Answers are grounded in this paper and cite their sources.
Ground Penetrating Radar (GPR) is a widely used nondestructive testing method for tunnel lining inspection, but its effectiveness is restricted by the trade-off between resolution and penetration across different frequency bands. Conventional fusion methods for dual-frequency GPR data rely on manual parameter tuning and heuristic strategies, often producing artifacts and discontinuities where the two datasets are merged, including at transitions between high- and low-frequency-dominant regions. Deep learning (DL) approaches can mitigate these issues but typically require extensive, diverse training datasets that are rarely available in tunnel inspections. To address these challenges, this study proposes DFNet, a DL framework for dual-frequency GPR fusion. The network partitions complex Fourier coefficients into high- and low-amplitude branches with a prescribed amplitude threshold, applies multi-scale convolutional encoding with attention mechanisms, and incorporates a Transformer encoder to capture both local and global dependencies. Fused time-domain representations are adaptively weighted through an information-aware loss function that integrates entropy and statistical frequency constraints, thereby preserving fine structural details and ensuring smooth transitions across tunnel layers. A numerical simulation and a controlled field test show that DFNet outperforms representative state-of-the-art fusion methods, delivering enhanced accuracy and robustness in subsurface characterization.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: