Reconstruction of Subsurface Temperature Field in the South China Sea From Satellite Observations Based on an Attention U-Net Model

In this study, an attention U-net network was proposed to reconstruct the subsurface temperature (ST) field with high temporal and spatial resolution in the South China Sea (SCS) from sea surface parameters observed by satellites. In addition to sea surface temperature, sea-level anomaly, and sea surface wind (SSW) field, the wind stress curl, which influences the 3-D structure of temperature through the induced Ekman pumping and transport, was also input into the model. The five-day average vertical temperature profiles with a spatial resolution of 0.5° from simple ocean data assimilation (SODA) reanalysis were used for training and evaluating the network. The results show that the attention U-net model performs quite well in ST reconstruction in the upper 100-m layers of the SCS. The additional input of wind stress curl helps to improve the model accuracy. The average root-mean-square error (RMSE)/bias of ST decreases from 1.08°C/−0.21°C to 1.01°C/−0.05°C. In particular, the RMSE near the thermocline is reduced significantly by up to 10.9%. The estimation error of the attention U-net model is much smaller than that of some linear and tree models in the SCS, especially in shallow waters and regions with complex dynamic processes. The case study also shows that our model is capable of capturing the evolution of mesoscale processes in the SCS. The combination of satellite observations with a high-precision ST reconstruction model will help us comprehensively understand the fine structure and variation of temperature and circulation in the marginal seas and open oceans.