Enhancing the Spatial Applicability of SMAP Soil Moisture Using Multi-Stage Machine Learning Downscaling
TL;DR
Results indicate that multi-stage machine learning downscaling can enhance spatial interpretability and strengthen the applicability of SMAP SM for agricultural monitoring, while preserving broad consistency with the original passive microwave retrieval and providing a practical baseline for regional refinement.
Abstract
The Soil Moisture Active Passive (SMAP) mission provides global soil moisture (SM) observations with high temporal frequency and broad spatial coverage, valuable for hydrological and agricultural applications. However, its coarse spatial resolution can limit direct use in heterogeneous agricultural landscapes where soil moisture varies over short distances. This study developed a multi-stage machine learning framework to improve the spatial applicability of SMAP Level-3 SM across the conterminous United States (CONUS). SMAP brightness temperatures and nighttime land surface temperature (LST) were used as core retrieval-relevant predictors, while vegetation-related controls and soil–vegetation background interactions were incorporated as supplementary information during downscaling and as regional controls during refinement. The Random Forest (RF) based framework first generated 500 m and 250 m downscaled SMAP soil moisture across the CONUS. A 100 m regionally refined soil moisture field was then produced by combining the downscaled SMAP soil moisture with regional environmental drivers. After aggregation to the 9 km SMAP Level-3 product grid, the 500 m downscaled product showed strong consistency with the corresponding SMAP values, with an r2 of 0.880 and an RMSE of 0.0292 m3/m3 across the CONUS. Direct comparison with International Soil Moisture Network (ISMN) in situ observations showed RMSE values of up to 0.099 m3/m3 depending on the site network. Following regional refinement based on hydro-topographic, soil, and vegetation-related controls, the overall five-fold cross-validated RMSE was 0.052 m3/m3, indicating improved predictive refinement within the available ISMN station network. These results indicate that multi-stage machine learning downscaling can enhance spatial interpretability and strengthen the applicability of SMAP SM for agricultural monitoring, while preserving broad consistency with the original passive microwave retrieval and providing a practical baseline for regional refinement.