An Integrated explainable machine learning framework for pavement condition assessment and maintenance prioritization: A Saudi Arabian case study

Abstract

Abstract Effective pavement condition assessment is critical for maintaining durable, safe, and well-performing road networks, particularly in arid regions such as Saudi Arabia, where heavy traffic loading, high temperatures, and surface distress accelerate pavement deterioration. Traditional mechanistic or deterministic approaches often struggle to capture the nonlinear and complex patterns of pavement deterioration. This study aims to develop and validate an integrated, explainable machine-learning (XML) framework for predicting the International Roughness Index (IRI) using 4 years of traffic, pavement distress, and weather data from a 190 km-long rural highway in Al-Qassim Province, Saudi Arabia. IRI is classified into three distinct classes (low, medium, and high) based on predefined thresholds. The dataset included ~ 27,700 observations on cracking, rutting, texture, pavement serviceability, pavement condition rating, traffic volume, speed, temperature, humidity, wind speed, and precipitation. The Synthetic Minority Over-sampling Technique (SMOTE) trained the data and addressed data imbalance. Seven machine learning models, namely Decision Tree (DT), Support Vector Machine (SVM), Gradient Boosting (GB), Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Deep Neural Network (DNN), were evaluated using stratified k-fold cross-validation and an independent testing set. To enhance the model’s transparency and interpretability, feature importance and SHAP (Shapley Additive Explanations) analyses are also performed. The study also developed a pavement deterioration curve to quantify IRI progression over time. Model assessment across multiple performance indicators (accuracy, precision, recall, F-1 score) showed that XGBoost and CatBoost outperformed other models, achieving the highest classification accuracy (0.87). Feature importance estimates from CatBoost indicated traffic loading as the most influential predictor of IRI, followed by rutting and cracking. SHAP yielded an interpretable visualization of the collective influence of significant predictors to shape IRI outcomes. The pavement deterioration curve indicated IRI approaching a critical threshold of 4.0 in 6.6–6.7 years, under prevailing conditions without major rehabilitation. The proposed framework provides an efficient data-driven tool for identifying high-risk pavement sections, prioritizing maintenance interventions, and supporting proactive pavement asset management in arid and heavy-traffic environments.