Abstract
Winter road maintenance is a critical component of transportation safety and mobility in cold regions, where snow and ice events can greatly reduce pavement friction, roadway capacity, and operational efficiency. As climate variability increases the frequency and unpredictability of winter storms, the need for data-driven, adaptive, and autonomous decision-support systems has grown substantially. This dissertation advances an integrated artificial-intelligence-enhanced framework designed to improve situational awareness, forecasting accuracy, and maintenance decision-making for winter road operations. The research brings together physics-based modeling, machine learning, and reinforcement learning to address long-standing limitations in road condition monitoring and treatment optimization. Five core components are developed and investigated. First, a cloud-based Smart Maintenance Decision Support System (SmartMDSS) is designed to unify real-time data acquisition, road image recognition, and rule-based maintenance recommendations, providing the foundational infrastructure for data-driven operations. Second, a comprehensive physics-based road surface model is formulated to capture the coupled interactions of snow, ice, water, salt, heat transfer, and maintenance actions. This model establishes a virtual environment capable of simulating realistic storm evolution and treatment effects under diverse meteorological and operational scenarios. Third, short-term road surface temperature forecasting is examined using three machine learning algorithms, Transformer, LSTM, and XGBoost to evaluate their predictive accuracy and operational suitability. Fourth, a Deep Q-Network reinforcement learning framework is developed to autonomously identify optimal plowing and salting strategies based on the evolving virtual environment, demonstrating the potential of DRL to enhance maintenance efficiency and resource allocation. Fifth, Reinforcement Learning from Human Feedback (RLHF) is incorporated to integrate operator expertise into the decision-making process, enabling personalized agent behavior that reflects local practices and improves trust, interpretability, and adaptability. The methodologies and findings presented in this dissertation demonstrate a unified approach where physics-based analysis, data-driven prediction, and human-in-the-loop reinforcement learning collectively support more resilient and adaptive winter maintenance strategies. The contributions offer an important step toward autonomous, scalable, and operator-informed decision-support tools that can help transportation agencies improve safety, reduce costs, and respond more effectively to winter storms.
