Road traffic crashes are a global challenge, claiming over 1.2 million lives annually, with low- and middle-income countries bearing a disproportionate burden. While fatalities have declined slightly globally since 2013, Africa experienced a 17% increase, and Ethiopia was among the 28 countries with increasing fatalities. Addis Ababa faces growing challenges in urban mobility and road safety due to rapid population growth and lateral expansion. The city accounts for nearly 40% of injury crashes and 13% of fatal crashes nationwide. In 2022 alone, Addis Ababa recorded 403 traffic fatalities, which is nearly ten times the number of fatalities in Berlin and four times higher than in London. Addressing this challenge requires a data-driven, spatially informed, state-of-the-art approach to accurately identify crash hotspots and hot periods, enabling road and traffic enforcement agencies to use their limited resources effectively. This study employed spatial pattern analysis on a three-year geo-referenced crash dataset from Addis Ababa. Cluster analysis, spatial autocorrelation, and Network Kernel Density Estimation (Network KDE) were conducted to examine crash severity patterns and identify severity-specific hotspots. The results demonstrate significant spatial clustering of crashes, with minor injury crashes concentrated in central areas, while severe and fatal crashes were more prevalent along arterial corridors and in peripheral zones. Temporal analysis further revealed that fatal crashes peaked in the evening near the inner and intermediate zones, while daytime peaks were more common in the outer area. Crash frequency was modeled using Poisson, Negative Binomial, and their geographically weighted counterparts, while crash severity was modeled using machine learning methods, including Random Forest variants and Support Vector Machine. Geographically Weighted Negative Binomial Regression provided the best fit for frequency modeling by capturing both spatial heterogeneity and overdispersion. For severity prediction, Random Survival Forest and Weighted Random Forest outperformed other approaches, particularly in detecting serious and fatal crashes. Victim type, collision type, and distance from the mean center were identified as key predictors of crash severity. The study contributes to the existing body of knowledge in road safety employing spatial analysis, particularly in low- and middle-income urban contexts. The proposed framework for hotspot identification extends the practical application of Network KDE, offering road agencies a systematic tool for prioritizing high-risk locations and providing direction for future research.
Wondwossen Gedamu (Mon,) studied this question.