Mesoscale investigation of salt migration and accumulation in asphalt mixtures using a deep-learning-based image segmentation framework

Salt migration and accumulation in asphalt mixtures, which critically affect long-term pavement durability, remain poorly understood. To address this gap, this study developed an integrated image-analysis framework that combines Otsu thresholding method (OTSU) and a nested U-shaped convolutional neural network (U-Net++) for semantic segmentation to characterize three-dimensional salt distribution from X-ray computed tomography (CT) images. Using this artificial-intelligence-assisted framework, the spatiotemporal evolution of salt was investigated in asphalt mixtures with different aggregate gradations, exposure modes, and exposure durations. The results showed that the proposed framework reduced the salt identification error from 7.34% to 2.12% compared with conventional single-threshold segmentation, enabling more reliable quantification of internal salt distribution. Salt migrated from surface-connected voids into the specimen interior and evolved from isolated clusters into an interconnected transport network. Continuous exposure caused more sustained salt ingress and more pronounced spatiotemporal evolution than dry–wet cycling. With increasing exposure duration, salt occupancy increased and migration pathways expanded, indicating enhanced salt accumulation, more efficient salt transport, and the development of a complex interconnected transport network. Among the investigated mixtures, the open-graded friction course (OGFC) exhibited the most extensive salt penetration and pathway development, followed by asphalt concrete (AC), whereas the stone mastic asphalt (SMA) showed the greatest resistance to salt transport. This study provides a quantitative framework for elucidating dynamic salt migration in asphalt mixtures and demonstrates the potential of deep-learning-based image segmentation for durability assessment and mixture design in saline environments.