Dual U–Vision–Transformer for reconstructing the three-dimensional eddy-resolving oceanic physical parameters from satellite observations

• Our model provides reliable estimations of 3-D parameters in marginal seas. • 3-D fine structures of various ocean phenomena are effectively reproduced. • Mode-water eddies can be distinguished from general anticyclonic eddies. • Detailed information from high-resolution satellite observations is fully used. • Our model outperforms other AI models in estimating 3-D ocean parameters. Three-dimensional (3-D) observations are crucial for understanding structures and evolution of ocean dynamic processes. This study proposes a dual U–Vision–Transformer (DUViT) model to reconstruct 3-D, eddy-resolving ocean temperature, salinity, and horizontal current fields from multi-resolution data observed by satellites at the sea surface. Daily parameter profiles from a reanalysis product with a 0.083° grid serve as labels for establishing and evaluating the model. Results demonstrate that the DUViT model can reproduce 3-D physical variables on scales from the basin scale to the mesoscale within the upper 2000 m of the South China Sea. Average root mean square errors between model estimations and reanalysis data are 0.039 ℃, 0.017 psu, and 0.012 m/s for temperature, salinity, and current profiles, respectively, with correlation coefficients above 0.9. Moreover, the reconstructed temperature and salinity profiles and their evolution agree well with in-situ observations with the correlation coefficients above 0.94. Case studies indicate that the DUViT model can effectively reproduce the 3-D structures of the circulation in the upper ocean and mesoscale eddies and distinguish mode-water eddies from general anticyclonic eddies. The integration of high-resolution satellite observations with the DUViT model enables accurate 3-D reconstruction of ocean phenomena across scales, which will offer new insights into the dynamics of marginal seas and open oceans.