Urban accidents and crashes are one of the major causes of death and injuries in metropolitan areas like Athens, Greece, where intense traffic strain and urban dense infrastructure increase exposure risk. In this research, we develop an interpretable ensemble machine learning approach to understand and interpret the severity of crashes, incorporating data from 2016 to 2020 provided by Smart Maps and city traffic records. Random Forests and XGBoost models were developed based on an integrated dataset that includes spatiotemporal, geo-infrastructural, and crash-related features. The evaluation of the models showed strong predictive capabilities, with XGBoost slightly outperforming Random Forests on all predictive accuracy indicators. To improve the model’s interpretability and transparency, we applied SHAP (SHapley Additive exPlanations) analysis to identify the most important crash severity predictor variables. After applying SHAP analysis to the best-performing model, it was discovered that the most significant metrics were crash severity, mean fatalities, and longitude (caplon). Further, analysis through Partial Dependency and SHAP revealed the nonlinear and spatial interactive effects of urban configuration/ traffic and spatial crashes. Subsequent use of Smart Maps geolocation data for spatial risk mapping depicted high-risk municipalities concentrated in central and southern Athens. Such spatial knowledge offers a decision-advocacy tool for urban planners and policymakers, helping them to focus on road safety improvements and optimize road infrastructure in targeted municipalities. This study shows the potential of interpretable ensemble learning, coupled with Smart Maps geospatial intelligence, to develop advanced analyses of road safety for municipalities and to take proactive management of crash risk.
| ID | pc676 |
| Presentation | |
| Full Text | |
| Tags |
