Semantic 3D reconstruction of tunnel lining defects from sparse GPR measurements

Three-dimensional (3D) characterisation of tunnel lining defects is essential for understanding failure mechanisms and informing maintenance decisions. Ground-penetrating radar (GPR) is widely used for tunnel lining inspection, but full-coverage measurements are often infeasible due to the curved tunnel geometry and obstructions from in-tunnel facilities. This paper proposes a systematic workflow, termed GPR-3DRE, for the semantic 3D reconstruction of tunnel lining defects from sparse GPR data. The workflow comprises two stages: (i) prediction of radargrams in unmeasured areas using a deep-learning-based model; (ii) construction of an attribute-enhanced 3D GPR volume, followed by segmentation and reconstruction of 3D defect models through an unsupervised machine-learning approach. The proposed workflow is validated through two field case studies. Comparative analyses demonstrate that each stage of GPR-3DRE surpasses existing state-of-the-art methods in accuracy and reliability, thereby providing an effective and practical solution for 3D reconstruction of tunnel lining defects from sparse GPR measurements.