GIS and Machine Learning for Identifying Crash Hotspots: Amman, Jordan as a Case Study

GIS is the most significant aid in visualizing and analyzing accidents because it easily provides a way to visualize hazardous areas on highways. The presented study uses accident data with high point density based on 166,000 points collected by the Police and Security Department in Jordan. After processing the data, 106,220 accidents that occurred during 2022 were identified. Geospatial analysis conducted using network-constrained kernel density estimation indicated high densities of accidents at the north and center of Amman along major urban roads predominantly with moderate incidence in the suburbs and low frequencies in rural areas. The study area was further reclassified into polygon features to provide accident counts correlated to road segments with increased crash frequency particularly ten road segments including Hussein Khawaja Street and Al Jama’a Street where speeds between 50 to 60 km/h enhance susceptibility. The Getis-Ord Gi* hot spot analysis generates an ICOUNT field and applies K-nearest neighbors with multiple comparison testing the strength of results. DBSCAN supports finding and categorizing clusters within the data, effectively managing noise or outliers with parameters like eps and MinPts. The study identifies the ideal epsilon through a k-distance plot, recommending values ranging from 800 to 1000 meters, while further tuning between 700 and 3000 meters enhances clustering performance. The analysis revealed 10 micro-risk zones, with a focus on sub-clusters 3 and 6, which had high accident rates at night with an average speed of 70 km/h. Sub-cluster 5 indicated increased female involvement during inclement weather, while sub-cluster 2 showcased daytime collisions correlated with behavior-driven risks at moderate speeds.