Beyond aggregation: a multi-scale, interpretable machine learning assessment of built environment effects on heavy-truck crash severity in Fortaleza, Brazil

This study investigates the spatial determinants of freight-related injury crashes in Fortaleza, Brazil, focusing on three outcomes: Possible Injury Crashes (PIC), Evident Injury Crashes (EIC), and Fatal Injury Crashes (FIC). A rich urban dataset incorporating land use, freight infrastructure, road design, and traffic characteristics was analyzed across three spatial aggregations: hexagons, aggregated neighborhoods, and census tracts. Two complementary modeling techniques were employed: Zero-Inflated Poisson (ZIP) models were used to account for excessive zeros in crash counts, while Extreme Gradient Boosting (XGBoost) provided a high-performing machine learning approach capable of capturing nonlinear relationships. XGBoost results show that the density of traffic signals, density of freeway lane mileage, and density of truck volumes are the dominant factors influencing freight-related injury risk across spatial scales. ZIP results highlighted the significant impact of freight proximity (e.g., distance to ports, airports, and intermodal terminals), infrastructure density (e.g., traffic signals, freeway lanes), and land-use intensity (e.g., job density and retail zones) on crash occurrence and severity. To interpret the complex patterns captured by XGBoost, Shapley Additive Explanations (SHAP) analysis was used. SHAP dependence and interaction plots revealed threshold effects and spatial interactions such as increased EIC and FIC risk in areas with high truck exposure and dense signal or freeway infrastructure. Notably, PIC risk was highest close to the airports, especially where primary roads and poles coexisted. The study demonstrates that both linear and nonlinear modeling frameworks are necessary to uncover the layered nature of freight-induced crash risks. Policy implications include optimizing traffic control density, enhancing safety in freight corridors, and prioritizing pedestrian protection near logistics hubs. By integrating statistical and explainable machine learning methods, the study offers actionable insights for data-driven urban safety and mobility planning.