Synergistic Multi-Model Approach for GPR Data Interpretation: Forward Modeling and Robust Object Detection

Abstract

Ground penetrating radar (GPR) is widely used for subsurface object detection, but manual interpretation of hyperbolic features in B-scan images remains inefficient and error-prone. In addition, traditional forward modeling methods suffer from low computational efficiency and strong dependence on field measurements. To address these challenges, we propose an unsupervised data augmentation framework that utilizes CycleGAN-based model to generates diverse synthetic B-scan images by simulating varying geological parameters and scanning configurations. This approach achieves GPR data forward modeling and enhances the scenario coverage of training data. We then apply the EfficientDet architecture, which incorporates a bidirectional feature pyramid network (BiFPN) for multi-scale feature fusion, to enhance the detection capability of hyperbolic signatures in B-scan images under challenging conditions such as partial occlusions and background noise. The proposed method achieves a mean average precision (mAP) of 0.579 on synthetic datasets, outperforming YOLOv3 and RetinaNet by 16.0% and 23.5%, respectively, while maintaining robust multi-object detection in complex field conditions.

1. Introduction

As a non-invasive electromagnetic (EM) detection technology, Ground Penetrating Radar (GPR) is a cornerstone technique in modern non-destructive testing of civil engineering structures, owing to its superior operational efficiency and versatility across multiple scenarios. Its principal engineering applications encompass: 3D reconstruction of subsurface utility networks [1], quantitative analysis of internal defects in reinforced concrete components [2,3,4], interlayer damage evaluation in concrete bridge decks [5,6,7], void detection in pavement structural layers [8,9,10], the recognition of moisture damage in asphalt pavement [11], high-accuracy intelligent detection of centimeter-level voids in cement pavement [12] and assessment of construction process conformance [13]. Notably, GPR has also enabled groundbreaking planetary discoveries, exemplified by its critical role in identifying ancient Martian coastal deposits through rover-mounted radar systems [14]. The system operates by emitting nanosecond-scale EM pulses into layered media to capture reflected waveforms generated by dielectric discontinuities, which are subsequently synthesized into time-domain radargrams containing subsurface target signatures. Of particular significance is the observation that typical detection targets such as buried metallic conduits, concrete reinforcement, and geotechnical voids exhibit distinct hyperbolic diffraction patterns in B-scan images [15,16]. Crucially, the reliability of GPR data interpretation remains dependent upon precise extraction and physical characterization of target reflection signatures. Traditional manual interpretation methods face significant challenges when processing large-scale GPR scan data, including signal distortion due to complex noise interference, low processing efficiency, and susceptibility to subjective errors [8].

To address the inherent limitations of traditional manual interpretation, recent advancements in artificial intelligence have provided new pathways for enhancing GPR technology. Deep learning (DL), particularly convolutional neural networks (CNNs), has demonstrated exceptional performance in computer vision tasks such as image classification [17,18], semantic segmentation [19,20,21,22], and object detection [23,24,25], owing to its hierarchical feature extraction and spatial correlation capture mechanisms [26]. This data-driven pattern recognition paradigm aligns closely with the core requirements of GPR image interpretation. These requirements span feature extraction to target discrimination. The paradigm establishes end-to-end mapping relationships. These relationships directly associate EM reflection signatures with the spatial distributions of subsurface targets. This approach overcomes the limitations of empirical human interpretation.

In this context, DL is reshaping GPR data processing methodologies, offering significant potential for automating the critical task of object detection in B-scan images. However, the successful deployment of DL, particularly for supervised object detection, is highly contingent upon access to large-scale, high-quality training datasets. To mitigate the critical issue of data scarcity, forward modeling has been employed to generate synthetic training data. While theoretically effective, conventional numerical methods exhibit low computational efficiency and frequently fail to capture the full complexity and variability of real-world scenarios (e.g., heterogeneous soil layers, target interactions, complex noise) [27]. Notably, recent years have witnessed the emergence of DL approaches addressing the inverse problem in GPR, such as reconstructing permittivity distributions or target geometries from B-scan data [27,28,29]. These studies demonstrate DL’s capacity to learn complex mappings governing EM wave phenomena.

Inspired by the success of DL in addressing GPR inverse problems and acknowledging the limitations of conventional forward modeling, we explore the under-explored application of DL specifically for efficient and flexible forward modeling. By leveraging the powerful pattern learning and generation capabilities of deep neural networks, particularly generative adversarial networks (GANs) [30], this approach offers the potential to overcome the computational bottlenecks and scenario constraints inherent in traditional methods. The methodology aims to rapidly generate diverse synthetic B-scans, tailored to enhance datasets for training robust object detection models.

This study proposes a synergistic framework integrating a CycleGAN [31]-based data augmentation module with an EfficientDet-based interpretation module. The CycleGAN component establishes an image-to-image mapping that directly translates permittivity distribution patterns into synthetic B-scan images while preserving hyperbolic signatures. This capability expands training data diversity without modifying existing scans, serving as a critical augmentation strategy.

The core interpretation task—detecting hyperbolic signatures in B-scans—is performed by the EfficientDet architecture. EfficientDet incorporates a Bidirectional Feature Pyramid Network (BiFPN) module [32], which optimizes multi-scale detection through bidirectional cross-scale feature fusion. By dynamically integrating multi-resolution features, the BiFPN enhances the detection of small targets (e.g., shallow voids) and hyperbolic signatures.

This integrated framework operates bidirectionally: Data augmentation diversifies training samples, while interpretation extracts semantic subsurface information. Leveraging CycleGAN-generated cross-domain data, the system learns robust representations under complex conditions, overcoming feature diversity limitations of conventional methods and addressing critical challenges in real-data scarcity and complex feature interpretation.

The principal contributions of this work are:

(i).

Unsupervised GPR Data Augmentation Framework: A CycleGAN-based generation methodology is proposed, synthesizing diverse B-scan images through simulation of varying geological parameter and scanning configurations(This process is also known as GPR data forward modeling). This significantly enhances the geological rationality and generalization of synthetic data.

(ii).

BiFPN-Driven B-scan Feature Recognition: A detection architecture incorporating BiFPN is engineered for multi-scale hyperbolic targets in GPR imagery. By utilizing cross-scale dynamically weighted fusion mechanisms, the framework optimizes feature representation for small targets and hyperbolae while suppressing time-varying noise, thereby striking a balance between geometric feature preservation and detection efficiency.

2. Methodology

GPR data interpretation confronts two critical challenges: First, well-annotated B-scan samples are scarce in practical engineering contexts, and the complex variability of subsurface medium EM parameters severely constrains data diversity. Second, targets in B-scan images (e.g., pipelines and voids) typically manifest time-varying hyperbolic signatures, whose vertex positions and geometric characteristics are susceptible to noise interference, rendering conventional detection algorithms inadequate in both accuracy and generalization for field applications.

To address these issues, this study proposes a hybrid data-driven and model-driven framework. As illustrated in Figure 1, the implemented solution establishes a three-stage methodological pipeline: (1) Multi-source dataset preparation combining open-access measured data and numerical simulation models as foundational inputs for CycleGAN training; (2) forward-augmented data generation via CycleGAN-based cross-domain translation, which enhances training set diversity while preserving consistency between hyperbolic signatures and subsurface EM properties; (3) multi-scale hyperbolic object detection employing the EfficientDet architecture with compound scaling, specifically optimized for time-varying geometric patterns in GPR imagery. This workflow constructs an integrated pipeline from raw data preparation to intelligent interpretation, directly resolving the dual challenges of data scarcity and hyperbolic feature instability.

2.1. Architecture of Data Augmentation Model

2.1.1. CycleGAN Model

DL models for GPR image interpretation depend critically on large-scale, diverse training datasets. However, acquiring well-annotated B-scan samples in practical engineering contexts presents significant challenges: First, the distribution of subsurface targets (e.g., pipelines and voids) is inherently complex, rendering measured data collection labor-intensive, while public datasets remain limited due to confidentiality constraints on critical infrastructure. Second, conventional simulation methods (e.g., the finite-difference time-domain (FDTD)-based gprMax platform) can generate high-fidelity synthetic data; however, their computational inefficiency and resource-intensive nature restrict feasibility for large-scale data generation.

