Answers are grounded in this paper and cite their sources.
Three key challenges hinder the identification of short-term heavy precipitation (STHP) over the Yangtze River Delta region: feature selection, sample imbalance, and poor physical performance of models. To address these issues, this study developed a physically consistent cascade recognition model based on ERA5 reanalysis data from 2020 to 2024 and high-density ground-based observations. The model first reduced 71 candidate factors to 15 core physical variables through a stepwise feature engineering procedure, then systematically compared four strategies for handling sample imbalance, and finally employed Shapley Additive Explanations (SHAP) attribution to validate the physical interpretability of the cascade model. Results show that feature selection reduced model training time from 8.89 s to 3.61 s and improved the critical success index (CSI) from 0.0059 to 0.0076. The proposed four-level cascade recognition model achieved the best performance, with a global optimal CSI of 0.0398, representing improvements of 52% and 23% over the baseline weighted adjustment method (0.0262) and the ensemble voting method (0.0323), respectively. SHAP-identified dominant factors exhibited a systematic evolution with increasing cascade levels: the initial stage is governed by A-index–dominated convective instability, jointly modulated by moisture factors such as 700 hPa specific humidity and precipitable water vapor (PWV); the final stage evolves into a dynamic triggering regime dominated by convective available potential energy (CAPE), vertical wind shear, and energy helicity index (EHI). By simulating the physical evolution of convection through task decomposition, the proposed cascade recognition framework effectively mitigates the long-tail effect in precipitation distribution, providing an effective solution for refined STHP monitoring and intelligent diagnosis of convective environments. While this study utilizes ERA5 reanalysis for diagnostic modeling, the framework offers a robust foundation for future transition to operational forecasting when integrated with Numerical Weather Prediction (NWP) outputs.
No takes yet. Share an insight, caveat, or question.
An et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: