Prediction of Asphalt Pavement Skid Resistance Based on Fusion of Pavement Image Texture Features

To address the lack of explicit physical priors in pure vision deep learning models and poor generalization under limited labeled samples for asphalt pavement skid resistance prediction, this study builds a multi-source dataset with pavement RGB images, British Pendulum Number (BPN) and three-dimensional texture parameters (MPD, Ssk and Rv) collected from Cida Road in Chongqing, China. A gated multimodal fusion model is proposed, adopting 1024 × 1024 high-resolution input with a Sobel edge branch and a mesoscale downsampling branch to enhance multi-scale texture characterization. A two-stage transfer learning strategy combined with data augmentation is applied for small-sample generalization. The recommended v6 model achieves RMSE of 1.782 BPN, MAE of 1.354 BPN and R2 of 0.719 on the independent point-level test set, reducing RMSE and MAE by 15.8% and 18.3% respectively compared with the Ridge regression baseline and outperforming the MPD-only regression (R2 = 0.632). A matched no-image control (R2 = 0.251) confirms that the image pathway provides a large positive contribution once adequate physical priors steer the gated fusion, whereas images alone are not predictive. Feature analysis reveals a strong positive correlation between MPD and BPN, while GLCM features provide limited information gain. This method offers a preliminary reference for rapid image-based screening of pavement skid resistance. Image acquisition is non-contact, but the MPD physical prior is still obtained with a dedicated 3D texture scanner; the present results constitute a feasibility study on a single road, and fully non-contact deployment on independent road sections remains future work.