A South China Sea Surface Absolute Dynamic Topography Prediction Model Based on Convolutional Long Short-Term Memory Network With Self-Attention Mechanism

Abstract

Absolute dynamic topography (ADT) obtained from satellite altimeter data mapping is widely used in marine environment monitoring and research. Traditional numerical ADT prediction models exhibit high computational demands and low operational efficiency. This study proposes a deep learning framework integrating U-Net architecture with a self-attention convolutional long short-term memory network (SA-ConvLSTM) to develop a high-precision ADT forecasting model for the South China Sea. The approach utilizes 0.08° high resolution multi-source satellite data. Training optimization incorporating teacher forcing and scheduled sampling enhanced model capability in representing complex ocean dynamics. The SA-ConvLSTM is shown to outperform the traditional ConvLSTM model and several existing models in terms of both forecast accuracy and computational efficiency. This framework is demonstrated significant potential for high-resolution marine forecasting and disaster early warning systems, offering an efficient alternative to traditional numerical models for regional ocean dynamic monitoring.