This study presents a methodology to retrieve surface soil moisture (SM) at 1-km scale over croplands and grasslands using Sentinel-1 radar backscatter, harmonized Landsat/Sentinel-2 NDVI, and ancillary environmental variables. A Random Forest (RF) model was trained on the coarse (25-km) SMOS-IC product (2017–2024) to learn relationships between predictors and SM and then applied to generate fine-scale estimates. Only SMOS-IC pixels dominated by cropland and grassland selected to improve the reliability of the estimation. The RF achieved good performance at the training scale (R = 0.79, RMSE = 5.68 vol.% and negligible bias). When transferred to 1-km grids, independent validation using International Soil Moisture Network observations (2015–2024) yielded an R value of 0.49, an RMSE of 11.2 vol.%, an unbiased RMSE of 9.55 vol.%, and a global bias of 5.8 vol.%. Spatial comparison with GLASS SM indicated generally strong correlations, especially in humid regions. However, limitations remain, including a systematic underestimation inherited from SMOS-IC and reduced performance over areas with fine-textured soils. Accuracy also decreases under very dense vegetation and following recent heavy rainfall events (>30 mm). Despite these challenges, the 1-km High-resolution SM dataset offers valuable spatial detail on soil moisture variability, supporting improved agricultural management and local-scale monitoring.
Karami et al. (Sat,) studied this question.