Data-Driven Identification of Pavement Distress Factors in Imbalanced Datasets Using a Machine Learning Approach

Data-driven pavement maintenance systems are advancing globally as agencies increasingly rely on evidence-based decision-making. A critical component of effective maintenance planning is identifying the actual factors causing each distress type, so that appropriate treatment selection can be supported. However, pavement datasets often suffer from class imbalance, which reduces the reliability of distress-cause analysis. This study presents an analytical framework that addresses two key challenges in pavement engineering: handling class imbalances in distress datasets and identifying localized factors associated with major asphalt pavement distresses. Pavement condition data from 241 km of provincial highways were analyzed across nine distress types. Class imbalances were addressed using Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN), with data validity assessed through PCA, t-SNE, and Chi-square tests. The best-performing models were selected based on minority-class recall, F-1 score, and AUC-ROC. Distress occurrence was found to be influenced by traffic load, drainage, precipitation, land-use patterns, and terrain, offering insights that enable more targeted and context-specific maintenance strategies.