GANs have demonstrated exceptional performance in image generation [33,34] and image editing [35,36], powered by their adversarial loss mechanism that aligns generated data with target domain distributions. This loss continuously evaluates the realism of synthetic samples through a discriminator network, achieving the long-pursued objective of photorealistic rendering in computer graphics. In GPR data augmentation tasks, this adversarial training framework is adapted to learn mappings from subsurface permittivity distributions to corresponding B-scan images. However, acquiring strictly paired training samples (e.g., permittivity maps spatially aligned with B-scan images) remains highly challenging in GPR applications. To overcome this limitation, we implement the CycleGAN framework as illustrated in Figure 2, which employs bidirectional generators ($G:X\to Y$ and $F:Y\to X$) to establish unpaired domain mappings, thereby accounting for spatial variability in subsurface medium parameters.

The generator $G$ translates subsurface relative permittivity distributions (domain $X$, e.g., laboratory-standard permittivity maps) into B-scan radar images (domain $Y$, e.g., EM echoes under complex field conditions), while the generator $F$ performs the inverse mapping from B-scan images back to permittivity distributions. Each generator employs an enhanced U-Net architecture with an encoder–decoder structure to extract and reconstruct input features. The discriminator network adopts a PatchGAN architecture, consisting of two independent branches ($D_X$ and $D_Y$). The PatchGAN architecture divides input images into overlapping $N\times N$ pixel patches. Each discriminator independently evaluates patch authenticity through patch-level analysis, enabling precise capture of localized features in GPR signals.

During CycleGAN training, the generators $G$ and $F$ are iteratively optimized through adversarial training to produce realistic images that deceive the discriminators $D_X$ and $D_Y$. Simultaneously, the discriminators strive to distinguish real samples from generated ones. This dynamic competition drives the generators to learn domain-specific features while preserving key properties of the input data, achieving high-fidelity cross-domain translation.

2.1.2. Loss Function

To establish effective cross-domain mapping, CycleGAN incorporates multiple loss functions that enforce realism in synthesized images and structural fidelity. The core loss components comprise adversarial loss, cycle-consistency loss, and identity consistency loss.

The adversarial loss is applied to the mapping functions $G$ and $F$, along with their corresponding discriminators $D_Y$ and $D_X$. Specifically, the adversarial loss for generator $G$ is defined as:

$${\mathcal{L}}_{GAN}\left(G,D_Y,X,Y\right)={\mathbb{E}}_{Y\sim p_{data}\left(y\right)}\left[\mathit{log}{D}_Y\left(y\right)\right]+{\mathbb{E}}_{X\sim p_{data}\left(x\right)}\left[\mathit{log}\left(1-D_X\left(G\left(x\right)\right)\right)\right]$$

(1)

This loss enforces domain realism for images synthesized by generator $G$. Generator $G$ seeks to maximize the discriminator $D_Y$’s assessment confidence for its generated outputs $G\left(x\right)$, while $D_Y$ learns to differentiate authentic samples $y$ from synthesized counterparts $G\left(x\right)$. Through adversarial optimization, $G$ assimilates target domain features. Similarly, generator F and discriminator $D_X$ are governed by a symmetric adversarial loss formulation:

$${\mathcal{L}}_{GAN}\left(F,D_X,Y,X\right)={\mathbb{E}}_{x\sim p_{data}\left(x\right)}\left[\mathit{log}{D}_X\left(x\right)\right]+{\mathbb{E}}_{y\sim p_{data}\left(y\right)}\left[\mathit{log}\left(1-D_X\left(G\left(y\right)\right)\right)\right]$$

(2)

