PSSA-GAN: a Network with Self-attention and Season Attention for High Spatiotemporal Resolution Ocean Subsurface Temperature Reconstruction

Due to the limited availability of subsurface and deep sea observation data, accurately reconstructing high spatiotemporal resolution ocean interior temperatures is both a crucial and challenging task. This study proposes Patch Self attention and Season Attention Generative Adversarial Network (PSSA-GAN) for reconstructing high spatiotemporal resolution ocean subsurface temperature fields. The model utilizes sea level anomaly (SLA), sea surface temperature (SST), and sea surface wind (SSW) data as inputs, extracting multi-scale features and seasonal signals from the remote sensing data to reconstruct the ocean subsurface temperature fields. To improve the model's ability to capture global features and to allow the model to incorporate seasonal variations in ocean temperature during training. The patch self-attention and season attention mechanism are integrated into the generator. To ensure the reconstructed temperature fields more align with the real ocean environment, a re-weighting loss function is designed. The loss function jointly leverages ocean reanalysis data and Argo observations as ground truth for model training. The daily temperature fields in the Pacific Ocean (104°E-127°W, 20°S-44°N) are reconstructed by PSSA-GAN. The construction results demonstrate that PSSA GAN consistently captures the spatial patterns across various depth layers, aligning well with reanalysis data and reflecting fine scale ocean variations in the tropical eastern Pacific. The average RMSE of the reconstruction results across 17 layers above 1062m is 0.5213°C, outperforming both the Attention U-net and Pix2Pix models. The evaluation experiments based on Argo in-situ observations show that the reconstructed daily mean temperature field achieves a numerical accuracy comparable to that of ocean reanalysis data.