Influence Analysis of Pavement Distress on International Roughness Index Using Machine Learning

Abstract

The International Roughness Index (IRI) is closely related to pavement distress. However, previous studies employing statistics and machine learning approaches would present challenges in comprehensively analyzing the influence of pavement distress on IRI considering their severities. This study introduces interpretable machine learning to investigate the influence of pavement distress on IRI, with a particular focus on the severity of pavement distress. The pavement distress and IRI data for flexible pavements obtained from the long-term pavement performance (LTPP) program were meticulously preprocessed. The developed random forest (RF) model demonstrated satisfactory predictive performance, with an RMSE of 0.2191 and an R² of 0.7874. The relationship between pavement distress and IRI, as captured by the developed model, was further analyzed using the SHapley Additive exPlanations (SHAP) method. The model interpretation identified the transverse crack, rutting, and alligator crack as the key factors influencing IRI. Notably, both transverse and alligator cracks exhibited significant contributions to IRI increment at medium and high severity levels, highlighting the importance of proactive maintenance for these distress types at lower severity levels. Additionally, a threshold in rutting depth was observed, which could increase IRI. A comparative analysis with the AASHTO MEPDG smoothness model demonstrated that the predictive performance of the RF model was notably superior.