Abstract: Establishing an accurate numerical model of the mesostructure of soil-rock mixtures using CT images is essential for studying their physical and mechanical properties. However,conventional image processing techniques for reconstructing the mesostructure often require manual intervention,which results in low efficiency and insufficient accuracy. To address these issues,this study proposes an image segmentation method for clayey soil-rock mixtures based on an enhanced UNet neural network,and a boundary-weighted loss function is designed to accurately segment CT scan images of typical clayey soil-rock mixture specimens. Additionally,a novel method for 3D reconstruction of soil-rock mixtures is introduced to transform the segmented images into a detailed 3D numerical model. The results show that the deep learning-based method significantly improves the efficiency and accuracy of CT image segmentation of clayey soil-rock mixtures compared to traditional techniques. The integration of the UNet model with the boundary-weighted loss function enhances segmentation accuracy,particularly at the target boundary. Meanwhile,the higher accuracy of the segmented image makes the reconstructed 3D model more detailed and accurate. This study ensures that numerical simulation tests,based on this foundation,minimize human interference and maintain high accuracy.