A case study on predicting asphalt pavement distress using advanced machine learning techniques and road surface inspection data
Abolfazl Yazdi, Mohammad Hosein Dehnad
Abstract
This study develops robust machine learning models to predict pavement distress using high-resolution road surface scan data. The dataset includes 4,685 samples representing various pavement conditions across different locations. Six key pavement indicators were modeled: Total Rut (mm), Aggregate Stripping, Linear Crack Length, Linear Crack Area, Average International Roughness Index (AVG IRI), and Average Macrotexture Depth (AVG MTD). The modeling pipeline involved data cleaning, standardization, recursive feature elimination (RFE), and hyperparameter tuning via Optuna. Seven regression models— RF, Gradient Boosting, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), XGBoost, LightGBM, and Linear Regression—were trained and evaluated using MAE, MSE, RMSE, and R² metrics with 5-fold cross-validation and a test set. Results showed ensemble methods, especially Random Forest and XGBoost, outperformed others by effectively capturing complex nonlinear patterns. Random Forest achieved the best accuracy with R² scores of 0.824 for crack area, 0.810 for crack length, and 0.774 for aggregate stripping. SVR and Linear Regression showed weaker performance with lower R² and higher errors. Gradient Boosting also performed well for rutting and texture indicators. The study highlights the importance of feature selection and model tuning to enhance predictive accuracy. The proposed ML framework provides a scalable, reliable approach for pavement distress prediction, supporting improved pavement management strategies and future research.