Soil Moisture Satellite Data Under Scrutiny: Assessing Accuracy Through Environmental Proxies and Extended Triple Collocation Analysis

Abstract 16 different satellite soil moisture (SM) datasets (passive, active, combined, and model data) were compared at the European scale. We hypothesized that SM should be reflected by a variety of environmental factors, such as topography, hydroclimatology, soil characteristics, and biomass. Robust correlation was used to explore the relationship among the satellite data products, and the Recursive Feature Elimination method combined with the Random Forest Regression (RFR) algorithm was used to find the most important variables. Variations in SM-values were analyzed using extended triple collocation analysis (ETC), while the accuracy metrics of the RFR models were summarized through UMAP dimension reduction. The result showed that generally, correlations among the SM products were low ( r < 0.5) with some exceptions. GLDAS had the weakest correlation with the other SM products. Using SM as the dependent variable in regression models, model testing showed that GLDAS’s SM was explained with the highest accuracy based on the Nash-Sutcliffe Efficiency (0.631), followed by the SMOPS (0.624). SSM demonstrated the lowest environmental influence (NSE: 0.288). Using UMAP, ETC, it was determined that SMOPS exhibited superior performance in terms of error variance and model accuracy; however, based on the ETC results, GRD.P was deemed the most suitable option. Results called the attention of varying SM values by products, being biased by various environmental factors and the applied technology of the satellites.