Distillation of CNN Ensemble Results for Enhanced Long-Term Prediction of the ENSO Phenomenon

2025-09-07

AI Summary

Current ENSO long-range forecasting commonly employs equal-weight ensemble averaging, overlooking inter-member skill heterogeneity and thus limiting prediction accuracy. To address this, we propose a posteriori assessment–based ensemble distillation method: leveraging outputs from a CNN-based ensemble forecasting system, we dynamically select high-skill ensemble subsets using the Niño3.4 index as reference, guided by RMSE and Pearson correlation coefficient. This approach departs from conventional arithmetic averaging. At an unprecedented 23-month lead time, it achieves substantial improvements—correlation increases to 0.71 (a relative gain of 172%, +0.43), and RMSE decreases by 22.5%. Notably, performance gains are most pronounced during climatological regime transitions, underscoring the critical importance of explicitly modeling ensemble skill heterogeneity for subseasonal-to-seasonal prediction.

Technology Category

• Machine Learning: Ensemble Methods
• Intelligent Robots: State Estimation
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

The accurate long-term forecasting of the El Nino Southern Oscillation (ENSO) is still one of the biggest challenges in climate science. While it is true that short-to medium-range performance has been improved significantly using the advances in deep learning, statistical dynamical hybrids, most operational systems still use the simple mean of all ensemble members, implicitly assuming equal skill across members. In this study, we demonstrate, through a strictly a-posteriori evaluation , for any large enough ensemble of ENSO forecasts, there is a subset of members whose skill is substantially higher than that of the ensemble mean. Using a state-of-the-art ENSO forecast system cross-validated against the 1986-2017 observed Nino3.4 index, we identify two Top-5 subsets one ranked on lowest Root Mean Square Error (RMSE) and another on highest Pearson correlation. Generally across all leads, these outstanding members show higher correlation and lower RMSE, with the advantage rising enormously with lead time. Whereas at short leads (1 month) raises the mean correlation by about +0.02 (+1.7%) and lowers the RMSE by around 0.14 °C or by 23.3% compared to the All-40 mean, at extreme leads (23 months) the correlation is raised by +0.43 (+172%) and RMSE by 0.18 °C or by 22.5% decrease. The enhancements are largest during crucial ENSO transition periods such as SON and DJF, when accurate amplitude and phase forecasting is of greatest socio-economic benefit, and furthermore season-dependent e.g., mid-year months such as JJA and MJJ have incredibly large RMSE reductions. This study provides a solid foundation for further investigations to identify reliable clues for detecting high-quality ensemble members, thereby enhancing forecasting skill.

Problem

• Improving long-term ENSO forecasting accuracy
• Identifying high-skill ensemble members for prediction
• Enhancing forecast performance during critical transition periods

Innovation

• Distills CNN ensemble subsets for enhanced prediction
• Identifies Top-5 members by RMSE and correlation
• Improves long-term ENSO forecasting accuracy significantly