Applying XAI to Optimized XGBoost Models: Elucidating Multi-type Predictor Impacts for Short-Term Precipitation Forecast

The inherent black-box nature of machine learning (ML) models limits their interpretability and broader application in heavy precipitation forecasting. Evaluating the reliability of these models involves analyzing the link between predictions and predictors. In this study, ERA5 reanalysis data, CERES satellite observations, and ground-based meteorological observatories were utilized to compile more comprehensive multi-type predictors for developing a Bayesian optimized XGBoost model for the nowcasting of heavy precipitation in the Guangdong-Hong Kong-Macao Greater Bay Area during the pre-summer rainy season. A comparison of model performance with different combinations of input features and classical machine learning algorithms demonstrated that the Bayesian optimized XGBoost model achieved the best overall performance, with an average Critical Success Index of 68.30%. Permutation Importance (PI) and shapley Additive Explanations (SHAP) methods were utilized to interpret feature effects in heavy precipitation forecasting. The results indicated that precipitable water vapor (PWV), cloud, relative humidity, and seasonal and diurnal variables had more significant effects on the model output as individual features. Furthermore, the collective influence of derivatives from PWV and meteorological parameters (e.g., temperature, relative humidity, pressure and dew point temperature) showed a significant enhancement over their individual impacts, indicating synergistic interactions among these predictors. Applying explainable artificial intelligence (XAI) to ML models helps understand how models utilize features for forecasting, enhances the reliability of forecasts, and guides feature selection and the mitigation of overfitting phenomena.