Work

Hybrid Dense-Sparse Networks for 3D Object Segmentation in GPR-based Subsurface Imaging

Year: 2026
Type: conference-paper
Abstract: This paper presents an end-to-end framework for 3D multiclass semantic segmentation of subsurface objects in reconstructed ground-penetrating radar (GPR) volumes. We first develop a reconstruction method that recovers 3D reflectivity volumes directly from raw measurements using a proximal gradient solver. We then propose a hybrid dense-sparse segmentation framework that exploits the spatial sparsity of subsurface scenes. Specifically, a dense three-dimensional residual U-Net architecture produces a coarse full-volume segmentation, and a sparse refinement network updates predictions inside candidate object regions. The refiner processes only the ~1.4% of voxels flagged as candidate foreground, matching the natural sparsity of the scene. The framework is evaluated on reconstructed volumes containing three metallic target classes (a flat plate, a cylinder, and a rectangular box) with varying positions, sizes, depths, and orientations. With test-time augmentation, the hybrid method achieves accurate segmentation across all three classes, reaching a mean foreground Dice of 0.9650 (plate 0.989, cylinder 0.947, box 0.960) and detecting every target type in every test scene.
Language: English
Subfield: Geophysics
Open Access status: hybrid
OpenAlex ID: W7204149808
FWCI:
Cites: 18
Cited by:
Related to: 10