Spatiotemporal neural network with attention mechanism for El Nino forecasts

To learn spatiotemporal representations and anomaly predictions from geophysical data, we propose STANet, a spatiotemporal neural network with a trainable attention mechanism, and apply it to El Nino predictions for long-lead forecasts. The STANet makes two critical architectural improvements: it learns spatial features globally by expanding the network's receptive field and encodes long-term sequential features with visual attention using a stateful long-short term memory network. The STANet conducts multitask learning of Nino3.4 index prediction and calendar month classification for predicted indices. In a comparison of the proposed STANet performance with the state-of-the-art model, the accuracy of the 12-month forecast lead correlation coefficient was improved by 5.8% and 13% for Nino3.4 index prediction and corresponding temporal classification, respectively. Furthermore, the spatially attentive regions for the strong El Nino events displayed spatial relationships consistent with the revealed precursor for El Nino occurrence, indicating that the proposed STANet provides good understanding of the spatiotemporal behavior of global sea surface temperature and oceanic heat content for El Nino evolution.

Automated summary

This study proposes a neural network model called STANet, which is used to learn spatiotemporal representations and anomaly predictions from geophysical data. The model improves upon existing architectures by learning spatial features globally and encoding long-term sequential features using a visual attention mechanism. Experimental results show that STANet outperforms state-of-the-art models in predicting El Nino events and provides a good understanding of the spatiotemporal behavior of global sea surface temperature and oceanic heat content.