Prediction of surface drifter trajectories in the South China sea using deep learning

The complex and dynamic marine environment of the South China Sea (SCS) presents formidable challenges for drifter trajectory prediction. While conventional approaches rely on parametric approximations that inevitably introduce errors, our study introduces a novel Informer-CNN hybrid architecture that significantly advances prediction capabilities. This innovative framework synergistically combines the Informer model's superior long-sequence modeling capacity with CNN's prowess in spatial feature extraction, achieving unprecedented accuracy in 6-24 h forecasts. High-resolution, multi-scale oceanographic data including sea surface currents, sea surface temperature, and sea surface salinity along with drifter trajectory information, are utilized as input to the Informer-CNN model to predict the drifter's latitude and longitude over 6- to 24-hour time horizons. Compared with traditional deep learning models such as RNN, LSTM, GRU, and Transformer, our model achieves a significant reduction in prediction error, demonstrating superior performance and robustness. Quantitative results reveal consistent performance across temporal scales: mean absolute errors of 4.33 km (6 h), 4.73 km (12 h), 9.05 km (18 h), and 13.35 km (24 h), accompanied by corresponding RMSE values of 0.05°, 0.06°, 0.11°, and 0.17°. This research establishes a new benchmark for data-driven marine trajectory forecasting while providing valuable insights into the complex interplay between ocean dynamics and floating object motion.