Comparative analysis of machine learning models for multi-depth pavement temperature prediction

In cold regions, temperature fluctuations critically affect the stability of granular base course and subgrade layers. To improve pavement design and maintenance, an instrumented test section was built in Edmonton, Alberta, with thermistors placed at depths of 0.25, 0.70, and 2.25 m. From May 2023 to December 2024, environmental data—including solar radiation, air temperature, and day of year—were collected. Three machine learning algorithms, Categorical Boosting, Extreme Gradient Boosting, and k-nearest neighbors, were tested for pavement temperature prediction. Models developed for individual depths consistently outperformed a single global model, with R 2 values above 0.9. Shapley additive explanations confirmed both robustness and interpretability, supporting their reliability for temperature prediction. These findings highlight the potential of depth-specific models to improve profiling in cold climates, enhancing pavement monitoring and enabling proactive maintenance strategies.