IDENTIFYING ROAD ACCIDENT HOTSPOTS USING LINEAR NETWORK POINT PROCESSES: A CASE STUDY IN THE KANDY POLICE DIVISION, SRI LANKA

Abstract

Road accidents are spatially constrained events that require analytical approaches that account for the underlying transport network. Traditional area-based methods often underestimate accidents along roads, as they ignore road geometry. Modeling crashes as points on a road network improves intensity estimation. Although conventional spatial point pattern techniques are widely applied in accident analysis, approaches that incorporate the linear structure of roadways are more suitable, as accidents are constrained to occur along these networks. This study aims to identify road accident hotspots using a Linear Network Point Processes (LNPP) approach, which is more suitable than traditional area-based methods for linear networks. By analyzing accident data from the Kandy Police Division in Sri Lanka, it seeks to enable precise hotspot detection and support targeted strategies for improving road safety. Accidents occur along roads; therefore, LNPP on a metric graph G = (V, E) was used, where V represents intersections and E represents road segments. The linear network kernel density estimation was used at a location x on a road segment to estimate event intensity per unit length along the network, based on the shortest path distances between observed points. Analyses used Gaussian and Epanechnikov kernels. The optimal bandwidth was selected using the Likelihood Cross-Validation (LCV) method and Scott's Rule. Diggle’s edge correction method was applied to reduce boundary bias. Hence, four models were fitted. A total of 2,043 crash events were represented as a point pattern on a road network comprising 108,584 line segments and 110,361 vertices. Both Gaussian and Epanechnikov kernels produced similar density estimates on the road network for a given bandwidth. However, the LCV method, which estimates kernel density on linear networks, provided a more accurate bandwidth than Scott’s rule. The smoothing bandwidths were 1.98 km and 5.07 km for the Scott and LCV methods, respectively. Residuals range from 0–45 for Scott’s rule and only 0–13 with LCV. Several high-density road accident segments were found, highlighting spatial heterogeneity in accident risk along the network. LCV outperformed Scott’s rule in bandwidth selection, reducing residuals and enhancing hotspot detection, demonstrating the effectiveness of LNPP for network-constrained traffic safety analysis.