Answers are grounded in this paper and cite their sources.
Three-dimensional ground-penetrating radar (3D GPR), as a high-efficiency, dataintensive, and wide-coverage nondestructive testing technique, is considered to be the optimal choice for underground defect detection. Nevertheless, the efficiency of 3D GPR data interpretation severely lags behind data acquisition. For interpretation methods based on individual two-dimensional (2D) slices, incorporating spatial information across the channels of 3D GPR data remains an important methodological consideration. Therefore, we propose a deep-learning framework based on point clouds to directly capture the scattering features of underground targets in 3D space for automatic interpretation. First, the 3D GPR data are converted into a 3D point cloud enriched with amplitude attributes. To mitigate the scarcity of field training data, CycleGAN is employed to simulate background clutter variations relevant to varying geologic conditions. Subsequently, PointNet++ classifies the prepared 3D GPR volumes into four categories: soil subgrade, pipeline, cavity, and manhole cover. Results demonstrate that the proposed framework effectively classifies road subsurface targets, and CycleGAN augmentation improves overall classification performance, with increased recall for cavities and pipelines. The proposed method combines spatial coordinates and intensity features within a sparse point-cloud representation for 3D GPR target classification, showing potential for the automated interpretation of large-scale GPR data sets in complex urban environments.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: