An explainable ensemble learning framework for flexible pavement roughness prediction under multi-climate stressors

The significant impact of climate change on pavement performance has been widely recognized. However, due to the complex interplay of multiple influencing factors, existing evaluation methods still lack high-precision, multi-dimensional predictive tools. To address this deficiency, this study proposes a novel machine learning–based framework for estimating the International Roughness Index (IRI) of pavements. A dataset comprising 1626 samples was constructed using the Long-Term Pavement Performance (LTPP) database, covering eight U.S. states and a broad range of climate zones. The dataset integrates road structure, traffic load, and key climatic variables such as temperature, precipitation, and humidity, ensuring strong regional representativeness. A comparative analysis was conducted across nine commonly used machine learning algorithms. After hyperparameter optimization and cross-validation, the Extreme Gradient Boosting (XGB) model demonstrated the highest predictive accuracy, achieving R² values of 0.9917 and 0.9930 on the training and testing sets, respectively. Furthermore, SHapley Additive exPlanations (SHAP) analysis identified precipitation, humidity, and freeze–thaw cycles as critical climatic drivers. Importantly, although some traffic-related variables showed strong correlations, they were deliberately retained due to their distinct engineering significance, providing a more comprehensive description of traffic-induced deterioration. By combining high predictive performance with explainable ML tools, the proposed framework not only ensures robustness to multicollinearity but also offers mechanistic insights into the progression of pavement roughness under diverse climatic conditions. To enhance practical applicability, a user-friendly graphical user interface (GUI) was developed, enabling rapid and accurate IRI prediction to support pavement maintenance and management decisions.