Investigating the formation mechanism of merging/diverging collision risk in short weaving segments: An integrated approach using spatial-temporal risk field and explainable machine learning

Merging and diverging operations in short weaving segments on urban expressways are high-risk scenarios that frequently lead to collisions. Investigating the formation mechanism of merging/diverging collision risk is essential for developing effective control strategies. However, the scarcity of micro-level collision risk events and the unclear correlation between macro- and micro-level risks hinder the integration of risk identification and control strategies. To address these challenges, we integrate the spatial–temporal risk field (STRF) with explainable machine learning to construct a systematic framework for investigating the formation mechanism of collision risks in short weaving segments. Using seven representative weaving segments with distinct configurations on the eastern and western expressways of Changchun, we collected 3.6 h of aerial video data and extracted vehicle trajectories through an improved YOLO algorithm. Based on STRF, we developed a micro-level collision risk event identification model that simultaneously captures rear-end, lateral, and environmental risks. STRF clearly demonstrates the risk distribution characteristics of merging and diverging in different weaving segments. We then built a three-dimensional macro-explanatory variable system encompassing lane-, segment-, and factor-level attributes. Through a three-stage refinement process, the optimized XGBoost model achieved superior performance across 14 training tasks, consistently outperforming all baseline models. By integrating SHAP interpretability analysis, we revealed both the commonalities and heterogeneities in how macro-level factors influence micro-level risks under varying movement patterns and weaving segment configurations. Finally, we proposed targeted macro-level control strategies tailored to distinct traffic flow characteristics, geometric layouts, and spatial distributions of risk points across different types of short weaving segments.