Machine learning prediction of summer extreme precipitation days in the middle and lower Yangtze River with SHAP explanation
Chunyan Xiao, Anmin Duan, Yuheng Tang, Bin Tang, Qilu Wang, Xianyi Yang
Abstract
Accurately predicting summer extreme precipitation days in the middle and lower reaches of the Yangtze River (YREPD) is critical for disaster mitigation yet remains challenging. This study introduces three advanced machine-learning algorithms—Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost)—to enhance seasonal forecasts of the leading mode of YREPD. Key predictors from preceding-season observational datasets were rigorously selected based on Spearman's rank correlation and physical mechanisms. Their ensemble achieves a temporal correlation coefficient (TCC) of 0.73 ( P < 0.05) and a root mean square error (RMSE) of 0.87 on test data. Compared to the Support Vector Regression (SVR) model (TCC = 0.59, RMSE = 0.99), the ensemble increases TCC by 23.7 % and reduces RMSE by 12.1 %. Moreover, the ensemble model significantly outperforms traditional multiple linear regression as well as persistence and climate baseline approaches. Furthermore, Shapley Additive Explanations (SHAP) analysis identifies Eurasian snow depth anomalies and tropical Atlantic sea surface temperature anomalies as dominant drivers. Crucially, the SHAP analysis uncovers prevalent nonlinear predictor relationships and significant predictor interactions, providing novel physical insights into YREPD variability. This work demonstrates the enormous potential of machine learning ensembles for advancing seasonal extreme precipitation prediction and provides valuable mechanistic understanding for future forecasting improvements.