Spatiotemporal Sequence Prediction of Ocean 3-D Temperature Field Based on DG-MSCAN Modeling

The spatiotemporal prediction of the ocean 3-D temperature is critical for studying ocean dynamics and climate change. However, most existing research focuses on sea surface temperature prediction. When 3-D temperature field prediction is addressed, it often relies on hierarchical or independent point approaches, neglecting spatial correlations and resulting in insufficient prediction accuracy. To address this limitation, this letter proposes a dynamic gate-driven multiscale attention network (DG-MSCAN), which builds upon Eidetic 3-D LSTM (E3D-LSTM). By integrating the temporal characteristics of LSTM with the spatial characteristics of 3-D convolution, DG-MSCAN captures both temporal and spatial properties simultaneously. Furthermore, to better adapt to the characteristics of the 3-D temperature field, the model incorporates three key improvements: multiscale feature fusion, dynamic gating attention, and lightweight state coupling. Based on the 3-D temperature field data obtained using the gravest empirical mode method and the proposed DG-MSCAN model, this letter conducted spatiotemporal prediction experiments. The results demonstrate that the RMSE of the proposed algorithm is reduced by 59.5%, 62.0%, and 66.8% compared to Conv4D, 3-Dimensional U-Net (3-D U-Net), and E3D-LSTM models, respectively.