Coupling semi-empirical and machine learning model in high-resolution remote sensing soil moisture retrieval

As a critical parameter of the Earth surface system, surface soil moisture (SSM) plays a pivotal role in investigating the water cycle and land-air interaction. Synthetic aperture radar (SAR)-based active microwave remote sensing offers an effective method for acquiring high-spatial-resolution SSM data. In high-resolution SSM retrieval studies, retrieval based on physical or semi-empirical physical models follows physical mechanisms, and machine learning models-based retrieval has strong learning and nonlinear modeling capabilities for multi-source datasets. Nowadays, the retrieval study of coupling physical mechanisms and machine learning has attracted much attention. To address the challenges and opportunities in high-spatial-resolution SSM retrieval studies based on SAR, we summarized multiple fusion models in this study, which were classified into three categories: complementary fusion model, predictive fusion model, and constrained fusion model, according to the relative importance of machine learning models and physical mechanisms in coupling. Several specific retrieval models for high-resolution SSM retrieval were designed based on the categories, and various comparative assessments of these models were carried out across multiple study areas. Evaluations revealed that the differentiable retrieval model, which falls under the constrained fusion model category, exhibited robust retrieval performance and spatiotemporal generalization capacity, with the highest R² values of 0.853 and the lowest ubRMSE values of 0.041 m³·m⁻³ within the study areas. It also demonstrated excellent retrieval performance under the forest cover type. The design and comparative evaluation of various fusion models in high-resolution SSM retrieval provide valuable references for related studies and offer insights for developing a series of new application modes of fusion models in the future.