JultikaUniversity of Oulu repository

Abstract

The primary focus of this study was to investigate the impact of weather and road conditions on the severity of accidents and to determine the feasibility of machine learning models in accurately predicting the likelihood of such incidents. The research was centered on two key research questions.

Firstly, the study examined the influence of weather and road conditions on accident severity and identified the most related factors contributing to accidents. We utilized an open-source accident dataset, which was preprocessed using techniques like variable selection, missing data elimination, and data balancing through the Synthetic Minority Over-sampling Technique (SMOTE). Chi-square statistical analysis was performed, suggesting that all weather-related variables are more or less associated with the severity of accidents. Visibility and temperature were found to be the most critical factors affecting the severity of road accidents. Hence, appropriate measures such as implementing effective fog dispersal systems, heatwave alerts, or improved road maintenance during extreme temperatures could help reduce accident severity.

Secondly, the research evaluated the ability of machine learning models including decision trees, random forests, naive bayes, extreme gradient boost, and neural networks to predict accident likelihood. The models’ performance was gauged using metrics like accuracy, precision, recall, and F1 score. The Random Forest model emerged as the most reliable and accurate model for predicting accidents, with an overall accuracy of 98.53%. The Decision Tree model also showed high overall accuracy (95.33%), indicating its reliability. However, the Naive Bayes model showed the lowest accuracy (63.31%) and was deemed less reliable in this context.

It is concluded that machine learning models can be effectively used to predict the likelihood of accidents, with models like Random Forest and Decision Tree proving the most effective. However, the effectiveness of each model may vary depending on the dataset and context, necessitating further testing and validation for real-world implementation.

These findings not only provide insight into the factors affecting accident severity but also open a promising avenue in employing machine learning techniques for proactive accident prediction and mitigation. Future studies can aim to refine the models further and potentially integrate them into traffic management systems to enhance road safety.