For a source domain sample $x$, generator $G$ maps it to the target domain as $y^{\prime }=G\left(x\right)$. Generator $F$ subsequently reconstructs the sample to the source domain as $x^{\prime }=F\left(y^{\prime }\right)$. The cycle-consistency loss minimizes the L1-norm distance between the original sample x and its reconstruction $x^{\prime }$:

$${\mathcal{L}}_{cyc}\left(G,F\right)={\mathbb{E}}_{x\sim p_{data}\left(x\right)}\left[\parallel F\left(G\left(x\right)\right)-x{\parallel }_1\right]+{\mathbb{E}}_{y\sim p_{data}\left(y\right)}\left[\parallel G\left(F\left(y\right)\right)-y{\parallel }_1\right]$$

(3)

This loss constitutes the fundamental innovation distinguishing CycleGAN from conventional unidirectional GANs. As depicted in Figure 2, the cyclic reconstruction mechanism enforces structural integrity and feature preservation during bidirectional domain transitions ($X\to Y\to X$ and $Y\to X\to Y$).

The identity consistency loss ensures that generators approximate identity transformations for samples already residing in the target domain. This loss is mathematically defined as:

$${\mathcal{L}}_{identity}\left(G,F\right)={\mathbb{E}}_{y\sim p_{data}\left(y\right)}[\parallel G(y)-y{\parallel }_1]+{\mathbb{E}}_{x\sim p_{data}\left(x\right)}[\parallel F(x)-x{\parallel }_1]$$

(4)

The total loss is a weighted sum of the above components:

$$\begin{gathered} \mathcal{L}\left(G,F,D_X,D_Y\right)={\mathcal{L}}_{GAN}\left(G,D_Y,X,Y\right)+{\mathcal{L}}_{GAN}\left(F,D_X,Y,X\right) \\ +{\lambda }_1{\mathcal{L}}_{cyc}\left(G,F\right)+{\lambda }_2{\mathcal{L}}_{identity}\left(G,F\right) \end{gathered}$$

(5)

Here, ${\lambda }_1$ and ${\lambda }_2$ are hyperparameters that balance the contributions of the cycle-consistency and identity losses, respectively.

2.2. Architecture of Object Detection Model

2.2.1. EfficientDet Model

The EfficientDet architecture adopts a single-stage object detection framework. Its core design principle optimizes the trade-off between detection accuracy and computational efficiency through an efficient backbone network and multi-scale feature fusion mechanism. The modular architecture comprises three cascaded components: a backbone network, a BiFPN, and a prediction head, as illustrated in Figure 3.

The backbone utilizes EfficientNet, pre-trained on ImageNet. It employs compound scaling to optimize depth, width, and resolution. This approach maintains robust feature representations while minimizing computational complexity. The backbone extracts hierarchical feature maps that provide multi-scale spatial and semantic information for subsequent fusion.

BiFPN functions as the core feature fusion module, recursively stacking bidirectional pathways to establish cross-scale feature interactions. This architecture integrates top-down and bottom-up information flows through weighted fusion of multi-resolution features, enabling synergistic combination of high-level semantics and low-level spatial details. The recurrent bidirectional topology significantly enhances multi-scale object characterization.

The prediction heads comprise parallel branches for classification and bounding box regression, employing shared weights across all feature levels. This parameter-sharing strategy reduces model complexity while enhancing robustness via multi-level supervision. The entire network undergoes end-to-end training to jointly optimize detection accuracy and inference efficiency.

2.2.2. BiFPN: Bidirectional Feature Pyramid Network for Multi-Scale Fusion

The evolution of multi-scale feature fusion methodologies has been marked by significant architectural advancements in feature pyramid networks. The foundational feature pyramid network (FPN) architecture (Figure 4a) establishes a unidirectional top-down pathway that propagates high-level semantic features to lower-resolution layers, enabling elementary cross-scale feature interactions. However, this unidirectional information flow intrinsically limits effective integration of low-level spatial details with high-level semantic representations, particularly degrading detection performance for small-scale targets.

