DNN-based inversion of Ground Penetrating Radar (GPR) data is gaining significant attention, primarily using simulated data for training due to the lack of ground truth in real-world scenarios. However, models trained on simulation data often perform poorly on real-world data due to domain discrepancies. This study applies an Unsupervised Domain Adaptation (UDA) method to GPR inversion, introducing a domain classifier within the main inversion DNN. By training the main DNN adversarially against the classifier to learn domain-invariant features, the DNN’s performance can be improved even without labeled data in the target dataset. Several simulation datasets with varying characteristics were generated to compare the performance of three domain classifiers. The experiments revealed that the proposed simple domain classifier structure, using 3×3 convolutional local alignment, outperformed both pointwise local alignment and global alignment classifiers. The proposed method achieved more than 80% improvement in metrics (MAE, SSIM, mIoU) over the non-UDA case, highlighting its effectiveness in addressing domain gaps related to frequency and attenuation. Furthermore, the DNN was trained using unlabeled data from real-world roads measured by on-vehicle GPR, and then applied to a bridge deck slab specimen with known design drawings. From both qualitative and quantitative perspectives, the proposed method demonstrates superior performance in permittivity and material estimation. The method enabled clear detection of rebars, and damage such as deterioration and cracks in the concrete slab. This research is expected provide new insights into the efficient subsurface damage detection of road infrastructures.

Detecting underground pipes and voids is vital for infrastructure safety. Using supervised learning for Ground-Penetration Radar (GPR) volumetric images faces difficulties when with limited unlabeled data. This paper introduces an unsupervised learning method for detecting underground objects from a GPR volumetric image without training dataset. The method employs singular value decomposition (SVD) for image enhancement, then merges textures and intensities using Gabor and Gaussian filters. A 3D-CNN model refined by graph-based supervoxels learning from the single 3D image, segments this image into a voxelwise label map. Following noise removal and binarization, a 3D binary underground-object map labels all voxels. Tested on a experimental field, the method outperformed previous method and identified 39 of 45 objects with 86.1% precision and a maximum IoU of 57.5%, emphasizing the efficacy of unsupervised learning in under-ground object segmentation.

Interpolation and high-resolution of missing traces of ground penetrating radar (GPR) images is a critical area of study. We propose a frequency-wavenumber (f-k) domain convolutional neural network specifically designed for the high-resolution task of GPR data acquired at large sampling intervals. The inputs and outputs of this network are the f-k domain spectra of GPR images. The architecture of the network is underpinned by aliasing theory, and it strategically addresses the relationship between the input and output spectra in the frequency domain. The proposed network was trained using 2000 simulation data generated via the FDTD method, and successfully reconstructed the f-k spectrum of data with fine 1 cm intervals from the spectra of coarsely sampled images taken at 7 cm intervals. In comparative performance evaluations, our network demonstrated a comparable proficiency to existing time-space domain networks for GPR data of earthwork sections. However, it exhibited superior accuracy in reconstructing images of bridge sections, where periodicity arises due to rebar responses. This indicates its enhanced capability in handling scenarios with periodic structural elements.

Vehicle-mounted multichannel ground penetrating radar (MC-GPR) is a revolutionary technology that facilitates the acquisition of volume images by arranging multiple antennas; however, its images are highly affected by noise due to different antenna characteristics. This study proposes reflectivity-consistent sparse blind deconvolution (RC-SBD) for appropriate denoising of ground penetrating radar (GPR) volume images. RC-SBD interprets the observed waveform as the convolution of the emitted wavelets and reflectivity, plus stationary clutter such as reflections from the vehicle itself. The method obtains denoised reflectivity by estimating the wavelets and clutter. The key feature of RC-SBD is that it extends the existing SBD method to 3-D, and introduces an assumption of reflectivity smoothness in the horizontal direction, expressed by the total variation (TV) regularization term. The estimation is formulated as a minimization problem involving$\ell _{2}$and$\ell _{1}$norms and is optimized using the Split–Bregman algorithm. Trade-off hyperparameters of the objective function are optimized via Bayesian optimization, maximizing the kurtosis of the calibrated volume image. Validation with synthetic data demonstrates accurate wavelet estimation and significant denoising of the volume image. Real-world data application further reveals considerable improvements in the channel-depth cross section, providing a clear visualization of structures like rebar and steel plates. Notably, the calibrated image remains stable across diverse datasets, including earthwork and bridge sections, showcasing the versatility and reliability of the proposed methodology.

Ground Penetrating Radar (GPR) is a promising tool for subsurface utility pipe detection due to its dense and highspeed 3D monitoring. However, because of enormous amount of radar data and difficulty of interpretation, inspection time and cost are the bottlenecks. In this research, a novel detection algorithm by the combination of 3D Convolutional Neural Network (3D-CNN) and Kirchhoff migration was proposed. The developed model estimated pipes' existences and directions. 3D-CNN utilized the 3D geometries of the pipes to achieve high classification accuracy compared to 2D-CNN. Kirchhoff migration was applied to localize pipes by extracting peaks. From the result of experimental field data, the algorithm provides the clear understandings of pipes' 3D positions and arrangement with reasonable calculation time.

In this article, we focus on ground-penetrating radar (GPR) for subsurface utility pipe detection. Due to the dense and high-speed 3-D monitoring, GPR is a promising tool. However, because of enormous amount of radar data and difficulty of interpretation, inspection time and cost are the bottlenecks. In this article, we propose a novel detection algorithm by the combination of 3-D convolutional neural network (3-D-CNN) and Kirchhoff migration. A 3-D-CNN architecture was trained utilizing transverse and longitudinal pipes’ measurement data. The classification accuracy of the developed model was about 91%, accurately estimating the pipes’ existences and directions. The 3-D-CNN improved the classification accuracy by about 6% compared to 2-D-CNN in the case of transverse pipes by considering the 3-D geometries of the pipes. After box-by-box search by 3-D-CNN, Kirchhoff migration was applied to cross section images and peaks were extracted. From the result of experimental field data, the algorithm provides the clear understandings of pipes’ 3-D positions and arrangement with reasonable calculation time.

In this paper, we focus on ground-penetrating radar (GPR) for infrastructural health monitoring, especially for the monitoring of reinforced concrete (RC) bridge slab. Due to the demand of noncontact and high-speed monitoring technique which can handle vast amounts of aging infrastructures, GPR is a promising tool. However, because radar images consist of many reflected waves, they are usually difficult to interpret. Furthermore, the spatial resolution of system is not enough considering the thickness of target damages, cracks, and segregation are millimeter-to-centimeter order while the wavelength of ordinary GPR ultrahigh-frequency band is over 10 cm. To address these problems, for the purpose of sensitive damage detection, we propose a new algorithm based on deconvolution utilizing a super high-frequency (SHF) band system. First, a distribution of reflection coefficient is inversely estimated by 1-D bridge slab model. Because concrete is found to be a lossy medium at SHF band, we consider the attenuation of signal in deconvolution. The algorithm is called “time-variant deconvolution” in this paper. After the validation by simulation, the effects of the algorithm and frequency band on damage detection accuracy are evaluated by a field experiment. Though the results show a 1-mm horizontal crack is not detected by measured waves, when it is filled with water, it is detected by time-variant deconvolution. Moreover, the 1-mm dried crack is detected only by time-variant deconvolution at SHF band, which greatly emphasizes the peaks of the reflection coefficient of the crack.