The Spatio-Temporal Predictability of Mesoscale Eddies via 3D Conv-LSTM with Intensified Memory

Oceanic mesoscale eddies, typically ranging from tens to hundreds of kilometers, at the temporal scales on the order of days to years, are ubiquitous in the global ocean. Accurately predicting the spatio-temporal variation of mesoscale eddies provides more opportunities to understand the propagation and evolution process of mesoscale eddies, which helps intensively investigate their significant roles in the redistribution of energy, heat, and salt. In this paper, we propose to develop the spatio-temporal predictability modelling of mesoscale eddies, by employing 3D Conv-LSTM architecture of intensified memory, together with the recursive and direct strategies. It has been demonstrated from the experiment results that our proposed scheme outperformed most of the state-of-arts approaches for the prediction of SLA time series, including Conv-LSTM, Eidetic 3D-LSTM, with RMSE, MAPE of the 14-day prediction duration, respectively, 0.0392, 5.042% in our developed model.