Ground-penetrating radar (GPR) 3-D modeling can intuitively reveal complex subsurface structures, facilitating a more comprehensive understanding and interpretation of the data. In practical engineering applications, the collected data often exhibit high sparsity and low consistency, posing challenges in directly constructing high-precision 3-D models. This article presents a method for generating dense arrays aimed at 3-D modeling. It employs a multifeature fusion net to extract temporal–space and temporal–frequency information from sparse GPR slices, aiming to achieve highly consistent coupled features. multivariate variational mode decomposition (MVMD) is utilized for multifrequency decomposition of input slice data. Additionally, a temporal–frequency feature encoder is designed, integrating continuous wavelet transforms (CWTs) to extract slice spectrograms. Temporal–space features are extracted by another encoder and weighted with temporal–frequency features to form a new multifeature representation. To establish strong correlations between sparse arrays, a correlation tracking network is proposed as the terminal model. Embedded within the network is the feature enhancement and alignment module (FEAM), which enhances bidirectional feature similarity and redistributes features to align feature maps. Contextual features and correlated volumes jointly update motion fields and intermediate features in a multiscale manner, synthesizing final intermediate data and enhancing slice array density. Experimental results demonstrate the network’s strong capabilities for continuous multilevel expansion across large interline spacings and robust performance on both simulated and real data.

In the context of ground-penetrating radar (GPR) surveys for underground engineering and pipeline identification, the processing of electromagnetic reflection data is pivotal for interpreting survey outcomes. The presence of substantial random noise and clutter within these data significantly complicates the imaging process. Despite the advancements brought by deep networks in GPR imaging technology, there remains a pressing need for more targeted techniques to enhance imaging reliability. This study introduces attention-driven deep network designed to enhance the perception of underground pipeline reflection features. The proposed network employs a dual-generative adversarial network (GAN) architecture: hyperbolic extraction (HE) GAN and target pipeline imaging GAN. The HE GAN leverages ResNet as the base model and utilizes localized perception hyperbolic attention to extract high-resolution hyperbolic waves. Meanwhile, the target pipeline imaging GAN, configured with U-Net and driven by rectified hyperbolic attention (RHA), incorporates multilevel attention mechanisms with skip connections to better capture and preserve fine details within the data. Experimental results demonstrate that the localized perception hyperbolic attention mechanism significantly enhances the response to hyperbolic wave features, effectively isolating these features while mitigating clutter and noise interference, thereby improving the reliability. RHA improves the accuracy of the pipeline imaging process.

Ground Penetrating Radar (GPR) is an efficient non-destructive testing tool used for detecting and locating buried pipelines. It helps to avoid interference with existing pipelines and determine optimal layouts, resulting in time and cost savings. However, applications in this domain often require the joint observation of sequential images, mapping from 2D B-scans to 3D spatial structures. Complex underground environments, equipment orientations, noise, and data deviations can introduce visual distortions, blurriness, and unclear structures in the collected data. Therefore, there is a need for a method to rapidly comprehend and visually analyze the true conditions of underground pipeline structures. GPR data is typically collected and stored in the form of two-dimensional B-scan sequences. In this paper, we propose a network framework that takes sparse original 2D B-scan sequences as input and outputs a dense three-dimensional target model. We first employ a Transformer model to interpolate the B-scan slice collection, generating dense 3D B-scan volume data. Subsequently, a from-coarse-to-fine back-projection strategy, based on the Transformer model, constructs a 3D volume data inversion-mapping model to transform 2D hyperbolic waves into three-dimensional pipeline information. Additionally, we apply a clutter removal mechanism based on Conditional Generative Adversarial Networks (CGAN) to Declutter and enhance the visualization of the desired hyperbolic wave structures, improving the accuracy of 3D visual imaging. Experimental results demonstrate that the proposed method is better suited for structural analysis of GPR pipeline data, particularly in complex real-world data experiments, affirming the effectiveness and practicality of the approach presented in this paper.

Ground Penetrating Radar (GPR) forward modeling holds significant importance in the realms of geological exploration, subsurface target detection, and scientific research. With the growing use of deep learning in GPR data processing, the need for extensive datasets that closely resemble real-world scenarios has become essential. However, acquiring such Approaching-reality datasets, is challenging. This paper introduces FM-GAN (Forward Modeling GAN), a network designed to generate B-scan images of underground pipelines across different mediums. Our adaptive spatial polarization generator, part of this network, processes dielectric constant images to create B-scan images that closely mimic real environmental data. It includes spatially adaptive normalization modules, Adaptformer modules, and an Double SimAM Feature Fusion Attention Module (DSFFA) for feature fusion and B-scan image generation. We train the network using paired data from underground dielectric constant models and B-scan data simulated through Finite-Difference Time-Domain (FDTD) techniques.Specifically, we apply Transformer fine-tuning methods to enhance its adaptability to real-world environments. This research combines deep learning, Conditional Generative Adversarial Networks (CGAN), and the Vision Transformer (ViT) to model underground pipeline dielectric constants. Empirical results show that our approach performs similarly to the FDTD method on both single and mixed dielectric constant data, significantly improving computational efficiency. This methodology holds promise for advancing underground structure imaging and interpretation, paving the way for innovative applications in underground surveys and ground-penetrating radar technology.

Underground pipelines are often detected and identified using ground-penetrating radar (GPR), which requires an inversion process to obtain specific pipeline properties. Unfortunately, this process is an ill-posed problem that involves inferring complex subsurface structures from a small number of observations, which leads to diverse and complex descriptions of the target. To address this challenge, this study proposes an inversion method based on a coupled-learning generative adversarial network. First, we extract hyperbolic waves from GPR buried-object B-scan images and filter out non-pipe and non-homogeneous media information from the resulting dielectric constant predictions. The training set comprises two data pairs: simulated clutter-free data with simulated clutter data pairing and simulated clutter-free data with dielectric constant data pairing. An experiment with real measured B-scan sequences showed that our proposed method provides clear and visually analyzed results about the location and direction of underground pipes. This approach enhances the accuracy and reliability of the inversion process and improves the clarity of subsurface pipeline property descriptions.

The acquisition of Ground Penetrating Radar (GPR) data is often impeded by clutter, which poses a significant obstacle to the effectiveness of target detection algorithms. This paper presents a novel approach to address this challenge by developing a flexibility-residual BiSeNetV2 for clutter suppression of GPR images. Our proposed network incorporates the flexibility-residual block into BiSeNetV2, allowing for adaptively selected convolutional kernel sizes based on the number of channels and network parameters required for different tasks, thereby ensuring effective mitigation of network degradation while minimizing the impact on time complexity. Moreover, we integrate an ECA attention mechanism into the network, which employs 1-dimensional convolutional local cross-channel interaction to extract inter-channel dependencies efficiently. As a result, the size of the 1-dimensional convolutional kernel can be adaptively selected according to the number of channels, determining the coverage of cross-channel interactions. Additionally, we adjust the ratio of multiple output losses in the network to optimize its suitability for our task. Experimental results demonstrate the effectiveness of our network for clutter suppression of cluttered images, and the network trained with the simulated dataset also performs better when processing measured GPR data.