GIS-Based Methodology for Crash Prediction on Single-Lane Rural Highways

Abstract

Due to the need to update the current guidelines for highway design to focus on safety, this study sought to build an accident prediction model using a Geographic Information System (GIS) for single-lane rural highways, with a minimum of statistically significant variables, adequate to the Brazilian reality, and improve accident prediction for places with similar characteristics. A database was created to associate the accident records with the geometric parameters of the highway and to fill in the gaps left by the absence of geometric highway plans through geometric reconstitution or semi-automatic extraction of highways using satellite images. The Generalized Estimating Equation (GEE) method was applied to estimate the coefficients of the model, assuming negative distribution of the binomial error for the count of observed accidents. The accident frequency and annual average daily traffic (AADT) were analyzed, along with the spatial and geometric characteristics of 215 km of federal single-lane rural highways between 2007 and 2016. The GEE procedure was applied to two models having three variations of distinct homogeneous segmentation, two based on segments and one based on the kernel density estimator. To assess the effect of constant traffic, two more variations of the models using AADT as an offset variable were considered. The predominant correlation structure in the models was the exchangeable. The principal contributing factors for the occurrence of collisions were the radius of the horizontal curve, the grade, segment length, and the AADT. The study produced clear indicators for the design parameters of roadways that influence the safety performance of rural highways.