability. In contrast, C-band SAR observations from Sentinel-1 offer fine spatial detail, though their sensitivity to surface roughness and vegetation requires careful calibration.

We develop a multi-sensor downscaling framework that combines SMAP morning soil moisture with ascending and descending Sentinel-1 VV/VH backscatter. To better represent soil moisture dynamics, historical VV backscatter minima and maxima are used to derive a Soil Moisture Index (SMI), alongside NDVI to account for vegetation. These variables are complemented by the Antecedent Precipitation Index (API) and evapotranspiration to consider surface water fluxes, as well as topographic information from a digital elevation model.

Model robustness is evaluated using a strict temporal split: data from 2020–2024 are used for training, while 2025 is reserved as an independent test year. Three non-parametric algorithms—Random Forest, XGBoost, and K-Nearest Neighbors—are assessed against in-situ measurements from the International Soil Moisture Network.

Including meteorological information and historical backscatter features leads to consistent performance gains across models. On the independent test set, coefficients of determination exceed 0.5, with XGBoost achieving the lowest RMSE and outperforming both Random Forest and KNN. These results demonstrate the value of combining complementary satellite observations and targeted feature engineering for reliable, high-resolution soil moisture mapping.