[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-P03] Spatiotemporal Attention-Enhanced Deep Learning for Improved ENSO Long-Lead-Time Forecasting

*Wen-Chieh Wu¹、Dong-Lin Li¹、Chih-Chieh Young² (1.Department of Electrical Engineering, National Taiwan Ocean University、2.Department of Marine Environmental Informatics, National Taiwan Ocean University)

キーワード：ENSO, lead time, spring predictability barrier, deep learning, sea surface temperature, sea surface salinity

Accurate long-lead-time predictions of the El Niño-Southern Oscillation (ENSO) that strongly affects global weather and climate remain a challenging task due to complex ocean-atmosphere interactions and the spring predictability barrier (SPB). In this study, to extend forecast lead times and mitigate SPB effects, we propose a deep learning-based ENSO prediction model with incorporation of sea surface salinity (SSS) and sea surface temperature (SST). In particular, based on the multiple oceanic variables, the multiscale spatiotemporal feature extraction framework with an attention mechanism can capture key physical processes to effectively improve the forecasting skills. We will experimentally investigate its ability to predict the Niño 3.4 index 24 months in advance and conduct interpretability analysis to elucidate different roles of SST and SSS. With growing availability of satellite data, the multi-variable data-driven model holds great promise for advancing long-term ENSO prediction.