Selecting and optimizing distress feature descriptors from multi-source images for pavement distress classification

Accurate and efficient early detection of pavement distress is crucial, as it prevents further deterioration, reduces repair costs, and enhances driving safety by mitigating hazards like potholes and cracks. However, automatically detecting pavement distress from image samples remains challenging due to complex asphalt textures, including variations in aggregate size, surface wear, and non-distress-related artifacts like shadows and tire marks. This study addresses these challenges by screening and optimizing texture descriptors to characterize local structures and patterns of image regions containing pavement distress. An automated detection system is developed using four categories of texture descriptors: (1) original image-based parameters, (2) multi-resolution analysis-based texture features, (3) local binary pattern (LBP) and their variants-based texture features, and (4) gray-level co-occurrence matrix (GLCM)-based texture features. Unlike previous studies that rely on a single type of descriptor, this integrated approach processes both 2D intensity images and 3D pavement surface models, capturing global and local texture patterns to improve robustness and feature discriminability for distress classification. A total of 196 optimal feature descriptors are selected to represent images from both imaging modalities, forming the basis of a distress classification model. A dataset comprising 4,932 samples across 8 distress categories, with slight variations in class distribution, is constructed to compare the performance of Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), and Support Vector Machine (SVM) in distress classification. Class-weighted training and balanced performance metrics were used to mitigate class imbalance effects and ensure robust model evaluation. Among these, the SVM classifier demonstrates the best predictive performance and robustness. Extensive experiments, including feature combination analysis and ablation studies, identify the optimal feature set comprising original image-based, multi-resolution analysis-based, and LBP variants-based descriptors. This feature set achieves a classification accuracy rate (CAR) of 0.896, an F1-score of 0.894, and area under receiver operating characteristic curve (AUC) of 0.938 across all distress categories. The selected descriptors effectively differentiate pavement distress types, providing a reliable automated classification system.