Multi-Scale domain adaptation for high-resolution soil moisture retrieval from synthetic aperture radar in data-scarce regions

Abstract

Remote sensing of soil moisture plays an important role in advancing various hydrology applications, with Synthetic Aperture Radar (SAR) being the most promising technique for high-resolution soil moisture estimation. The growing adoption of machine learning methods has further enhanced this field, though their effectiveness heavily relies on the availability and quality of in-situ measurements. A recent study has demonstrated that pretrained models at 9-km resolution, based on the Soil Moisture Active Passive (SMAP) soil moisture products, can be transferred to 1-km using fewer in-situ measurements. Despite the success of this cross-resolution framework, its performance in data-scarce regions at high resolutions remains poor. To address this limitation, a multi-scale domain adaption (MSDA) method was proposed for soil moisture retrieval from data-scarce regions at a resolution of 50 m, taking the pretrained 9 km models as the starting points. Two modifications were made: the integration of multi-scale losses at both 9-km and 50-m resolutions, and the application of domain loss to bridge the gap between training and testing datasets. The MSDA was evaluated in both transductive and inductive modes where transductive mode involves adapting the model using a portion of the unlabeled test data, and inductive mode involves generalizing the model to entirely new, unseen data. A total of 66,547 daily averaged soil moisture measurements from 480 stations of 7 networks across the Contiguous United States were used. In the transductive mode, the use of a single training station in the MSDA achieved an R and RMSE of 0.67 and 0.088 m³/m³ respectively, which were improved to 0.81 and 0.071 m³/m³ when using data from 45 training stations. An acceptable R and RMSE of 0.76 and 0.078 m³/m³ was achieved in the inductive mode. The joint use of the two modifications achieved significantly better results (p < 0.01), with a relative improvement of 5.8 – 20.0 %, overall, and a lower risk of performance deterioration in data-scarce scenarios.