Effective attention model for global sea surface temperature prediction

Abstract

Sea Surface Temperature (SST) is an essential parameter of the ocean–atmosphere system, so accurately predicting the future SST is of great significance. To address the low prediction accuracy due to insufficient consideration of spatiotemporal correlations in existing machine learning models, this paper proposes an Effective Attention Model (EAM) to achieve daily global SST prediction (predicting the SST of the next 15 days using those of the past 15 days) and monthly global SST prediction (predicting the SST of the next 12 months using those of the past 12 months). The EAM adopts an attention mechanism to avoid information redundancy and improve prediction accuracy. This is the first work to realize high-accuracy global SST prediction with a high spatial resolution of 0.25° on two time scales (daily and monthly) based on deep learning. Experiments show that (1) the prediction results have a consistent global spatiotemporal distribution with the ground truth, with the Root Mean Squared Difference (RMSD) of daily prediction ranging from 0.212 to 0.932 ℃ and the RMSD of monthly prediction ranging from 0.666 to 0.782 ℃. The EAM is superior to the deep learning models (ConvLSTM, 3D-CNN, and LSTM) and the numerical model (HYCOM); (2) Besides global prediction, the EAM also achieves good performance in a climate change sensitive area (Niño 3.4 sea area), with the lowest RMSD of 0.231 ℃. This study may not only benefit the research on climate events (El Niño and La Niña) but also serve to predict other marine disasters (such as heat waves and storms). In the future, the proposed attention module can also be adapted to other prediction scenarios such as meteorology.