Explainable AI–Enhanced Pavement Deterioration Modeling Using Hybrid Classification–Regression Deep Learning Architecture

Effective Pavement Management System (PMS) planning depends on the ability to anticipate both the severity and physical extent of multiple distress types under real-world data constraints. This work introduces an interpretable dual-stage Deep Learning (DL) framework that sequentially classifies distress severity and then predicts the corresponding crack lengths, alligator cracking areas, and pothole quantities. The first stage uses an Artificial Neural Network (ANN) optimized with Focal loss to overcome severe class imbalance, while the second employs an ANN with Huber loss and targeted non-zero weighting to counteract outlier dominance and the zero-inflation inherent in sparse distress records. Data scarcity is addressed through the Synthetic Minority Over-sampling Technique (SMOTE) and Gaussian noise augmentation. To assess whether the learned predictive relationships are consistent with established Mechanistic-Empirical (M-E) pavement knowledge, an Explainable Artificial Intelligence (XAI) framework combining Shapley Additive Explanations (SHAP), Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) curves is integrated. These post hoc analyses do not establish causality; rather, they show that high-severity alligator cracking is the most influential predictive indicator associated with pothole area, a pattern consistent with established pavement deterioration mechanisms. Beyond prediction quality, the design is substantiated through systematic ablation experiments: a shared-backbone Multi-Task Learning (MTL) variant is shown to degrade regression accuracy due to gradient interference between conflicting loss objectives, while network capacity is empirically scaled, using a bottleneck regularization strategy for data-scarce targets such as pothole area and count, to prevent overfitting without compromising expressiveness. The framework achieved F1-scores of 0.843, 0.930, and 0.982 for linear crack, alligator crack, and pothole severity, respectively. In addition, R2 values of 0.941, 0.954, 0.823, and 0.885 were obtained for linear crack length, alligator crack area, pothole area, and pothole count, respectively. However, the pothole-severity result should be interpreted as a high point estimate obtained from a small and highly sparse sample rather than as evidence of near-perfect population-level classification performance. Paired statistical tests indicate no detectable systematic bias in the evaluated predictions. The proposed framework provides strong and stable performance for the linear and alligator cracking targets, whereas the pothole area and count results should be regarded as preliminary because of the extremely limited number of available pothole observations. Accordingly, the framework offers a promising and interpretable basis for pavement maintenance planning within the Long-Term Pavement Performance (LTPP) context, while further external validation is required before operational deployment for sparse poth