The Soil Moisture and Ocean Salinity (SMOS) satellite mission has offered the longest continuous record of sea surface salinity (SSS) observations, since 2010. This extensive dataset contributes the study of large-scale salinity-related phenomena. However, the effective resolution of the L3 SMOS SSS is still unable to resolve mesoscale phenomena due to the limitations of the SMOS footprint, swath width, revisit time, and retrieval noise, and its data in the coastal zone are largely missing. Therefore, the SMOS Sea Surface Salinity Super-Resolution Reconstruction (S5R2) network using deep learning-based super-resolution (SR) is here proposed to achieve high quality reconstruction of L3 SMOS SSS by fusing multiple ocean remote sensing variables. First, we improve the Self-Attention of Transformer. One is to introduce CNNs attention to highlight key local regions and channels in the global dependency. The other is to propose a land filtering mechanism to focus attention on the ocean. Second, we improve the search efficiency of optimal input variables by limiting the direction and step size of the search through importance scores of random forest and correlations. S5R2 improves the spatial resolution of the L3 SMOS SSS from 1/4° to 1/12° and the temporal resolution from 10 days to 1 day, and removes noise, while completing the missing data for the coastal zone. The wave number spectrum analysis verifies that the effective resolution of the reconstructed product is improved from ~100km to ~35km. Comparing five satellite SSS products and six SR methods, S5R2 achieves the best performance. The root-mean-square error of the reconstructed product was reduced from 0.581 to 0.237 psu in the Kuroshio Extension region and from 0.583 to 0.465 psu in the Gulf Stream region. S5R2 achieves near-real-time, high-quality reconstruction of L3 satellite-derived SSS products, and further contributes to the satellite-based monitoring and research of SSS in the coastal zone.

SMOS,sea surface salinity (SSS),deep learning,Convolutional Neural Networks (CNNs),Transformer,wavenumber spectrum analysis,Feature selection