Measuring ENSO Predictability Sources Using Machine Learning Techniques

Abstract

Important global climate anomalies, such as the El Niño-Southern Oscillation (ENSO), can impact socio-economic systems, water resources, and agriculture. Strategies for disaster preparedness and climate adaptation rely on accurate ENSO event predictions. Here, we use machine learning techniques to investigate potential causes of ENSO prediction. The ENSO indices (Niño 3.4 and Niño 4) were modelled using historical records of sea surface temperature (SST), sea level pressure (SLP), and subsurface ocean temperatures. To unravel the data’s patterns, including its non-linear linkages and temporal dependencies, we employed Random Forest (RF), Gradient Boosting (GB), and Long Short-Term Memory (LSTM) networks. To evaluate model performance, we used RMSE, MAE, and correlation coefficients for performance metrics, as well as feature importance metrics and seasonal analysis to evaluate phase-dependent predictability (i.e., modelling winter ENSO and summer ENSO). The results indicate that the LSTM model yields performance levels superior to the tree-based models and predicts the highest levels of ENSO prediction accuracy; the strongest predictors were SST anomalies in the Niño 3.4 region and subsurface temperatures. In terms of seasonal predictability, we found that ENSO events during winter months were more predictable than summer months, which aligns with phase-locking behaviour. Overall, this research shows that machine learning can provide reliable understanding of ENSO dynamics and identify the important climate drivers, which together provides a step towards better forecasting and early warning capacity.