Deep Learning Model for ENSO Forecasting Using Multiple-Scale Spatiotemporal Information

The variability of the El Niño/Southern Oscillation (ENSO) is associated with a diverse range of climate-related extremes and impacts on ecosystems. As such, the ability to provide robust and accurate long-lead forecasts would be invaluable for effective policy management. Current research uses only interseasonal scale spatiotemporal information to predict the ENSO for the target month. However, the interannual information is largely ignored by existing methods. In this study, we propose a novel architecture based on a vision transformer (ViT) model to combine spatiotemporal information from both interseasonal and interannual scales for predicting the target month Niño3.4 index. The model further incorporates a monthly aware (MA) token to effectively capture seasonal variations. Our results demonstrate that the proposed model achieves accurate forecasts up to 20 months in advance, with significant improvements in prediction skill for lead times of 6–15 months across all seasons when interannual information is included. In addition, the attention maps provide insights into the physical connections driving ENSO predictions. Furthermore, the MA token outperforms a single token in capturing global spatiotemporal features. These findings highlight the potential of combining interseasonal and interannual information within deep learning frameworks to advance ENSO prediction and deepen our understanding of its dynamics.