Intigrating Geospatial Hotspot Analysis With Ensemble Learning For Accident Severity Prediction
Students & Supervisors
Student Authors
Supervisors
Abstract
Road traffic crashes continue to be a major public health problem in Dhaka, one of the world’s most crowded mega cities. Defective traffic flux patterns, the presence of infrastructural bottlenecks and non-standardized enforcement practices together lead to a higher incidence of collisions, which in turn highlights the need for an accurate prediction model and identification of high-risk areas. The current work presents a unified pipeline that combines geospatial hotspot detection with machine-learning based severity classification and explicitly accounts for high cardinality categorical variables and severe class imbalance. Bangladesh Road Accident Dataset that was collected from Dhaka Metropolitan Police reports of 2017-2022 was preprocessed with feature engineering, target encoding, and SMOTETomek resampling. Baseline models of Random Forest, LightGBM and XGBoost were benchmarked against a stacking ensemble. Empirical results show that the ensemble had an accuracy of 71% with an ROC-AUC of 0.91 for fatal crashes and a Cohen’s Kappa of 0.613. The methodology developed here provides a replicable, policy-relevant framework for enhancing road safety in Dhaka and other fast-urbanizing cities.
Keywords
Publication Details
- Type of Publication:
- Conference Name: 28th International Conference on Computer and Information Technology (ICCIT)
- Date of Conference: 19/12/2025 - 19/12/2025
- Venue: Long Beach Hotel
- Organizer: IEEE Bangladesh Section