Enhancing Municipal Pavement Condition Predictions Using Machine Learning Models

Pavement condition prediction is essential for effective street maintenance. Reliable prediction models enable municipalities to anticipate pavement deterioration and plan maintenance proactively. This paper examines the use of machine learning (ML) models to enhance prediction accuracy for municipal streets. Seventy-two models were developed to predict the 2022 pavement condition index (PCI) using Linear Regression (LR), Random Forest (RF) and Neural Network (NN) algorithms. The study data was collected through windshield surveys in 2014, 2018 and 2022 in Skellefteå Municipality, Sweden. The models were trained and tested separately on the 2018 dataset and the combined 2014 and 2018 datasets, incorporating several combinations of variable sets for residential and non-residential (main, collector, and industrial) streets. Additional models were developed for non-residential streets categorized by maintenance treatments to analyse the predictive capability of the models. RF consistently outperformed other models, with the RF (A+D+S) model achieved the highest accuracy for residential streets, with a marginal improvement using the combined dataset. The RF (A+D+S+T) model performed best for non-residential streets and was also the most robust for pavement segments with surface levelling and milling & resurfacing. Among the variables, WDV (Weighted Distress), Status2018 (the PCI rating based on the 2018 assessment), LAR (longitudinal/transverse cracking), A (pavement age), LT (light traffic), and SU (surface unevenness) were the most significant contributors in the best-performing models. These findings provide municipalities with practical guidance for data-driven maintenance, helping to prioritize interventions and optimize resources. The insights also support informed decisions on maintenance frequency and treatment selection to extend pavement service life.