Comprehensive evaluation of multiple machine learning classifiers for predicting rutting severity levels

The accurate prediction of rutting severity levels is vital for the maintenance and safety of flexible pavements, facilitating timely, cost-effective measures to avert further degradation and prolong road life. In this study, we explore the efficacy of various machine learning algorithms in classifying rutting severity, drawing on a substantial dataset from the Long-Term Pavement Performance (LTPP) program. Through the use of feature selection techniques such as MRMR, Chi-Square, ANOVA, and Kruskal Wallis, we assessed the accuracy of decision trees, Naive Bayes, SVMs, ensemble trees, and neural networks. The Quadratic SVM with MRMR-selected features achieved the highest accuracy of 95.51%, followed closely by the Bagged Trees model with 95.45% using ANOVA-selected features. Conversely, the RUSBoosted Trees model, when paired with Kruskal Wallis features, exhibited the lowest accuracy at 68.73%. Normalized confusion matrices elucidated how different feature selection methods influenced the models’ classification accuracy. These insights stress the critical role of feature selection in the performance of machine learning algorithms and establish a foundation for deploying effective predictive modeling in pavement management. Our findings advance the methodology for implementing enhanced maintenance strategies, ultimately contributing to more durable and safer roadway systems.