Accurate analysis and forecasting of short-term heavy rainfall (hourly rainfall ≥ 20 mm) are crucial for extending warning, enabling targeted preventive measures, and supporting efficient resource allocation. In recent years, machine learning techniques combined with atmospheric physical variables have offered promising new approaches for analyzing and predicting and forecasting short-term heavy rainfall. However, these methods often lack transparency, which hinders the interpretation of key atmospheric physical variables that drive short-term heavy rainfall and their coupling mechanisms. To address this challenge, the present study integrates the interpretable SHAP (SHapley Additive exPlanations) framework with machine learning to examine potential relationships between widely used atmospheric physical variables and short-term heavy rainfall, thereby improving model interpretability. CatBoost models were constructed based on multiple feature-input strategies using 71 physical variables across five categories derived from ERA5 reanalysis data, and their performance was compared with two benchmark algorithms, XGBoost and LightGBM. The SHAP method was subsequently applied to quantify the contributions of individual features and their interaction effects on model predictions. The results indicate that (1) the CatBoost model, utilizing all 71 physical variables, outperforms other feature combinations, with an AUC of 0.933, and F1 score of 0.930, and a Recall of 0.954, significantly higher than the XGBoost and LightGBM models; (2) Shapley value analysis identified 500 hPa vertical velocity, the A-index, and precipitable water as the most influential features on model performance; (3) The predictive mechanism for short-term heavy rainfall is fundamentally bifurcated: negative instances are classified through the discrete main effects of individual features, whereas positive event detection necessitates a sophisticated coordination of intrinsic main effects and synergistic interactions. Among the feature categories, the horizontal and vertical wind fields, stability and energy indices, and humidity-related variables exhibited the highest contribution ratios, with wind field features demonstrating the strongest interaction effects. The results confirm that integrating atmospheric physical variables with the CatBoost ensemble learning approach significantly improves short-term heavy rainfall identification. Furthermore, incorporating the SHAP interpretability framework provides a theoretical foundation for elucidating the mechanisms of feature influence and optimizing model performance.
An et al. (Tue,) studied this question.