Multi-Source Fusion 3D Voxel Deep Learning Based on 3D U-Net for Intelligent Karst Detection in Highway Engineering of Karst Regions

Abstract

Artificial intelligence and three-dimensional spatial intelligence perception technologies provide new computational paradigms for complex underground engineering surveys. To improve spatial accuracy and risk identification in highway engineering of karst regions, this study proposes an intelligent karst detection method based on multi-source three-dimensional voxel modelling and an improved 3D U-Net architecture. Drilling data, geophysical exploration signals, and geological imagery are unified into a voxelized geological representation, and multi-channel volumetric features are constructed to characterize karst spatial heterogeneity. The improved 3D U-Net is designed to learn karst morphological structures and anomaly response patterns directly from fused voxel blocks, enabling automatic and spatially consistent karst identification. A karst-developed highway section is used for validation. Quantitative results show that the proposed method achieves an intersection-over-union (IoU) of 0.73, significantly outperforming traditional single-source and non-voxel-based approaches. Furthermore, constraining borehole layout based on identification results effectively reduces misclassification and redundant surveys. The proposed framework demonstrates strong potential for enhancing spatial reliability and decision efficiency in detailed highway engineering surveys of karst regions.