Uncertainty-aware prediction of asphalt pavement roughness using probabilistic machine learning

Accurate prediction of asphalt pavement deterioration is critical for data-driven infrastructure asset management. However, conventional data-driven models commonly provide deterministic estimates, offering limited insight into prediction confidence. This limitation restricts their application in risk-aware decision-making, where understanding the reliability of each prediction is essential. This study addresses this challenge by developing an uncertainty-aware framework for predicting pavement roughness using data from the Long-Term Pavement Performance program. A curated dataset of 1,246 test sections across 50 states in the United States was compiled, covering diverse climate regions, structural configurations, traffic loading, and surface characteristics. Probabilistic modeling approaches, including Bayesian neural networks and Gaussian process regression, were used to predict pavement roughness while generating prediction intervals. Deterministic approaches, including random forest, extreme gradient boosting, support vector regression, and artificial neural networks, were also developed under the same validation and optimization framework for benchmarking. All models were evaluated using five-fold cross-validation and optimized using a Bayesian optimization framework. The results showed that extreme gradient boosting achieved the highest point prediction accuracy, while Gaussian process regression and Bayesian neural networks provided competitive accuracy, with coefficients of determination of 0.9133 and 0.9226, respectively, along with explicit uncertainty estimates. Gaussian process regression also demonstrated a reliable uncertainty calibration, with empirical coverage of 95% prediction intervals ranging from 88.89% to 96.31% across climate regions. Initial roughness, rutting, and cracking were identified as the dominant predictors. Regional analyses further revealed generally lower prediction errors in non-freezing regions, and spatial analyses identified geographic variations in pavement deterioration.