A Physical-Environment-Driven Multi-Stream Deep Neural Network for Short-Term Heavy Precipitation Potential Identification

Abstract

Accurate identification of short-term heavy precipitation (STHP), governed by multiscale atmospheric processes, is critical for effective disaster prevention and mitigation. Conventional statistical methods often fail to capture the complex nonlinear relationships inherent in multidimensional atmospheric systems, whereas deep learning (DL) approaches exhibit significant advantages in multi-factor fusion. However, existing DL architectures typically process heterogeneous inputs through single-branch networks, resulting in feature redundancy, and suffer from limited interpretability—the "black box" problem—that undermines operational reliability. To address these challenges, this study proposes a physics-driven multi-stream deep neural network (MSDNN) framework, complemented by interpretability analysis using the integrated gradients (IG) method. Using observational data from 976 meteorological stations in Anhui Province and ERA5 reanalysis data from 2021 to 2024, we categorized 71 environmental physical variables into five distinct input streams according to physical characteristics: water vapor conditions, dynamic conditions, thermal instability, composite indices, and height layer properties. The split-merge architecture of MSDNN enables integrated processing of these five feature categories, achieving accurate identification of STHP. Results show that the MSDNN model achieved a threat score (TS) of 0.851 and a matthews correlation coefficient (MCC) of 0.844 on the test set, significantly outperforming both conventional ensemble learning methods (LightGBM, Random forest) and single-branch deep neural network (DNN). IG-based attribution analysis revealed that dynamic factors contributed most substantially to model performance (45%), followed by thermal (25%), vertical structure (15%), moisture-related (10%), and composite indices (5%). Furthermore, this study identified critical nonlinear thresholds: contribution polarity reversal at 80% relative humidity (700 hPa), strong sensitivity to upward vertical velocity (500 hPa). These quantified feature interactions provide data-driven physical insights into precipitation triggering mechanisms, elucidating the synergistic roles of moisture transport, dynamic lifting, and thermal instability. The MSDNN-IG framework establishes a technical pathway for severe convective weather identification that harmonizes accuracy with physical transparency, enhancing both the credibility and practical utility of AI methods in operational warning systems. Introduction