Soil moisture retrieval model from satellite remote sensing under vegetation cover based on coupled radiative transfer and ensemble learning

Soil moisture (SM) is a key parameter in agricultural water management and drought monitoring, and accurately estimating SM is crucial for promoting sustainable agricultural development. However, under vegetative cover conditions, the interference of the vegetation canopy and the limitations of optical and microwave remote sensing data pose challenges to the precision of SM remote estimation. To address the issues of low SM estimation accuracy under vegetative cover and insufficient temporal and spatial resolution of ground remote sensing, this study focuses on the Xinjiang Production and Construction Corps' 8th Agricultural Division, coupling radiative transfer models with ensemble learning to propose a high-resolution soil moisture inversion method based on spectral scale conversion. The method first converts Sentinel-2 MSI data into leaf-scale reflectance using the 4-SAIL model, and then employs the PROSDM model to achieve the conversion between field-measured leaf bidirectional reflectance spectra and hemispherical reflectance spectra, validating the leaf reflectance derived from Sentinel-2. Finally, the leaf equivalent water thickness (CW) is estimated using the PROSPECT-5B model, and SM is estimated by combining the water cloud model framework with the Stacking ensemble algorithm. The results indicate that: (1) In the conversion and validation of canopy-leaf reflectance, the first derivative of the measured leaf bidirectional reflectance shows a good fit with the hemispherical reflectance (R²=0.8136), and the validation accuracy of the leaf reflectance derived from Sentinel-2 reaches R²= 0.9408 and RMSE= 0.0232; (2) The integrated algorithm based on radiative transfer models and the water cloud model outperforms machine learning models in SM inversion accuracy, with a coefficient of determination R² of 0.707 for the validation set, improving by an average of 49.78% compared to machine learning models; (3) The mixed model constructed by integrating Sentinel-1 radar backscatter coefficients, Sentinel-2 multispectral, and hyperspectral data can elucidate the interaction between the vegetation canopy and soil, validating the complementary potential of multi-source remote sensing data for SM inversion under complex vegetative cover conditions. This study emphasizes the potential of physics-based inversion methods at fine temporal and spatial resolutions, providing new technical means for real-time monitoring of soil moisture in vegetated areas and sustainable management of land resources.