Predicting and Understanding the Pacific Decadal Oscillation Using Machine Learning

Abstract

The Pacific Decadal Oscillation (PDO), the dominant pattern of sea surface temperature anomalies in the North Pacific basin, is an important low-frequency climate phenomenon. Leveraging data spanning from 1871 to 2010, we employed machine learning models to predict the PDO based on variations in several climatic indices: the Niño3.4, North Pacific index (NPI), sea surface height (SSH), and thermocline depth over the Kuroshio–Oyashio Extension (KOE) region (SSH_KOE and Ther_KOE), as well as the Arctic Oscillation (AO) and Atlantic Multi-decadal Oscillation (AMO). A comparative analysis of the temporal and spatial performance of six machine learning models was conducted, revealing that the Gated Recurrent Unit model demonstrated superior predictive capabilities compared to its counterparts, through the temporal and spatial analysis. To better understand the inner workings of the machine learning models, SHapley Additive exPlanations (SHAP) was adopted to present the drivers behind the model’s predictions and dynamics for modeling the PDO. Our findings indicated that the Niño3.4, North Pacific index, and SSH_KOE were the three most pivotal features in predicting the PDO. Furthermore, our analysis also revealed that the Niño3.4, AMO, and Ther_KOE indices were positively associated with the PDO, whereas the NPI, SSH_KOE, and AO indices exhibited negative correlations.