This study developed a machine learning–based predictive framework to preemptively classify four major types of bridge deck pavement damage—cracking, potholes, alligator cracking, and pavement wear—and elucidated their physical deterioration mechanisms using SHAP (SHapley Additive exPlanations) analysis. A comprehensive dataset was constructed by integrating 12,972 pavement inspection records collected from 3,162 bridges in South Korea between 2004 and 2024 from the Bridge Management System (BMS), together with traffic data from the Traffic Monitoring System (TMS) and climatic variables from the Korea Meteorological Administration (KMA). Three machine learning algorithms—XGBoost, Random Forest, and Logistic Regression—were compared under three class imbalance strategies (No-Resampling, SMOTE, and ADASYN), with hyperparameters optimized using Optuna and model robustness validated through 10-fold cross-validation and a hold-out test dataset. XGBoost consistently outperformed the other algorithms across all damage types. For cracking and potholes, which exhibit distinct morphological boundaries, the original dataset (No-Resampling) yielded the highest F1-Scores of 0.8783 and 0.8227, respectively. For progressive damages with ambiguous boundaries—alligator cracking and pavement wear—SMOTE oversampling achieved the best performance, with F1-Scores of 0.8150 and 0.7644, respectively. All models achieved high precision (0.89–0.98), minimizing false-positive detections that could lead to unnecessary maintenance expenditures. SHAP analysis quantitatively identified the dominant deterioration drivers for each damage type, including thermal stress and heavy truck traffic (AADTT) for cracking, moisture infiltration and freeze–thaw cycles for potholes, cumulative fatigue under sustained loading for alligator cracking, and time-dependent UV aging combined with winter maintenance operations for pavement wear. These findings provide data-driven evidence to support the transition from reactive to preventive bridge infrastructure asset management.
Lee et al. (Mon,) studied this question.