Improving Road Accident Severity Classification with Cluster-Based Severity Resampling: A Hybrid Approach

In response to the rising number of vehicles and corresponding increase in road accidents within urban environments, this study introduces a hybrid approach that combines clustering and classification of road accident severity to aid city planning and enhance public safety in smart cities. To address the imbalance in real-world accident data, which particularly affects the accurate prediction of fatal accidents often underrepresented in datasets, we employed clustering as a foundational structure for data segmentation. We developed a method based on two clusters for severity resampling. Within each cluster, we applied targeted resampling techniques, enhanced by SMOTE and ADASYN, to achieve a balanced representation of severity classes. Our comparative analysis reveals that our approach significantly outperforms models trained on both the original imbalanced dataset and those modified by standard resampling techniques. Specifically, we observed an increase in G-mean from 0.02 and 0.29 to 0.70 and an improvement in AUC-ROC by 35%, and 38% respectively. These enhancements highlight the efficacy of integrating clustering techniques into accident severity classification. This robust and adaptable methodology consistently improves predictions of accident severity in urban environments, demonstrating its effectiveness across various methodological setups.