Predicting motorcycle crash severity on Thailand's curved roadways: A deep learning approach

Motorcycle crashes on curved roadways present unique safety challenges due to complex vehicle dynamics and increased risk exposure. In Thailand, where motorcycles dominate transport, crash data from 2016 to 2022 reveals a significantly higher fatal/severe injury rate on curved segments (56.77%) compared to straight segments (44.41%), despite curves comprising a smaller portion of the road network. This study develops Convolutional Neural Network (CNN) models to predict binary crash severity outcomes on curved roadways using 2679 crash records from Thailand's Highway Accident Information Management System. The research systematically compared four CNN architectures with varying complexity (depths 3 to 10+) and found that simpler architectures (Layout I with depth 3) outperform deeper configurations (accuracy = 0.634, summary score = 53.61%). SHapley Additive exPlanations (SHAP) were applied to identify key risk factors and interaction effects. Results show that factors such as large trucks, head-on collisions, depressed medians, darkness conditions, speeding, two-lane roads, and work zones are associated with higher predicted crash severity on curves, while urban settings, side-swipe crashes, and barrier medians are associated with lower predicted severity. SHAP interaction analysis identified combinations associated with elevated severity predictions, notably large trucks operating in darkness and head-on collisions regardless of median type. These findings support targeted interventions including differential speed management for large vehicles, enhanced illumination strategies, barrier median installations, and modified work zone protocols for curved segments. This research advances both methodological approaches for crash severity prediction and practical applications for motorcycle safety in countries with similar transportation contexts.