Optimization and Interpretability Analysis of Different IRI Prediction Models Based on LTPP Data

Abstract This study focuses on the construction, optimization, and interpretability analysis of different International Roughness Index (IRI) prediction models based on the Long-Term Pavement Performance (LTPP) database. A multidimensional feature system was established, and 15 key variables related to pavement structures, traffic loading, and climatic factors were identified using the Boruta algorithm in combination with Gini and permutation importance measures. Several machine learning models, including random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost), were developed and compared. To enhance predictive performance, Bayesian optimization (BO) was employed to fine-tune hyperparameters, and the BO-XGBoost model achieved the best results on the testing set [ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="normal upper R squared equals 0.9549" display="inline" overflow="scroll"> <mml:msup> <mml:mrow> <mml:mi mathvariant="normal">R</mml:mi> </mml:mrow> <mml:mrow> <mml:mn>2</mml:mn> </mml:mrow> </mml:msup> <mml:mo>=</mml:mo> <mml:mn>0.9549</mml:mn> </mml:math> , root mean square error (RMSE) = 0.0335, mean absolute percentage error (MAPE) = 4.77%], while maintaining high computational efficiency. Comparative experiments confirmed that BO-XGBoost outperformed models optimized with differential evolution (DE) and grid search (GS), as well as other candidate models, in terms of both accuracy and stability. Furthermore, Shapley additive explanations (SHAP)-based interpretability analysis revealed that surface layer thickness, traffic loading, and average humidity were the most influential factors affecting IRI.