To overcome these limitations, BiFPN (Figure 4b) implements an optimized recurrent fusion mechanism. This architecture first streamlines the feature hierarchy by pruning single-input nodes exhibiting marginal contribution to detection outcomes, selectively retaining core layers supporting multi-directional connectivity. It subsequently constructs a bidirectional propagation framework:

The top-down pathway, visually represented as the green part, integrates upsampled high-level semantic features with high-resolution low-level features, effectively conveying rich contextual information. Conversely, the bottom-up pathway, depicted as the orange part, facilitates hierarchical refinement by propagating low-level spatial features (conveying precise positional information) through downsampling operations to semantically enriched higher layers. Critically, each fusion node dynamically aggregates dual-source inputs: intrinsic features from the current resolution level and cross-scale features from adjacent layers accessed via skip connections. Furthermore, blue edges represent the additional connections added between input nodes residing on the same layer, enhancing intra-level feature exchange. This integrated design, combining the bidirectional flows and the novel same-layer connections, enables simultaneous feature reuse and information enhancement, achieving the balanced spatial precision and contextual awareness essential for hyperbolic target detection in GPR B-scan imagery.

3. GPR Dataset Construction

This study establishes an integrated multi-source GPR dataset for subsurface object detection, integrating three distinct data modalities: numerically simulated B-scan images, synthetic B-scan images, and field-measured data. Each sample is paired with corresponding EM property characterization through relative permittivity distribution maps. To mitigate insufficient subsurface target samples, 1000 high-fidelity simulated samples were generated using the gprMax3.0 [37] EM simulation platform. Augmented with 98 open-source field datasets [38], these simulated samples trained a CycleGAN model to synthesize 500 additional data pairs (synthetic data), expanding the original dataset. The standardized composite dataset comprises 1598 samples (1000 simulated + 500 synthetic + 98 measured). From this repository, 1469 B-scan images were annotated with pixel-level hyperbolic target delineations using LabelImg1.8.1, forming an object detection dataset with detailed composition in

Table 1. Composition and characteristic parameters of multi-source GPR dataset.

Dataset	Source	Quantity	Train/Val	Description
Simulated data	gprMax3.0 [37]	1000 pairs	1386 pairs/212 pairs	Paired permittivity maps and B-scan images
Measured data	Open-source [38]	98 pairs
Synthetic data	CycleGAN	500 pairs
Object Detection data	LabelImg1.8.1	1469 images	1214/255 images	Annotated B-scans

.

3.1. Simulated Dataset

This study addresses the data augmentation requirements for GPR B-scan images by constructing a multi-domain fusion dataset based on a collaborative framework integrating EM simulation and DL techniques. To overcome challenges such as the scarcity of real GPR data and high annotation costs, the open-source software gprMax3.0 [37] was utilized to establish refined subsurface models. Large-scale simulated B-scan images were generated using the FDTD method as the foundational dataset.

Key modeling parameters were configured as follows: The computational domain comprised a 2.5 m × 0.5 m 3D space. Spatial discretization employed an isotropic grid ($\Delta x=\Delta y=\Delta z=2.5 \mathrm{m}\mathrm{m}$), with a 15 ns time window configured to align with typical road inspection scenarios. Stratigraphic layers included (top-to-bottom): 10 cm air layer, 5 cm asphalt layer, and 10 cm concrete layer. Grid resolution was set at 0.0025 m × 0.0025 m, yielding 1000 × 100 computational cells. EM excitation utilized a z-polarized Hertzian dipole with a Gaussian pulse waveform (center frequency: 900 MHz). Transmitting and receiving antennas were positioned 2.5 mm above the concrete surface. During acquisition, the transceiver system traversed synchronously along the horizontal axis at 20 mm intervals, capturing 1000 A-scans per B-scan. To ensure diversity, tubular targets with randomized geometries and positions were embedded subsurface, ultimately yielding 2000 B-scan datasets through parametric scripting. Relative permittivity and conductivity values per medium are detailed in

Table 2. Model medium parameter information.

Model Medium	Concrete	Asphalt	Free_Space (Cavity)
Relative permittivity	6.4	4	1
Conductivity	0.01	0.005	0

.

To suppress interference from direct waves in radar images, exponential gain compensation is applied to enhance the hyperbolic response characteristics in B-scan data, while median filtering is employed to effectively separate direct wave components. This processing method significantly improves the discernibility of target reflection signals in GPR imaging data. The final simulated B-scan data are shown in Figure 5a. To enable the model to focus on the geometric distribution of subsurface targets in relative permittivity maps, the input permittivity maps are uniformly processed by rendering solid layers (excluding air and cavities) in black and air/cavity layers in white. The processed relative permittivity maps are illustrated in Figure 5b.

