Random Forest-Based Soil Moisture Downscaling in Southeastern Hungary

Soil moisture is a key variable in the water and energy cycle and has an important role in agricultural drought monitoring. Existing remote sensing products, such as SMAP, are limited in their regional applicability due to coarse spatial resolution. To overcome this, downscaling techniques are widely used. Machine learning methods, especially Random Forest, have shown strong potential for integrating multi-sensor auxiliary data to retrieve highresolution soil moisture. In this study, NASA's SMAP Level- 3 Enhanced product at 9 km resolution was captured via Google Earth Engine. Auxiliary predictor variables, including Sentinel-1 SAR backscatter coefficients (VV and VH polarizations), MODIS-derived vegetation indices (NDVI and EVI), and land surface temperature, were incorporated. Monthly datasets from April to October for 2021 and 2022 were used over south-eastern Hungary to compare normal conditions and the drought-affected 2022 growing season. Random Forest models were trained to establish statistical relationships with SMAP soil moisture and applied to produce 1 km downscaled maps. The model showed strong performance, consistent with original SMAP retrievals, while capturing the spatial heterogeneity of drought conditions. Validation against ground-based measurements revealed uncertainties arising from the depth mismatch between SMAP's 0–5 cm sensing depth and in-situ observations collected at 10 cm, which may be particularly pronounced during drought conditions.