A novel random forest approach to downscale SMAP soil moisture products to 100 m resolution using temporally closest satellite observation data

Surface soil moisture is an important factor in controlling the water and energy budget, as well as other hydrological and land surface processes. However, the coarse resolution of satellite data monitoring soil moisture presents a problem that can be addressed by downscaling. This study presents a novel approach for downscaling the coarse-resolution Soil Moisture Active Passive (SMAP) soil moisture product using a random forest model with data from Landsat, MODIS, Sentinel-1, and Sentinel-2 satellites. Key variables include vegetation indices, land surface temperature (LST), low-resolution microwave data, and elevation. Leveraging Google Earth Engine (GEE), individual models are developed for each SMAP image, using the closest finer satellite data to account for temporal variations and enhance prediction accuracy. The downscaled product was evaluated across various spatiotemporal scales and land cover types, showing strong correlations with precipitation and irrigation events, high efficacy in water body detection, and differentiation between crop types and moisture conditions. Comparisons with soil moisture time series from Spain's REMEDHUS network indicate good agreement, with an R-value of 0.697 and an RMSE of 0.098 m³/m³, very close to its much coarser resolution counterpart SMAP/Sentinel-1 1 km product with RMSE 0.07 m³/m³, highlighting the downscaled product's robustness and accuracy. Developing the model for each target soil moisture product, as opposed to a single model for all images, reduces the time and volume of the training phase while maintaining prediction accuracy. This study's findings suggest that downscaling soil moisture data to 100 m resolution significantly enhances the ability to monitor and manage soil moisture at a finer scale. This improvement has broad implications for precision agriculture, hydrological modeling, and environmental monitoring, potentially leading to better resource management, improved crop yields, and more accurate hydrological predictions.