Developing an improved extreme gradient boosting model for predicting the international roughness index of rigid pavement

Abstract

The International Roughness Index (IRI) is a crucial indicator for evaluating pavement performance. However, existing prediction models for IRI in rigid pavements primarily rely on linear regression or simple machine learning techniques, which necessitates improvements in both training efficiency and interpretability. In addressing this issue, a proposed IRI prediction model combines the Extreme Gradient Boosting (XGBoost) method with advanced hyperparameter optimization utilizing the Tree-structured Parzen Estimator (TPE) technique. Furthermore, the model incorporates the Shapley Additive Explanations (SHAP) interpretability framework. To train and validate the model, 146 historical observation records were extracted from the Long-Term Pavement Performance database. The Boruta algorithm was employed to select the final input variables. Results demonstrated that after TPE hyperparameter optimization, the XGBoost model achieved an R2 value of 0.896, and an RMSE value of 0.187. Surpassing performance compared to other ensemble learning models. The Boruta algorithm successfully identified all significant variables, thereby improving the model's accuracy on the testing set. Additionally, the SHAP framework confirmed the primary factors influencing IRI and provided explanations for the model's results at both global and local levels. These findings highlight the ability of the proposed model to accurately predict IRI, providing precise guidance for pavement maintenance while reducing the workload associated with data collection and management for pavement engineers.