A cross-resolution transfer learning approach for soil moisture retrieval from Sentinel-1 using limited training samples

Abstract

Synthetic Aperture Radar (SAR) data is increasingly popular as a data source for global near-surface soil moisture mapping, but large-scale applications are still challenging due to the complex scattering process and the cumbersome data preprocessing. The emergence of deep learning methods has allowed advances in the remote sensing of large-scale surface parameters, but its application in SAR soil moisture retrieval has suffered from the availability of ground soil moisture measurements. Accordingly, this study proposed a cross-resolution transfer learning framework, with the assumption that sophisticated models for different spatial resolutions share a similar model architecture and trainable parameters. A robust high-resolution model can thus be trained with fewer samples by using coarse models. Accordingly, 25 deep learning models were pre-trained taking ∼387,000 Soil Moisture Active Passive (SMAP) Level-3 9 km enhanced passive soil moisture measurements as the truth, with an average validation RMSE of 0.03 m3/m3. They were then transferred to finer grids of 0.1–1 km using a small number of in-situ samples. A total of ∼190,000 daily soil moisture measurements from the international soil moisture network (ISMN) were used to evaluate the proposed framework in three scenarios. The results show that 1) 5000–6000 random samples are sufficient to achieve a target RMSE of 0.06 m3/m3; 2) training samples from a short period (2 or 4 months for Sentinel-1) of 2021 resulted in an overall RMSE of ∼0.068 m3/m3 in an independent period of 2016–2020; 3) the transfer learning also improved the retrieval accuracy (10–30% in relative) over areas without ground samples used for training but failed to yield an acceptable accuracy over mountainous areas. The promising results from this study confirmed the effectiveness of using “pre-trained models + scenario specific models” for regional to global soil moisture retrieval from Sentinel-1.