Super-resolution reconstruction of subsurface temperature field in South China Sea using satellite observations

The accurate reconstruction of high-resolution sea subsurface temperature structures is essential for comprehending meteorological models and evaluating climate change impacts. However, the diversity of the sea subsurface environment and complex physical processes make it challenging to directly reconstruct high-resolution subsurface temperature data from low-resolution satellite observations at once and with high accuracy. This study proposes an indirect two-phase transformer-based model to achieve super-resolution reconstruction of sea subsurface temperatures at a 1/12° resolution, utilizing 1/4° resolution satellite observations from the South China Sea. The two-phase method decouples the super-resolution reconstruction task to reduce the complexity of model optimization, leading to a more accurate result. In the first phase, the model inverts the subsurface temperature data from satellite observations to depth profiles. In the second phase, the model performs super-resolution based on the inverted sea subsurface temperature data in depth profiles. Experiments are conducted using the Copernicus Marine Environment Monitoring Service temperature dataset, and the performance of the proposed model is compared against Attention U-net, Very Deep Super-Resolution, and Super-Resolution Convolutional Neural Network. The results indicate the superior performance of the proposed model, achieving a root mean square error of 0.3524 °C, a structural similarity index of 0.9854, and a peak signal-to-noise ratio of 42.5031 in the super-resolution reconstruction of a 27-layer sea subsurface layer temperature profile, covering depths from 0 to 200 m. This demonstrates the model's effectiveness in enhancing the resolution of sea subsurface temperature data, which is crucial for improving the understanding of marine environments and climate dynamics.