Performance prediction is a crucial step in the predictive maintenance of pavements, which could affect working plans and budget allocation of transport authorities. The International Roughness Index (IRI) is a widely adopted performance index that reflects the pavement condition at some level and the ride quality. Machine learning algorithms, which have the capacity to handle nonlinear problems, have been extensively used to predict pavement performance. However, such methods usually suffer from the issues of hyperparameter optimisation and the lack of interpretation ability. Pavement deterioration involves inherent uncertainty in material properties, measurement accuracy, and environmental and operational conditions. It is thus critical to characterise the predictive uncertainty so that rational maintenance decisions will be made. Nevertheless, most existing approaches focus on single-value prediction of pavement performance and fail to consider this inherent uncertainty. In this study, a reduced-order Gaussian Process Regression (GPR) is proposed for the probabilistic prediction of the IRI of flexible pavements. The Long-Term Pavement Performance (LTPP) dataset is used to illustrate its feasibility. These datasets are collected from the LTPP Specific Pavement Studies located in different climatic zones in the US. The proposed method is then compared with the traditional GPR. The comparison results indicate that the reduced-order GPR model exhibits higher predictive performance and better uncertainty quantification ability than the traditional one.