A transferable deep learning approach for crash severity prediction based on road environmental factors

This is the authors' abstract. We don't add key points for this paper.

Predicting crash severity and identifying influencing factors are vital prerequisites for implementing safety countermeasures. While previous studies have largely concentrated on individual, difficult-to-control factors (e.g. distractions), this research shifts attention to road environmental influences to inform the design of road safety strategies. Specifically, this study proposes a four-step predictive model, which captures road environmental factors through web crawlers based on time and location information, to forecast traffic crash severity and enhance proactive risk assessment in urban transportation systems. An artificial neural network (ANN) model was employed at the last step for predicting crash severity. The study trained the model using the Xichang dataset and validated it with the Suqian dataset, demonstrating model transferability. The developed model achieves high predictive accuracy, with F1-scores of 85.46% for the training dataset and 82.12% for the validation dataset. The proposed framework enables real-time generation of traffic crash distribution maps and crash hotspot identification within a region. Particularly in developing countries, the proposed model can address challenges such as limited data availability in accident records and the migration application of models. These insights can provide traffic authorities with data-driven risk mitigation strategies while simultaneously issuing predictive alerts to drivers for enhanced situational awareness.