Reconstructing three-dimensional salinity field of the South China Sea from satellite observations

Abstract

High-resolution salinity information is of great significance for understanding the marine environment. We here propose a deep learning model denoted the “Attention U-net network” to reconstruct the daily salinity fields on a 1/4° grid in the interior of the South China Sea (SCS) from satellite observations of surface variables including sea surface salinity, sea surface temperature, sea level anomaly, and sea surface wind field. The vertical salinity profiles from the GLORYS2V4 reanalysis product provided by Copernicus Marine Environment Monitoring Service were used for training and evaluating the network. Results suggest that the Attention U-net model performs quite well in reconstructing the three-dimensional (3D) salinity field in the upper 1000 m of the SCS, with an average root mean square error (RMSE) of 0.051 psu and an overall correlation coefficient of 0.998. The topography mask of the SCS in the loss function can significantly improve the performance of the model. Compared with the results derived from the model using Huber loss function, there is a significant reduction of RMSE in all vertical layers. Using sea surface salinity as model inputs also helps to yield more accurate subsurface salinity, with an average RMSE near the sea surface being reduced by 16.4%. The good performance of the Attention U-net model is also validated byin situmooring measurements, and case studies show that the reconstructed high-resolution 3D salinity field can effectively capture the evolution of underwater signals of mesoscale eddies in the SCS. The resolution and accuracy of sea surface variables observed by satellites will continue to improve in the future, and with these improvements, more precise 3D salinity field reconstructions will be possible, which will bring new insights about the multi-scale dynamics research in the SCS.