Integrating multi-source physical knowledge into deep learning for satellite soil moisture downscaling

Fine-resolution soil moisture (SM) is essential for understanding land–atmosphere interactions and managing hydrological hazards at local scales. Passive microwave remote sensing provides reliable SM observations but is limited by its coarse spatial resolution. Spatial downscaling offers an important means of addressing this limitation. However, many existing downscaling methods either rely on physical models with over-simplifications of soil hydrological knowledge, or data-driven models (particularly Deep Learning, DL) that lack interpretability, while current physics-guided DL approaches often emphasize a limited number of physical relationships or constraints. In response, this study proposes a framework to integrate multi-source physical knowledge into DL for satellite SM downscaling. Specifically, the framework incorporates physically informed drivers, a DisPATCh-informed physical regression branch, a hydrologically motivated consistency loss, and a constrained SM boundary loss, while using a gating module to adaptively fuse physically guided and data-driven representations. Evaluations based on in-situ and aircraft SM observations, together with coarse-scale consistency checks against SMAP SM, show that the proposed framework can effectively downscale SMAP SM from 9 km to 500 m. Using region-specific models across ten environmentally diverse regions, MPhy achieved station-averaged Pearson’s R and ubRMSE values of 0.76 and 0.05 m3/m3, respectively; the corresponding values for the pure-DL baseline Dens were 0.68 and 0.06 m3/m3. Detailed experiments in WGEW further showed that MPhy generally performed better than the pure-DL baseline, DisPATCh, single-physics-informed models, and selected conventional machine-learning methods. Overall, the results support improved physical plausibility and convergence efficiency, as well as framework robustness across the evaluated environments.