• A Novel Multi-Source Attention-Enhanced Transformer U-Net is proposed for SSS retrieval. • It can achieve accurate SSS retrieval without reliance on in-situ observations or salinity background fields. • The proposed method significantly outperformed the official SMAP 8-day SSS product without data assimilation. • It demonstrated robust retrieval performance across seasonal cycles and diverse spatial domains. Accurate mapping of sea surface salinity (SSS) is essential for diagnosing the marine hydrological cycle and enhancing coupled atmosphere–ocean forecasts. However, current L-band satellite products remain compromised by radio-frequency interference, land-sea contamination and imperfect roughness corrections. Therefore, this study developed a multi-source attention-enhanced Transformer U-Net (MSAtt-TransUNet), which synergistically exploits Soil Moisture Active Passive (SMAP) L-band brightness temperature (TB) and auxiliary environmental factors, adopting a physically guided loss function that preserves the salinity front to retrieve SSS in the dynamically complex Northwest Pacific. MSAtt-TransUNet reduced Root Mean Square Error (RMSE) by 12. 6% relative to the baseline U-Net. Compared with in situ Argo data, the proposed model attained an RMSE of 0. 235 psu and a Pearson correlation coefficient (R) of 0. 844, markedly outperforming the operational SMAP Level-3 product (0. 338 psu and 0. 728). Validation against the BOAArgo gridded dataset further confirms that MSAtt-TransUNet had better performance than SMAP official, while remaining robust across spatio-temporal domains. Notably, spatial diagnostics against the Hybrid Coordinate Ocean Model (HYCOM) analysis indicate that MSAtt-TransUNet effectively reconstructed salinity gradients across the Northwest Pacific. Shapley additive explanation and channel ablation experiments assigned the highest importance to dual-polarized TB, followed by wind speed and sea surface temperature. The model achieved these gains without assimilating any in situ measurements or model priors, offering a highly accurate, robust, and interpretable alternative for near real-time SSS monitoring and regional data-assimilation applications.
Han et al. (Sun,) studied this question.