Soil moisture downscaling using machine learning with scale-aware, physically informed feature constraints

Machine learning enables the downscaling of coarse-resolution remotely sensed soil moisture to finer spatial scales, yet most existing methods lack physical constraints and neglect scale-dependent feature effects. The study proposes a physics-guided machine learning framework that embeds scale-aware feature constraints into a random forest model. These constraints, derived from the Disaggregation based on Physical and Theoretical Scale Change algorithm, adaptively adjust the influence of topography and precipitation across spatial scales. The framework was applied to downscale the European Space Agency's Climate Change Initiative soil moisture product (25 km) to 1 km over the Naqu region of the Tibetan Plateau. Validation against in situ measurements and a 1-km benchmark products showed improved accuracy and stronger physical consistency compared with unconstrained models. Sensitivity analyses confirmed the benefits of integrating physical knowledge and revealed scale-dependent trade-offs in predictor influence. The framework is transferable to heterogeneous landscapes and offers a robust paradigm for generating high-resolution soil moisture products to support hydrological modeling, agricultural water management, and climate impact assessment.