Sea Level Anomaly Prediction Using Attention-based ConvLSTM in the Northern South China Sea

Abstract This study developed an attention-based ConvLSTM model to predict sea level anomaly (SLA) in the northern South China Sea (NSCS). The model integrates two variants of spatial attention mechanisms to dynamically focus on key ocean features, such as mesoscale eddies, and addresses the limitation of traditional methods that treat all features equally. Using 30 years of daily SLA data from CMEMS (1993-2023) with a spatial resolution of 1/8°, our model demonstrated excellent performance, It achieves a root mean square error (RMSE) of 0.0102 meters, with 94.4% of predictions under 0.02 meters, significantly outperforming stacked LSTM (RMSE:0.4739 meters) and standard ConvLSTM (RMSE:0.0158 meters).The attention mechanism is particularly effective in handling complex ocean dynamics, and the model has been proven to be suitable for SLA prediction in the northern South China Sea.