Decoding sub-seasonal predictors of extreme heat with interpretable machine learning

Jagger Alexander, Zong-Liang Yang

As climate change accelerates, heat waves are becoming more frequent, intense, and deadly. A better understanding of sub-seasonal predictors of extreme heat is crucial for adaptation efforts. This study utilizes an interpretable machine learning model, implementing Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP), to evaluate the predictive strength of various climate factors—including local weather, global climate indices, geopotential heights, soil moisture, and sea surface temperatures—on heat waves, as classified by consecutive days of extreme daily maximum temperatures across six North American cities over the past 45 years. This model demonstrates strong predictive performance for extreme heat on the sub-seasonal time scale, with sea surface temperatures and soil moisture features emerging as more influential than atmospheric features, though key regional differences in feature importance and feature dependence are shown through variation between chosen cities. Bivariate relationships between MJO phase and amplitude are also uncovered through analysis of model predictions. This method shows promise for rapid application to other regions and also serves as a foundation for integration with dynamical modeling approaches, advancing sub-seasonal extreme heat forecasting more broadly.