Spatiotemporal Downscaling of SMAP Soil Moisture With an Integrated Machine Learning and Land Data Assimilation Framework

Soil moisture is essential for precision agriculture, flood prediction, and irrigation management. Current satellite-based soil moisture datasets generally suffer from coarse spatial resolution, and substantial efforts have been devoted to generating high-spatial-resolution soil moisture products. Machine learning methods lack physical constraints, while high-resolution land surface process simulations are limited by inaccurate surface parameterization. This study proposes a novel framework integrating machine learning and a land data assimilation system (LDAS) to downscale coarse-resolution satellite soil moisture products. The framework combines the advantages of statistical and physical models and improves the spatial resolution and temporal variability of soil moisture active passive (SMAP) soil moisture by incorporating multisource high-resolution surface variables. First, the random forest (RF) model is adopted to characterize the complex nonlinear relationship between coarse-resolution SMAP data and land surface variables. The trained RF model subsequently predicts soil moisture at a 0.05° × 0.05° resolution using fine-resolution inputs. Second, the RF-based downscaled soil moisture results are assimilated into the LDAS to optimize system parameters, yielding high-accuracy soil moisture datasets with improved spatiotemporal resolution. The integrated downscaling framework is validated against in situ measurements from three soil moisture monitoring networks across the Tibetan Plateau. The proposed framework outperforms traditional in situ-based machine learning downscaling methods, with a ubRMSE lower than 0.04 m3 m-3, an R higher than 0.7, and an MBE lower than 0.03 m3 m-3. This study provides an effective integration strategy of machine learning and LDAS, which offers great potential for generating global high-resolution soil moisture products.