Assessing Pavement Conditions Impact on Traffic Crashes: An Interpretable Machine-Learning Approach to Identify Crash Risk Factors and Performance Thresholds

Pavement surface condition may have a significant impact on traffic accidents, but its influencing mechanisms remain unclear, and safety-oriented maintenance thresholds are still lacking. To addresses these gaps, this study develops a highway crash regression model by comparing generalized linear models with tree-based machine learning (ML) approaches to establish the relationship between various factors and crash counts. An interpretability algorithm is employed to examine the combined effects of pavement, traffic, and meteorological conditions, with a particular focus on pavement-related factors. Minimum pavement performance requirements for safety considerations under varying traffic levels are then established, providing guidance for safety-oriented maintenance decision-making. Results reveal that annual average daily traffic (AADT) has the greatest impact, with higher traffic volumes correlating with increased crash counts. Additional risk factors include a higher proportion of trucks, bridge presence, frequent rainfall, and road ages exceeding ten years. Poor skid resistance, excessive roughness, and deep rutting detrimentally affect safety, particularly under high traffic volumes, whereas transverse cracks may reduce crashes by promoting heightened driver caution. For high-AADT roadways (≥30,000), this study identifies minimum performance thresholds of 40 for the side-way force coefficient (SFC), 2 m/km for the international roughness index (IRI), and 10 mm for the rutting depth (RD), whereas lower AADT levels do not necessitate specific requirements. These findings enhance the understanding of pavement-related crash risks and provide actionable insights for developing safety-oriented maintenance planning.