Estimating Winter Road Friction Coefficients: An Integrated Approach using Machine Learning, Explainable AI, and Model Transferability

Abstract

Ensuring road safety and efficient traffic movement through Winter Road Maintenance (WRM) operations is a pressing concern, particularly during harsh weather conditions. The challenge of accurately monitoring road friction coefficients, which play a crucial role in WRM, often leads to impractical and expensive solutions. To tackle such challenges, we introduce in this thesis a two-phase methodological framework that focuses on the development and optimization of machine learning (ML) models for road friction coefficient estimations, thereby bridging the gap between theoretical research and real-world application. In the first phase, we concentrate on the precise estimation of road friction coefficients. Utilizing meteorological and geographic data from the Road Weather Information Systems (RWIS), we developed a Regression Tree model that achieved a high accuracy of 93.3%. To ensure spatially continuous friction estimations, we employed Ordinary Kriging interpolation to handle missing weather data. By categorizing road friction coefficients into distinct risk levels, we were able to provide critical insights on road surface conditions, achieving nearly 90% accuracy. The second phase of our research emphasizes the refinement and augmentation of our models. We conducted a comparative analysis of ML algorithms, including Regression Tree, Support Vector Regression (SVR), Random Forest, and Extreme Gradient Boosting (XGBoost). We found a positive correlation between complexity and accuracy with XGBoost emerging as the most reliable model. To provide deeper insights into these models' inner workings, we leveraged SHAP explainable artificial intelligence (AI). We also examined the transferability of the XGBoost model, and after calibrating it for a new dataset, the updated model exhibited a significant improvement in accuracy on the new data, thereby affirming its robustness, adaptability, and transferability. The framework presented in this thesis goes beyond theoretical modeling to offer tangible and innovative solutions that can be readily applied in the field of WRM. By utilizing meteorological and geographic data in tandem with advanced ML models, this research creates a pathway for more effective estimation and interpretation of road friction. The synergy of explainable AI, accurate estimation methods, and the successful transfer of models to new data sets demonstrates the potential of modern technology to transform traditional practices, ultimately for improved mobility and safety of winter travelling public.