High-Resolution Mapping of Soil Moisture by AMSR2 Data Disaggregation Based on Sentinel-1 and Machine Learning

Thanks to the frequent revisiting, satellite microwave radiometers have great potential for surface soil moisture (SM) monitoring. However, their spatial resolution is not sufficient for hydrological studies in small catchments as well as applications to precision farming. In this study, a disaggregation technique based on machine learning is proposed: the technique combines Sentinel-1 (S-1) SAR data with SM generated from advanced microwave scanning radiometer 2 by the IFACs HydroAlgo algorithm, with the aim of enhancing the SM spatial resolution from the original 10 km to about 30 m. To this scope, two machine-learning techniques have been considered for the implementation, namely artificial neural networks (ANNs) and random forests (RF). Training is carried out by aggregating and coregistering S-1 data with the HydroAlgo SM at 10-km resolution. After training, the ANN and RF algorithms are applied pixel-by-pixel to the S-1 images at full resolution for generating the enhanced SM maps. The method has been implemented and validated in two agricultural areas located in Central Italy, where a series of experiments has been carried out between 2019 and 2020 for collecting the main soil and vegetation parameters at the same time of satellite overpasses. To assess the actual resolution of the output SM, the validation against in situ measurements has been carried out by aggregating data at 10, 30, 50, and 70 m. The results confirmed the effectiveness of the proposed method: validation carried out at 30 m obtained R≃0.82 and RMSE≃0.05 m3/m3 that represent a noticeable improvement with respect to the results obtained by HydroAlgo at 10 km (R≃0.56 and RMSE >> 0.1 m3/m3). Validation results also pointed out the superior performances of the ANN based with respect to the RF-based disaggregation.