Downscaling of soil moisture in satellite-borne GNSS-R based on multi-model fusion

Soil moisture (SM) inversion using Global Navigation System Reflectometry techniques has garnered significant attention in recent years. However, current Cyclone Global Navigation Satellite System (CYGNSS) SM products generally suffer from low spatial resolution, which limits their application in regional and local scale studies. Therefore, this study proposed a SM downscaling method based on multi-model fusion, aimed at enhancing the spatial resolution of CYGNSS data from 36 km to 3 km. The method integrates CYGNSS observations, SM active passive (SMAP) (36 km) SM data, and auxiliary environmental variables to construct four individual machine learning models: random forest, light gradient boosting machine, categorical boosting, and gradient boosting trees as well as the Stacking fusion model. The results of each model downscaling SM comparison with International SM Network stations show that the Stacking model outperforms the single model in correlation coefficient ( R ), root mean square error, and mean absolute error, while simultaneously exhibiting superior inversion accuracy. In the spatial correlation analysis using moderate resolution imaging spectroradiometer enhanced vegetation index and VIIRS land surface temperature products, the Stacking model results have the highest correlation and more detailed spatial distribution. In addition, the Stacking model downscaled SM also showed good spatial consistency and accuracy in the comparison and verification with SMAP/Sentinel-1 SM (3 km). The results show that the Stacking model, which integrates multiple machine learning algorithms, significantly improves the spatial resolution of the CYGNSS SM products and effectively improves the inversion accuracy, which is of certain reference value for the development of high-resolution SM products.