The use of deep learning to interpret ground penetrating radar (GPR) data for automated target detection has become a rapidly developing field. However, the performance of deep learning networks is significantly affected by the quantity and diversity of the available GPR data for training. The finite-difference time-domain (FDTD) method is commonly used to supplement datasets. However, there are considerable differences between FDTD-simulated and real GPR data, which limit the effectiveness of using FDTD data for training. To address these challenges, we propose a generative adversarial network (GAN)-based domain adaptation network for detecting target reflections in GPR data. This network combines the You Only Look Once (YOLO) network with a domain adaptation (DA) module, forming a GAN-based DA-YOLO that aligns features extracted from simulated and real GPR data to enhance the performance of the network on real GPR data. To further improve the network’s performance on real GPR data, a pseudo-labeling strategy is employed. This approach leverages the pseudo-labels generated by the network to iteratively refine and improve its accuracy. To validate the proposed method, we conduct comparative experiments to evaluate the impact of key factors, including network architecture, pseudo-labeling thresholds, and source-target domain discrepancies, on feature alignment and the improvement of detection performance for target domain data. The effectiveness of the proposed method is validated using real-world GPR data collected from a railway tunnel inspection. The results demonstrate that the proposed DA-YOLO network achieves the best detection performance on real data, with recall, precision, F1 score, and mAP values of 81.25%, 86.67%, 83.87%, and 69.30%, respectively, outperforming other networks that do not utilize domain adaptation methods. This study provides a solution for situations where real GPR data is scarce and only FDTD-simulated data can be used as training data.

The crosshole ground-penetrating radar (GPR) technique is widely used to characterize subsurface structures, yet the interpretation of crosshole GPR data involves solving non-linear and ill-posed inverse problems. In this work, we developed a generative adversarial network (GAN)-based inversion framework to translate crosshole GPR images to their corresponding 2D defect reconstruction images automatically. This approach uses fully connected layers to extract global features from crosshole GPR images and employs a series of cascaded U-Net structures to produce high-resolution defect reconstruction results. The feasibility of the proposed framework was demonstrated on a synthetic crosshole GPR dataset created with the finite-difference time-domain (FDTD) method and real-world data from a field experiment. Our inversion network obtained recognition accuracy of 91.36%, structural similarity index measure (SSIM) of 0.93, and RAscore of 91.77 on the test dataset. Furthermore, comparisons with ray-based tomography and full-waveform inversion (FWI) suggest that the proposed method provides a good balance between inversion accuracy and efficiency and has the best generalization when inverting actual measured crosshole GPR data.

Crosshole ground penetrating radar (GPR) is an efficient method for ensuring the quality of retaining structures without the need for excavation. However, interpreting crosshole GPR data is time-consuming and prone to inaccuracies. To address this challenge, we proposed a novel three-dimensional (3D) reconstruction method based on a generative adversarial network (GAN) to recover 3D permittivity distributions from crosshole GPR images. The established framework, named CGPR2VOX, integrates a fully connected layer, a residual network, and a specialized 3D decoder in the generator to effectively translate crosshole GPR data into 3D permittivity voxels. The discriminator was designed to enhance the generator’s performance by ensuring the physical plausibility and accuracy of the reconstructed models. This adversarial training mechanism enables the network to learn non-linear relationships between crosshole GPR data and subsurface permittivity distributions. CGPR2VOX was trained using a dataset generated through finite-difference time-domain (FDTD) simulations, achieving precision, recall and F1-score of 91.43%, 96.97% and 94.12%, respectively. Model experiments validate that the relative errors of the estimated positions of the defects were 1.67%, 1.65%, and 1.30% in the X-, Y-, and Z-direction, respectively. Meanwhile, the method exhibits noteworthy generalization capabilities under complex conditions, including condition variations, heterogeneous materials and electromagnetic noise, highlighting its reliability and effectiveness for practical quality assurance of retaining structures.