Predicting Wintertime Anomalous Eddies in the Bay of Bengal using a TAU-Enhanced SimVP Model and Satellite Remote Sensing Data

Recent studies based on multi-source satellite remote sensing have revealed the widespread presence of abnormal coldcore anticyclonic eddies globally, with a particularly high occurrence during the Northeast Monsoon in the Bay of Bengal. This study enhances the Simpler yet Better Video Prediction (SimVP) model by replacing its translator module with a Temporal Attention Unit (TAU). The model is trained using 14 years (2004-2017) of satellite-derived Absolute Dynamic Topography (ADT) and Sea Surface Temperature (SST) data and subsequently applied to predict eddy features from 2018 to February 2025. The proposed SimVP-TAU achieves low shortterm forecast errors (ADT RMSE: $\mathbf{1. 9 3} \mathbf{~ c m}$; SST RMSE: $\mathbf{0. 2 7}{ }^{\circ} \mathbf{C}$). Evaluation results show a high wintertime anticyclonic eddy matching rate of $\mathbf{8 3. 7 \%}$ and a polarity classification accuracy of $76.7 \%$. The experiments demonstrated the model’s strong generalizability to satellite remote sensing data and confirmed its feasibility for forecasting anomalous eddies. This work broadens the application of satellite remote sensing in oceanographic research and offers a novel perspective on multi-source, multiscale remote sensing data fusion forecasting techniques, thereby advancing the intelligent evolution of remote sensing methodologies.