Hybrid MLP–LSTM deep predictive modeling of pavement distress severity and density and condition indices

Accurate prediction of pavement performance is essential for optimizing maintenance and rehabilitation strategies within Pavement Management Systems (PMS). However, traditional mechanistic–empirical models often fail to represent the nonlinear and time-dependent interactions among structural, environmental, and traffic factors. This study proposes an integrated predictive framework that combines Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks to jointly predict quantitative distress parameters (length, area, and count) and classify distress severity levels. The Long-Term Pavement Performance (LTPP) database was utilized to develop 13,080 records with 38 features related to pavement distress, traffic, climate, and structure. Preprocessing included normalization, missing value imputation, and Synthetic Minority Over-sampling Technique (SMOTE) to reduce class imbalance. Models were trained for 100 epochs using focal loss to enhance learning for rare distress categories. Results show that MLP achieved superior regression accuracy (R² ≈ 0.83–0.92), especially for linear parameters like crack length. Conversely, LSTM yielded higher classification accuracy (F1-score ≈ 0.86–0.91) due to its ability to capture temporal dependencies. The proposed framework also generated realistic deterioration trends for the Pavement Condition Index (PCI) and Surface Cracking Index (SCI) over a 20-year horizon (2013–2033), supporting integration within Cognitive Digital Twin (CDT) systems for proactive pavement management.