Uncertainty-aware pavement performance prediction based on Gaussian process regression

Deterministic pavement performance models fail to capture uncertainties inherent in the deterioration process, limiting risk-informed maintenance decision-making. This study employs Gaussian process regression (GPR), a non-parametric Bayesian inference method, to predict pavement performance while explicitly quantifying prediction uncertainty. Separate GPR models were established for eight maintenance and rehabilitation (M&R) treatments and three performance indicators, using data from 10 expressways in southeastern China. Gradient Boosting Decision Trees (GBDT) were employed for feature selection, and model performance was evaluated using point and interval prediction metrics. Results show that the proposed GPR models achieved robust predictive accuracy, with coefficient of determination (R2) values exceeding 0.7 for all models and 70.83% exceeding 0.8 on testing sets. Prediction interval coverage probability (PICP) values approached the 95% confidence level with narrow intervals, confirming reliable uncertainty quantification. Out-of-sample validation on three additional expressways further confirms model generalisability. Sensitivity analysis revealed that both predictive accuracy and uncertainty quantification were largely insensitive to kernel selection. Compared with Bayesian Neural Network (BNN), random forest and Extreme Gradient Boosting (XGBoost), GPR demonstrated more consistent prediction performance and better-calibrated uncertainty estimates. The findings highlight the effectiveness of GPR in providing uncertainty-aware predictions, offering a reliable tool for pavement maintenance management.