Texture-based image analysis and explainable machine learning for polished asphalt identification in pavement condition monitoring

This study presents a framework for detecting polished asphalt pavement surfaces by integrating texture-based image analysis with interpretable Machine learning (ML). Polishing, caused by aggregate degradation and bitumen aging, alters surface texture and reduces skid resistance, posing a safety risk. A real-world dataset of 12,480 pavement images was analyzed using 24 texture features derived from the Gray Level Co-occurrence Matrix (GLCM), capturing directional spatial patterns of surface roughness. Several ML models were trained and optimized with the Hyperopt framework, with a Backpropagation Neural Network (BPNN) achieving the highest classification accuracy of 96.1%. Feature contributions were interpreted using SHapley Additive exPlanations (SHAP), providing physical insight into texture-driven polishing mechanisms. Although a ResNet50-based CNN achieved slightly higher accuracy (98.7%), its high computational cost limits practical deployment. The proposed GLCM-ML approach offers an interpretable, efficient, and physics-aware tool for pavement condition monitoring, with potential to enhance predictive modeling of surface texture evolution.