This study proposes an explainable extreme gradient boosting (XGBoost) model for predicting the international roughness index (IRI) and identifying the key influencing factors. A comprehensive dataset integrating multiple data sources, such as structure, climate and traffic load, is constructed. A voting-based feature selection strategy is adopted to identify the key influencing factors, which are used as inputs for the prediction model. Multiple machine learning (ML) models are trained to predict the IRI with the constructed dataset, and the XGBoost model performs the best with the coefficient of determination ($R^2$) reaching 0.778. Finally, interpretable techniques including feature importance, Shapley additive explanations (SHAP) and partial dependency plots (PDPs) are employed to reveal the mechanism of influencing factors on IRI. The results demonstrate that climate conditions and traffic load play a critical role in the deterioration of IRI. This study provides a relatively universal perspective for IRI prediction and key factor identification, and the outputs of the proposed method contribute to making scientific maintenance strategies of roads to some extent.

Balancing detection performance and computational efficiency is critical for sustainable pavement disease detection in energy-constrained scenarios. However, existing visual methods often struggle to adapt to structural transformations and capture critical features of pavement diseases in complex environments, while their computational demands can be resource-intensive. To address these challenges, this paper proposes a structure perception and feature attention network (SPFAN). The network includes a structure perception module that employs the updated deformable convolution technique. This technique enables the model to dynamically adjust and focus on the actual pavement disease regions, improving the accuracy of feature extraction, especially for diseases with irregular shapes and sizes. Additionally, the convolutional block attention module (CBAM) is integrated to optimize feature map attention across channel and spatial dimensions, enhancing the model focus on critical disease features without significantly increasing complexity. To further improve robustness, the generalized intersection over union (GIoU) loss function is adopted, ensuring better stability across targets of varying shapes and sizes. Experimental results on real-world pavement disease images show that the mAP@0.5 of the proposed SPFAN increases from 66.2% to 71.2%, an improvement of 7.55%, while the F1-score also increases by 9.03%, compared to the baseline YOLOv8n model. Furthermore, while achieving significant accuracy improvements, the proposed method maintains a similar parameter count as the baseline, preserving its low computational demands and high efficiency, making it suitable for real-time pavement damage detection in energy-constrained environments.