Ocean salinity is an essential climate variable that influences the velocity field through the buoyancy term, thereby regulating ocean dynamic processes and climate change. Satellites have been providing quasi-global sea surface salinity (SSS) observations since 2010; however, their effective resolution, exceeding 100 km, cannot resolve mesoscale eddies and fronts, and they observe only the sea surface. Meanwhile, most 3D reconstruction methods can only provide monthly salinity files and fail to resolve the daily structures. This study develops a deep learning framework that fuses multivariate satellite observations to overcome both limitations through super-resolution reconstruction and daily 3D reconstruction, thereby achieving an eddy-resolving daily 3D salinity field for monitoring ocean dynamic processes.

At the surface level, the SMOS SSS super-resolution reconstruction (S5R2) network super-resolves the SMOS L3 SSS product from 1/4° to 1/12° (eddy-resolving). Employing a hybrid Transformer-CNN architecture with a land filtering mechanism to suppress land interference during attention operations. An optimal variable selection algorithm efficiently determines the best subset of satellite inputs. S5R2 outperforms mainstream L3/L4 satellite SSS products and 12 super-resolution algorithms, achieving 20% and 60% reductions in RMSE over the Kuroshio Extension and Gulf Stream, respectively. The effective resolution of SMOS improves from 75–100 km to 20–30 km, enabling the salinity satellite to resolve and track mesoscale eddies.

At the subsurface level, the Wavelet-Enhanced 3D Mamba (3D-WaveMa) network reconstructs daily 1/4° 3D salinity fields within the upper 1000 m of the Pacific from satellite SSS, SST, and SSH. 3D-WaveMa can capture long-range 3D dependencies of multi-scale processes. Meanwhile, K-means clustering-guided transfer learning enhances local reconstruction across nine subregions, and the Random Forest-based bias correction model revises the reconstructed 3D field griddly. Compared with mainstream reanalysis products, daily and monthly accuracy improves by 13%–56% and 31%–64%, respectively, outperforming standard deep learning architectures for 3D salinity reconstruction.

The two networks are integrated into a daily eddy-resolving 3D salinity reconstruction system. Case studies show it resolves the 3D structure of mesoscale eddies and their weekly evolution, and captures large-scale salinity anomalies during a La Niña event. By transforming coarse satellite surface observations into high-resolution 3D fields, this framework provides satellite-based 3D pseudo-observations with accuracy comparable to reanalysis products, offering new capabilities for feature detection of multi-scale processes, and AI-accelerated data assimilation and forecasting.