Resilient road safety modeling through spatially disaggregated explainable AI

Area of Science:

• Transportation Safety
• Machine Learning Applications
• Spatial Analysis

Background:

• Traffic accident severity exhibits significant spatial disparities, impacting public health and equity.
• Road safety is intrinsically linked to the UN Sustainable Development Goals (SDG 3 and SDG 11).
• Existing pooled analyses often mask critical urban-rural differences in crash outcomes.

Purpose of the Study:

• To develop an explainable, spatially segmented machine learning framework to analyze urban-rural heterogeneity in traffic accident severity.
• To identify distinct risk factors influencing crash outcomes in urban and rural settings.
• To inform context-specific interventions for enhanced transport safety and equity.

Main Methods:

• Utilized disaggregated traffic accident data from Kent, UK (2022-2024).
• Employed a spatially segmented machine learning approach, treating urban and rural areas as distinct analytical units.
• Compared five machine learning models, selecting Random Forest for its superior performance, and interpreted results using SHapley Additive exPlanations (SHAP).

Main Results:

• Random Forest model demonstrated the best performance in predicting crash severity.
• SHAP analysis revealed that behavioral risk factors are more influential in urban areas.
• Infrastructure-related risk factors were found to be more dominant in rural environments.

Conclusions:

• Urban and rural areas exhibit distinct patterns of traffic accident risk factors.
• Context-sensitive, evidence-based interventions are necessary to ensure transport equity.
• Findings support sustainable governance by enabling spatially adaptive risk mitigation and resilient transport infrastructure.