Crash Severity Prediction in Cold Regions Using Machine Learning: Integrating Weather, Lighting, and Road Surface Conditions

This study presents a machine learning–driven framework for predicting crash severity in cold-region environments, where inclement weather, limited daylight, and adverse surface conditions are significant contributing factors. The analysis incorporates historical crash data from Alaska (1990–2012) and Michigan (2016–2023), focusing on structured crash reports that include temporal, environmental, and roadway-related features. Models such as Random Forest were applied to examine the predictive power of variables including hour of the day, month, year, weather condition, lighting condition, road surface type, and roadway classification. Results indicate that crash severity is highly influenced by temporal and environmental factors, with performance metrics reflecting strong classification of less severe crashes. The inclusion of weather and lighting significantly improves model robustness, especially in rural and icy road contexts. To assess generalizability, the model trained on Alaskan data was validated using more recent data from Michigan, maintaining acceptable accuracy levels despite differing regional characteristics. Future work involves incorporating natural language processing (NLP) techniques to analyze crash narratives, aiming to capture contextual information often missed in structured datasets. This integrative approach seeks to support more effective traffic safety interventions and policy planning in cold-climate transportation systems.