3.2. Measured Dataset

The measured GPR data were sourced from a publicly available open-source [38] dataset. The original measurements were collected by researchers using a commercial GSSI Utility Scan Pro GPR system in an outdoor sandy environment. The dataset includes multiple B-scan radar images and corresponding subsurface relative permittivity distribution information, making it suitable for algorithm validation and subsurface target reconstruction. During data acquisition, a 400 MHz antenna and control unit were employed to capture data along predefined scanning paths. Each scan trajectory spans 1 m, containing 88 A-scan signals with a sampling time window of 20 ns. The dataset underwent standardized preprocessing, including normalization, resizing, and mean subtraction, and provides denoised reference B-scan images and ground-truth relative permittivity maps. The final processed measured B-scan images are shown in Figure 6a. Similar to the simulated dataset, to enable the model to focus on the geometric distribution of subsurface targets in relative permittivity maps, the measured dataset’s permittivity maps were converted to black-and-white representations, with solid layers rendered in black and air/cavity layers in white. The final processed relative permittivity maps are illustrated in Figure 6b.

3.3. GPR Data Augmentation Results by CycleGAN

The CycleGAN model was trained using the Adam optimizer (initial learning rate = 0.0002, β₁ = 0.5) for 100 epochs, with the first 50 epochs at a fixed learning rate and the subsequent 50 epochs linearly decaying to zero. The loss function integrated LSGAN (least squares GAN), cycle-consistency loss (λ₁ = 10), and identity consistency loss (λ₂ = 0.5) to jointly optimize the plausibility of generated images and geometric alignment in cross-domain mapping.

To objectively demonstrate the computational efficiency advantage of the proposed CycleGAN-based forward modeling approach over traditional numerical simulation (e.g., gprMax), a comprehensive time cost analysis was conducted. As shown in

Table 3. Computational efficiency comparison of forward modeling approaches.

Method	Quantity	Time Cost	Resolution	Notes
gprMax (v3.0)	1000 pairs	20 days	3543 × 1772 pixels	Numerical simulation (B-scans + permittivity maps)
Proposed CycleGAN	-	10 h	-	Model training
	200 B-scans	2 min	1000 × 180 pixels	Image generation (inference)
Measured data	98 pairs	-	128 × 128 pixels	Open-source dataset; field acquisition excluded

, experiments were performed on a workstation equipped with an NVIDIA RTX 4070 GPU and an Intel i5-13600KF CPU. The conventional gprMax (v3.0) platform, based on the FDTD method, was used to generate high-fidelity B-scan images. Under a typical road inspection scenario configuration (center frequency: 900 MHz antenna, spatial grid resolution $\Delta x=\Delta y=\Delta z=2.5 \mathrm{m}\mathrm{m}$, time window: 15 ns, 1000 A-scans per B-scan), generating a high-fidelity simulated dataset comprising 1000 pairs of samples (B-scan images and corresponding relative permittivity maps) took approximately 10 days.

In contrast, the proposed CycleGAN model, once trained (training time: 10 h), demonstrated extremely high efficiency in generating new synthetic B-scan images: generating 200 B-scan images with equivalent resolution (1000 × 100 pixels) and complexity to those from gprMax3.0 required only about 2 min. The measured data were sourced from an open-access public dataset; the original acquisition time involving fieldwork is not included in this computational efficiency comparison. Although the CycleGAN model training requires a significant computational investment (10 h), once trained, it exhibits orders of magnitude efficiency improvement (minutes vs. days) in generating new samples. This “train once, generate efficiently” paradigm is particularly advantageous for scenarios demanding large-scale, diverse synthetic data to support DL model training, effectively overcoming the primary bottlenecks of computational intensity and long time consumption inherent in traditional numerical simulation methods.

