A SHAP-interpretable LightGBM model for reliable IRI prediction: bridging data-driven analytics and mechanistic insight in urban pavement management

Accurate prediction of the International Roughness Index (IRI) is essential for effective urban pavement management. Yet, traditional empirical models often fail to capture the complex nonlinear interactions governing pavement deterioration, particularly in data-scarce local networks. To address this limitation, this study develops a localized, data-driven framework for predicting the IRI of asphalt pavements in Tehran using a multi-source dataset comprising 2,313 road segments (200 m units; approximately 462.6 km). Six supervised machine-learning algorithms were evaluated, including CatBoost, XGBoost, LightGBM, Random Forest, K-Nearest Neighbors, and Decision Tree. Among them, LightGBM delivered the best predictive performance, achieving a coefficient of determination (R²) of 0.83 and a mean absolute percentage error (MAPE) of 4.235%. SHAP analysis was further employed to interpret model behavior and identify the most influential predictors. The results showed that air voids (Va) had the strongest effect on IRI prediction. By relying on locally available data and interpretable machine learning, the proposed framework provides an accurate and transparent decision-support tool for urban pavement management.