High-Frequency GPR Data Reconstruction With Conditional GAN and Contrastive Unpaired Translation

High-resolution ground-penetrating radar (GPR) data acquired using the high-frequency antenna enables operators to identify closely spaced targets or slight variations within the subsurface, but with limited penetration depth due to the significant attenuation of electromagnetic waves. To mitigate this inherent trade-off, we introduce a conditional generative adversarial network (cGAN) with the contrastive unpaired translation (CUT) framework, for reconstructing high-frequency, in other words, high-resolution, GPR images from low-frequency inputs. We use an enhanced U-Net with attention mechanism as generator to improve global context modeling. Additionally, patch-level contrastive learning is integrated to ensure structural consistency between the low-frequency inputs and high-frequency outputs. After training the network using the shallow parts of a real dual-frequency dataset, we directly reconstruct the high-frequency GPR data from the low-frequency measurements. Results indicate that our proposed method surpasses the CycleGAN-based resolution enhancement method in effectively producing structurally coherent and detailed high-resolution images for both shallow and deep regions. It also demonstrates the generalizability of the proposed method, as the network does not see the data related to deep regions during the training stage. Therefore, our method offers a promising way to reconstruct high-frequency GPR data from low-frequency measurements, significantly enhancing the interpretability of deep subsurface GPR imaging outcomes.