Motorcycle crashes impose a substantial public safety burden in low- and middle-income countries, yet conventional injury severity models often overlook gender-related heterogeneity in crash mechanisms. This study proposes a three-layer hierarchical explainable machine-learning framework to investigate motorcycle injury occurrence and injury severity among motorcyclists on Thai national highways using 14,464 crash records from the Highway Accident Information Management System (HAIMS) during 2017-2022. The framework comprises an injury occurrence model, a population-level injury severity model, and gender-stratified injury severity submodels for male and female riders. Ensemble machine-learning algorithms are optimized using Bayesian hyperparameter optimization and class-imbalance resampling, while SHapley Additive exPlanations (SHAP) enhanced model interpretability. CatBoost achieved the best predictive performance in three of the four prediction models, whereas Random Forest performed best for the male-specific model. SHAP analysis revealed distinct gender-specific injury severity mechanisms. Fatal injury pathways among male riders are primarily associated with heavy-truck interactions, nighttime travel, rainy-season conditions, and head-on collisions, whereas female riders are more strongly characterized by moderate injury pathways involving passenger-car interactions, daytime exposure, and mixed urban traffic environments. Because all female observations in the dataset are injury cases, these findings should be interpreted as severity-conditional rather than reflecting overall injury risk. The results demonstrate that gender-stratified modeling provides important analytical value by revealing heterogeneity masked by the pooled models, as pooled models mask important heterogeneity in injury severity mechanisms. The proposed framework provides an interpretable basis for evidence-based motorcycle safety management by supporting targeted interventions within a 6Es framework, thereby contributing to more effective road safety policies in Thailand and other motorcycle-dependent low- and middle-income countries.
In developing countries, urban tourism systems increasingly rely on large-capacity buses, whose stopping, boarding, and manoeuvring operations often create localised congestion and traffic bottlenecks. However, contemporary routing algorithms can create a paradox whereby solutions that are optimal for individual tourist buses may collectively induce emergent network-wide congestion when deployed simultaneously across the entire fleet. This paper introduces a cross-scale evaluation framework that explicitly links vehicle-centric routing decisions with macroscopic transport network performance. A microscopic traffic simulation of Nha Trang, Vietnam, developed in SUMO and calibrated using empirical intersection signal timings and stochastic bus dwell times, serves as the experimental testbed. Three routing paradigms are systematically benchmarked: a fixed-route baseline, a metaheuristic approach based on a Genetic Algorithm (GA), and an AI-driven method using Double Deep Q-Network (Double DQN). Stress testing with near-critical background traffic, evaluated across multiple random seeds to account for stochastic demand, reveals that uncoordinated static routing results in a severe 71% loss in network throughput. While static metaheuristic optimisation reduces this loss to 37%, it remains vulnerable to dynamic bottlenecks. Conversely, the Double DQN agent continuously redistributes the spatial load, suppressing average vehicle time loss by over 60% and limiting throughput degradation to a statistically significant 24% (p < 0.01). These findings demonstrate that optimising routes independently for individual buses is insufficient; real-time adaptive routing is a potential mechanism to help preserving network-level capacity, sustain Macroscopic Fundamental Diagram (MFD) flows, and mitigate localised CO2 emission hotspots.
Traffic congestion often arises from intrinsic traffic instabilities that generate stop-and-go waves and reduce traffic efficiency. While nonlocal perception and connected vehicle technologies have each been shown to improve traffic stability, their combined macroscopic effects in heterogeneous traffic streams remain incompletely understood. This paper develops a nonlocal cooperative multiclass macroscopic traffic model within a second-order continuum framework. The model integrates heterogeneous vehicle classes, nonlocal density perception, and vehicle-to-vehicle (V2V) communication. Cooperative interaction is represented through a density-normalised communicated velocity that distinguishes instantaneous sensing of vehicle presence from delayed reception of velocity information. Communication delays are incorporated into the velocity alignment term, enabling their influence on stability to be examined analytically. A linear stability analysis yields explicit conditions describing the combined effects of nonlocal perception, cooperative communication, and communication delay on traffic stability. Nonlocal perception weakens the amplification of density disturbances, while cooperative communication further suppresses traffic oscillations through velocity alignment and spatial averaging. Although communication delays reduce effective damping, stability is maintained for moderate delay magnitudes. Numerical simulations support the analytical results and illustrate the distinct roles of anticipation and cooperation. Nonlocal perception mitigates oscillations but does not fully suppress stop-and-go waves near capacity, whereas cooperative V2V communication provides substantial stabilisation even at moderate penetration rates. Strong cooperative regimes produce the greatest reduction in traffic oscillations, although these benefits gradually diminish as communication delays increase. These findings highlight the complementary roles of nonlocal anticipation and cooperative connectivity in improving stability in mixed-autonomy traffic streams.
