Mesoscale Eddy Prediction by LSTM Networks Based on Physical Features

Mesoscale eddies play a critical role in ocean circulation and biogeochemical processes, yet predicting their dynamic characteristics remains challenging due to nonlinear interactions and background errors in traditional methods. This study proposes a physics based Long Short-Term Memory (LSTM) network to predict key eddy features, amplitude, radius, and maximum circularly averaged speed (MCAs), by integrating multi-source observational data and hydrodynamic principles. Utilizing 28 years (1993-2020) of daily eddy trajectories from the global META3.1exp atlas and high-resolution reanalysis data (JCOPE2M) in the Northwest Pacific (15°-35°N, 115°-135°E), we systematically evaluate the effects of temporal sequence length and physical variables on prediction performance. The model demonstrates superior accuracy compared to conventional LSTM approaches, with mean absolute errors (MAE) for 1-7 day predictions increasing from 0.72 cm to 1.37 cm (amplitude), 8.85 km to 18.02 km (radius), and 0.80 cm/s to 2.46 cm/s (MCA). Key innovations include: 1) Dynamic reconstruction of spatiotemporal label-feature relationships to mitigate error accumulation, 2) Incorporation of sea surface temperature (SST) and height (SSH), which improve prediction accuracy by 5.33-5.92% and 3.65-5.47%, respectively, outperforming eddy kinetic energy inputs. Seasonal analysis reveals lower model accuracy in summer versus winter, particularly for amplitude (MAE: 1.29 cm vs 1.03 cm) and radius (15.3 km vs 13.2 km). Interannual error patterns correlate with El Niño events, highlighting climate-ocean coupling effects. This work advances eddy prediction through physics-guided machine learning, providing a framework for operational ocean forecasting. Future extensions could incorporate three-dimensional eddy structures and additional environmental drivers to enhance predictive capability.