Deep-learning-based prediction of mesoscale eddy distribution in the South China sea

Prediction of the distribution of mesoscale eddies plays a crucial role in sea navigation, guiding ships and submarines, and in route planning. In the past, the prediction of the trajectory of mesoscale eddies has primarily focused on their paths, often overlooking their distribution. However, given that the diameter of mesoscale eddies can span several hundred kilometers, their spatial distribution holds significant importance for the navigation and route planning of ships and submarines. To address this critical gap, this study concentrates on the South China Sea, a region renowned for its intricate sea conditions. The study introduces a novel BiST-LSTM model tailored to predict the distribution of mesoscale eddies. Leveraging the capabilities of the long short-term memory (LSTM) network, the model incorporates bidirectional spatial-temporal streaming and an attention mechanism. In this investigation, both the CMEMS dataset and a dataset constructed using the PYEDDYTRACKER method are employed to validate the efficacy of the proposed model and to compare its performance against other existing models. Experimental findings consistently demonstrate that the BiST-LSTM model outperforms alternative image prediction models in terms of evaluation metrics. Recognizing the impact of offshore typhoons on mesoscale eddy distribution, wind speed data from the ERA5 dataset are integrated to enhance the model's predictive capabilities. This augmentation significantly bolsters the model's accuracy in forecasting the distribution of mesoscale eddies, thereby reinforcing its utility in sea navigation, route planning, and ensuring safe passage during typhoon events.