Large-Scale Soil Moisture Retrieval under Limited In Situ Data Coupling Physical and Machine Learning Model

Large-scale surface soil moisture data are critical for hydrological and climatic studies at large regional scales. The accuracy of large-scale soil moisture retrieval relying solely on physical models is constrained by model complexity and inaccurate parameters. Nowadays, machine learning models are widely used, but their excellent retrieval performance depends heavily on extensive accurate labeled data and faces criticism for lacking physical interpretability. Using in situ data as labeled data can enhance the accuracy of retrieval models. However, current soil moisture sites are predominantly concentrated in some key regions, and it is challenging to perform high-quality soil moisture retrieval in regions where sites are sparse. Facing the above challenges, this study proposed a fusion model utilizing limited in situ data to achieve high-accuracy soil moisture retrieval on a large regional scale. Based on the SMAP SCA-V algorithm, the retrieval model employed a differentiable modeling approach, integrating physical models like the τ–ω model, the Q–H model, the Fresnel equation, and the Mironov mixing dielectric model with the neural networks. This integration ensured model accuracy and improved generalization, achieving high-accuracy soil moisture retrieval across China at a 1-km resolution with labeled data from a limited number of sites. The differentiable retrieval model demonstrated strong performance in the Shandian River Basin and Naqu study areas, with R of 0.906 and 0.927, respectively, and attained an R of 0.925 and an ubRMSE of 0.035 m 3 ·m −3 in the overall evaluation. In the comparative analysis with the SMAP product, the differentiable retrieval model demonstrated comparable spatial distribution characteristics and effectively captured the temporal trends of soil moisture variation. The differentiable retrieval model creates the conditions for high-accuracy soil moisture retrievals in countries and regions with a small number of soil moisture monitoring sites or publicly available in situ data.