• ASTRA delivers long-term SMAP-like SM from AMSR2 via image-to-image translation. • Through multiband inputs, performance improves over AMSR2 even in dense vegetation. • ASTRA back-extends SMAP-like SM to 2012 and expands global spatiotemporal coverage. • Time-varying ω , h derived from τ-ω inversion explain wet bias in dense vegetation. Soil moisture (SM) plays a central role in regulating land–atmosphere exchanges of water and energy and is therefore essential for climate and hydrological studies. The Advanced Microwave Scanning Radiometer 2 (AMSR2) has provided continuous multi-frequency observations since 2012; however, its shorter wavelengths compared with the Soil Moisture Active Passive (SMAP) L-band are more strongly attenuated by vegetation, and its retrieval algorithm relies on globally fixed parameters, leading to systematic biases in densely vegetated regions. Here, we introduce ASTRA (AMSR2-to-SMAP image-to-image translation), a deep learning framework that transforms AMSR2-based products into SMAP-consistent SM estimates. ASTRA employs a Swin-Unet architecture trained on SMAP SM retrieved with the single-channel algorithm at vertical polarization, using AMSR2 Land Parameter Retrieval Model SM, vegetation optical depth, and skin temperature as inputs. During the test period from April 1, 2015, to December 31, 2016, ASTRA achieved a correlation coefficient of 0.814, an unbiased root mean square deviation of 0.027 m 3 m⁻ 3 , and a bias of 0.005 m 3 m⁻ 3 relative to SMAP, demonstrating strong consistency with L-band retrievals. Validation against in situ observations from the International Soil Moisture Network (ISMN) further confirmed that ASTRA outperforms AMSR2 SM and approaches SMAP-level accuracy across diverse land-cover conditions. Beyond statistical harmonization, we investigated physical interpretability by inverting the τ - ω radiative transfer model to estimate τ and ω , using ASTRA-derived SM as a constraint. The resulting time-varying estimates of single-scattering albedo and surface roughness substantially reduced the wet bias of the original AMSR2 C-band product and improved agreement with SMAP and ISMN data. ASTRA thus provides a pathway to extend SMAP-consistent SM records back to the pre-SMAP era while offering insights into physically meaningful parameterization for long-term SM monitoring.
Lee et al. (Thu,) studied this question.