[A-TT30] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT30-05] Predictability of El Niño-Southern Oscillation (ENSO) Index Beyond 2-Year Lead Time Using Seasonally Optimized Deep Learning Models

*Kalpesh Ravindra Patil¹, Takeshi Doi², J. V. Ratnam², Swadhin Behera² (1.Young Research Fellow, APL, VAiG, JAMSTEC, 2.APL, VAiG, JAMSTEC)

Keywords:Deep learning, CNN, ENSO prediction, Long-lead ENSO forecast

Recent advancements in deep learning have significantly improved the forecasting of the El Niño-Southern Oscillation (ENSO) index. These models now achieve lead times of approximately 1½ to 2 years, surpassing fully coupled global ocean–atmosphere models. In our study, we delve deeper, leveraging seasonally optimized deep learning models to explore the potential predictability of the ENSO index beyond the 2-year lead time. Our findings reveal intriguing patterns: The ENSO index exhibits moderate to poor but statistically significant correlation skill in forecasting the all season ENSO index. Interestingly, this correlation skill is heightened during the peak season (December-February) when analyzed in terms of seasonal correlation. Furthermore, our models successfully capture several peak ENSO events. These events exhibit a weaker intensity, yet they demonstrate significant skill in both deterministic (anomaly correlation coefficient, ACC) and probabilistic (relative operating characteristic curve, ROC) measures. Notably, our study results outperform earlier deep learning investigations, suggesting the potential predictability of the ENSO index during the peak season up to about a 2½ years ahead. In summary, this study underscores the significance of seasonally optimized deep learning models in extending ENSO predictability beyond the conventional 2-year lead time. These findings have implications for climate forecasting and risk assessment.