Machine-Learning-Based Prediction of Winter Traffic-Accident Risk Using Meteorological Factors: A Case Study of Five ASOS Stations in Gyeonggi-do, South Korea

This study aimed to develop a machine-learning-based predictive model for winter traffic-accident risk using meteorological variables and analyze the influence of weather factors on accident occurrence. The study area consisted of five cities in Gyeonggi-do, South Korea—Dongducheon, Paju, Yangpyeong, Icheon, and Suwon—where automated surface observing system (ASOS) stations are located. Winter traffic-accident data from 2015 to 2023 were collected from the traffic accident analysis system (TAAS), and meteorological observations were obtained from the Korea Meteorological Administration. The dependent variable was constructed based on accidentoccurrence levels; ten meteorological variables, including temperature, solar radiation, wind speed, humidity, and cloud cover, were the independent variables. Softmax Regression, Random Forest, and XGBoost models were employed for prediction, and their performances were compared in terms of prediction accuracy and feature importance. The results revealed that the Random Forest model demonstrated the highest predictive performance across all regions. In addition, solar radiation, wind speed, and cloud cover significantly influenced accident occurrence beyond the widely assumed impact of temperature alone. These findings highlight the potential of machine-learning models as decision-support tools for winter road-safety management and proactive accident-prevention policies.