Predicting Pavement Condition Index Using an ML Approach for a Municipal Street Network

Abstract

Machine learning (ML) models are increasingly getting attention in predicting pavement maintenance methods to improve decision-making. This study investigates the use of ML at the municipal level to predict the street pavement condition index (PCI) rating over a 4-year span. Several supervised learning models, namely linear regression (LR), random forest (RF), and neural network (NN), were applied to the visually assessed pavement condition data of Skellefteå municipality, Sweden. Pavement distress, pavement age, and traffic data were used in several combinations to evaluate and compare the performance of the models. The RF model was based on paired variables of pavement age and pavement distress data. The results were comparatively accurate with R2=0.59 and Spearman’s coefficient=0.74 for residential streets in the model testing stage. Similarly, for main, collector, and industrial (MCI) streets, the RF model, based on pavement age and traffic variables, performed best with R2=0.79 and Spearman’s coefficient=0.88 during the model testing stage. The importance of input variables varies with the level of the model’s sophistication and pavement performance goal; however, pavement age is the dominant variable. The prediction models can be useful in effectively managing street networks among municipalities, even those with scarce resources.