Accurately obtaining high-resolution ocean subsurface thermal structure (OSTS) is essential for resolving mesoscale dynamics in the tropical Indian Ocean (TIO), yet observations remain sparse and uneven. We present a Transformer-based neural network model, the Downscaling Vision Transformer (DSVIT), which integrates prior knowledge to reconstruct high-resolution OSTS in the TIO. Inputs include sea surface temperature (SST), absolute dynamic topography (ADT), and wind stress curl (WSC), as well as temporal, geographic, and climatological information. DSVIT enhances a standard Vision Transformer (ViT) with a geographic positional prior and a physics-aware loss that emphasizes thermocline and surface variability. On an independent test set, DSVIT achieves a Root Mean Square Error (RMSE) of 0.29 °C and a Coefficient of Determination (R2) of 0.9962 for reconstructed subsurface temperature, outperforming traditional recurrent neural network (RNN) and convolutional neural network (CNN) models. Moreover, a key innovation of this study lies in its novel downscaling strategy, which effectively improves the EN4 subsurface temperature resolution from 1° to 1/4° by altering the input segmentation. Compared with traditional interpolation and assessed against independent high-resolution products, the downscaled outputs exhibit lower RMSE and higher R2, indicating enhanced physical consistency and mesoscale representation. SHapley Additive exPlanations (SHAP) analysis further reveals that climatology and SST are the dominant predictors, followed by ADT. This study provides a novel approach for downscaling OSTS and offers valuable insights for advancing oceanic and climatic research.
扫码关注我们
求助内容:
应助结果提醒方式:
