Large-Scale Soil Moisture Retrieval in Complex Terrain: A Physics-Informed Deep Learning Approach

Kurzfassung

Monitoring soil moisture in Alpine regions is essential for understanding hydrological dynamics, drought events, and the impacts of climate change in mountain environments. However, reliable monitoring remains challenging due to strong spatial heterogeneity, complex topography, sparse in situ observations, and rapidly changing land-surface conditions. While spaceborne Synthetic Aperture Radar (SAR) offers a valuable opportunity for large-scale soil moisture monitoring, the influence of vegetation, surface roughness, local incidence angle, and land-cover variability makes the direct soil moisture retrieval particularly challenging, especially in complex mountain regions where the SAR signal is also affected by geometric and radiometric distortions. As a result, current SAR-based algorithms for soil moisture estimation often have limited accuracy or are not readily applicable in mountain regions. In this study, we present HydroSAR, a physics-informed deep-learning framework for large-scale soil moisture retrieval tailored to heterogeneous and topographically complex terrain. The framework integrates three complementary components: spatially aware deep-learning methods to exploit spatial context and dependencies in SAR and ancillary observations; hydrological-model information to provide physically meaningful soil moisture patterns; and semi-empirical SAR backscatter modelling to explicitly represent interactions among the radar signal and soil and vegetation. HydroSAR consists of two interconnected branches: a hybrid forward-modelling framework for SAR backscatter simulation and a physics-based inversion framework for soil moisture retrieval from individual SAR acquisitions. In the forward branch, a deep-learning network estimates spatially distributed, physically constrained static model coefficients, while temporally varying soil moisture and vegetation information are provided by the WFLOW hydrological model and an LAI climatology. The semi-empirical backscatter model is embedded within the training framework, and its parameters are calibrated by minimizing the discrepancy between simulated and observed Sentinel-1 VV backscatter time series. The calibrated coefficient maps are then used in the inverse physics-based retrieval branch, where the semi-empirical model is analytically inverted to estimate soil moisture from an individual Sentinel-1 VV acquisition and the corresponding vegetation information. This two-stage configuration separates the computationally intensive calibration step from the lightweight retrieval step, enabling soil moisture estimation from single SAR acquisitions. The framework is developed using ascending and descending Sentinel-1 OPERA RTC time series over the Alpine arc (5.6-12.7°lon, 44.0-48.0°lat), covering approximately 147612.70 km2. WFLOW simulations at 6-hour temporal resolution are matched to the SAR acquisition times, while ancillary information on topography, acquisition geometry, land cover, vegetation, soil properties, and Sentinel-1 VH backscatter is incorporated to support parameters estimation. Preliminary results show a RMSE = 1.6 dB and a correlation coefficient r = 0.7 for VV backscatter reconstruction at 100 m spatial resolution. For soil moisture retrieval, RMSE = 0.07 m³ m⁻³, ubRMSE = 0.06 m³ m⁻³, and r = 0.4 are obtained when evaluated against in situ measurements at 0–2 cm depth from stations located at the Long-Term Socio-Ecological Research site in Val Mazia, South Tyrol.

The HydroSAR framework was developed within the ESA-funded “SMEM: Soil Moisture Ensemble Method” project.