Street view-based infrastructure assessment of motorcycle hotspots: A national road network analysis in Thailand

Motorcycle casualties in Thailand are strongly concentrated on a small proportion of the road network, yet road infrastructure assessment at the national scale remains constrained by incomplete road-inventory data and the cost of field audits. This study develops an integrated framework for identifying motorcycle crash hotspots and examining their visually observable road-environment correlates. Motorcycle crash records from the Thailand Road Accident Management System (2019-2024) were analysed using Network Kernel Density Estimation (NKDE) on 100 m road segments. The NKDE output is interpreted as severity-weighted crash concentration, not exposure-normalised crash risk. For hotspot and matched control segments, representative Mapillary Street-View Imagery (SVI) was processed using a Vision-Language Model (VLM) pipeline to extract 27 iRAP-aligned attributes. Only manually validated, reliable VLM-derived attributes were retained and combined with route-level traffic-volume proxies in an XGBoost classifier, with SHAP used to interpret feature contributions. The selected NKDE configuration showed strong concentration: the top 1% of network length captured 52.2% of severity-weighted motorcycle crashes. Model interpretation indicates that hotspot classification is associated with commercial access activity, traffic-exposure proxies, constrained median and roadside environments, and rigid roadside objects. These associations are interpreted cautiously as hotspot indicators rather than causal effects. The framework provides a scalable approach for linking realised motorcycle casualty concentration with SVI-derived infrastructure hazard indicators in data-constrained settings.