Optimizing pavement performance prediction with stacking regressor models

Abstract

Pavement performance assessment is crucial for ensuring road durability and safety, influencing maintenance and rehabilitation decisions. This study introduces a stacking regressor with CatBoost and XGBoost as base estimators and Linear Regression as the final estimator to assess pavement conditions such as surface curvature index (SCI), base curvature index (BCI), base damage index (BDI) etc. Inputs include cracking percentage, plasticity index, temperature, maximum dry density, CBR, soil type, and layer thickness. The dataset of 2001 samples from Indian roads were trained and validated. The stacking regressor outperformed classical methods and machine learning models like Random Forest, CatBoost, XGBoost, and LightGBM, achieving the highest R² scores and the lowest MSE and MAE. K-Fold Validation MSE for the Stacking Regressor was 0.0208, compared to 0.0224 for CatBoost, 0.0269 for XGBoost, and 0.0281 for Linear Regression. The R² scores for the best outputs, SCI and BCI, were 0.8055 and 0.7753, compared to 0.7624 and 0.7202 for XGBoost, and 0.7982 and 0.7702 for CatBoost, confirming the stacking regressor’s efficacy in predicting pavement conditions with the highest precision.