LAUSR

Abstract To perform crash analysis on large road networks, the network has to be segmented into smaller units. Existing approaches split road networks into road segments such that they result in (a)segments with constant length, (b) segments where attributes are homogeneous, (c) segments where crash-locations are interdependent. This study proposes a novel hybrid segmentation approach, especially to cater for network-level analysis. Road entities at network scale include intersections and mid-blocks which are dissimilar in their functionality as well as geometry. The effective zones of intersections and mid-blocks needs to be delineated, as there exists spatial correlation between crashes in them. The first level of the hierarchical hybrid segmentation uses crash locations to define intersections’ neighborhood. Subsequently, mid-blocks are segmented using attribute consistency. To illustrate its applicability on road networks in Low-and-Middle Income Countries (LMIC), we exemplify the methodology employing remote extraction of spatial attributes. The potential of Geographic Information System (GIS), Remote Sensing (RS), and Artificial Intelligence (AI) have been used for remote extraction of spatial features, including curvature, sight distance, land use, and network attributes. We compare the performance of hybrid segmentation with three other typical segmentation approaches. Crash susceptibility models using binary logistic regression, relating crashes and segment-specific attributes obtained with each of the four segmentation approaches were developed. The model developed with the proposed hybrid segmentation outperforms models with other segmentation approaches, in terms of goodness-of-fit and prediction accuracy. The proposed methodology can be readily used by practitioners for crash analysis.