Soil moisture at 30 m from multiple satellite datasets fused by random forest

As a key variable used to characterize the climate process between the land surface and atmosphere, the surface soil moisture (SM) plays an irreplaceable role in the fields of hydrology, meteorology and agriculture. However, the relatively coarse spatial resolution of SM products currently limit the application in water resource management at the field scale. In this study, we proposed a high-spatial-resolution SM retrieval framework based on the random forest algorithm (RF-SM) to integrate in situ SM dataset from in the International Soil Moisture Network (ISMN), Landsat 8 optical and thermal observations, soil properties from SoilGrids V2.0, meteorological variables from the fifth generation of the European ReAnalysis (ERA5) dataset and four coarse-scale SM products including the Soil Moisture Active/Passive (SMAP), the Soil Moisture and Ocean Salinity (SMOS), European Space Agency Climate Change Initiative (ESA CCI) and U.S. National Climate Assessment Land Data Assimilation System (NCA-LDAS). Compared to three other machine learning (ML) algorithms [extremely randomized trees (Extra-Trees), gradient boosting regression tree (GBRT) and extreme gradient boosting (XGBoost)], the random forest (RF) algorithm exhibited the best performance against a subset of 100 validation sites with a Kling–Gupta efficiency (KGE) of 0.69 and root-mean-square-error (RMSE) of 0.063 m3/m3. In terms of different land cover types and typical sites, RF-SM also showed a better accuracy than any of the individual SM product. Finally, the retrieval framework was applied to map the 30-m resolution SM spatial distributions in five substudy areas in the U.S. The results suggest that it is feasible to retrieve accurate SM at a 30-m spatial resolution from multiple satellite datasets based on the RF algorithm, which has important practical significance for agricultural drought monitoring at the field scale.