Are the Current Expectations for SAR Remote Sensing of Soil Moisture Using Machine Learning Overoptimistic?

High-resolution surface soil moisture is essential for advancing various applications. The increased synthetic aperture radar (SAR) missions over the past decade present an opportunity to obtain large-scale, high-resolution soil moisture data. Machine learning methods are increasingly used for this purpose, but they generally suffered from the availability of ground-based observations. The real performance in view of a global product is still unclear. Consequently, commonly used machine learning methods were evaluated in this study in simulated global mapping scenarios with few training data, using a global dataset of 209 318 samples from 1021 locations worldwide, and a unique regional dataset with intensive ground and airborne-derived soil moisture from L-band passive microwave observations. Three evaluation scenarios based on the global dataset were involved, with ≤5% samples used for training. The target accuracy of 0.06 m3/m3 was only met in the dependent evaluation scenario, where the training and testing samples were randomly split. In the temporal evaluation scenario and spatial evaluation scenario, where training and testing samples came from different time periods or locations, the best models achieved median root-mean-square errors (RMSEs) of only 0.078 and 0.089 m3/m3, respectively. The evaluation on the regional dataset showed consistently worse accuracy statistics (RMSE > 0.1 m3/m3 and R < 0.41). Moreover, all methods failed to capture the spatial patterns of soil moisture, compared to airborne-derived passive soil moisture maps. These findings, therefore, suggest that current expectations for SAR-based soil moisture estimation using machine learning may be overoptimistic, requiring more robust approaches for scenarios with sparse ground measurements.