To objectively demonstrate the computational efficiency advantage of the proposed CycleGAN-based forward modeling approach over traditional numerical simulation (e.g., gprMax), a comprehensive time cost analysis was conducted under a typical road inspection scenario. Experiments were performed on a workstation equipped with an NVIDIA RTX 4070 GPU and an Intel i5-13600KF CPU (Table 3). The conventional gprMax platform, based on the FDTD method, was used to generate high-fidelity B-scan images. For a dataset comprising 1000 sample pairs (B-scan images at 3543 × 1772 pixels and corresponding permittivity maps), the generation process took approximately 20 days. In contrast, the proposed pre-trained CycleGAN model generated 200 synthetic B-scan images at a resolution of 1000 × 180 pixels in only about 2 min. It is important to note that while the gprMax simulations model the complex physical processes resulting in high-resolution, physically accurate outputs, the CycleGAN generates synthetic data primarily for efficient data augmentation to support DL model training. The measured validation data originated from an open-access public dataset; its field acquisition time is excluded from this computational comparison.

The key observation is the dramatic difference in sample generation throughput after the initial investment. Although training the CycleGAN model required a significant 10 h computational investment, the subsequent generation of new synthetic samples is orders of magnitude faster (minutes vs. over two weeks) compared to FDTD simulation. This approach, characterized by a substantial upfront training cost followed by highly efficient generation, fundamentally shifts the computational bottleneck. The one-time training investment enables the rapid, scalable production of synthetic data tailored for DL. This efficiency is particularly transformative for applications demanding large volumes of diverse training data, effectively overcoming the prohibitive computational cost and time constraints inherent in traditional physics-based simulation methods like gprMax for iterative data generation tasks.

To validate the effectiveness of CycleGAN in expanding GPR B-scan image datasets, this study systematically compared its performance with the traditional pixel-aligned model (pixel to pixel) [39]. Note that the ground-truth data in Figure 7 (second row) were generated using classical FDTD-based simulations via gprMax3.0, providing the benchmark for evaluating CycleGAN outputs.

As shown in Figure 7c, while pixel-to-pixel-generated B-scan images exhibited basic hyperbolic contours, the vertex positions deviated significantly from the spatial coordinates of targets in the input permittivity maps; in multi-target scenarios (Figure 7a), it failed to match the number of input targets, with exacerbated positional errors at vertices, indicating its inability to model the complex coupling effects of EM reflections among multiple targets. In contrast, CycleGAN, through the joint optimization of cycle-consistency loss and adversarial training, achieved superior geometric fidelity in cross-domain mapping.

As illustrated in Figure 7b,d, CycleGAN-generated B-scan images strictly preserved the geometric distribution of targets in the input permittivity maps and precisely replicated the spatiotemporal evolution of hyperbolic vertices; for instance, when the input contained two adjacent pipelines (Figure 7b), the generated image displayed independent hyperbolic reflections with vertex spacing fully consistent with the input’s spatial distribution, confirming its capability to achieve high-fidelity visual mapping that preserves the geometric relationships implied by the permittivity distribution.

Further validation on field-measured scenarios (Figure 8) demonstrated CycleGAN’s robust generation capability under strong clutter interference. Despite significantly higher background noise in measured B-scan images compared to synthetic data, CycleGAN-generated images accurately identified target reflection features while dynamically suppressing clutter, maintaining precise localization of hyperbolic vertices. In contrast, Pixel to Pixel exhibited consistent limitations in measured data: generated hyperbola vertices retained positional deviations and failed to reflect multi-target quantitative relationships, merely approximating hyperbolic shapes. This indicates that pixel to pixel struggles to adapt to the complex requirements of GPR B-scan data augmentation tasks.

4. Results and Discussion

4.1. Experimental Setup and Evaluation Metrics

The experimental platform was developed on the PyTorch 2.4.0 framework with an Intel i5-13600KF CPU and NVIDIA RTX 4070 GPU hardware. Mixed-precision training acceleration was implemented via CUDA 12.1 to fully exploit GPU parallel computing capabilities.

The model was trained using the AdamW optimizer with an initial learning rate of 1 × 10⁻⁴. A dynamic learning rate adjustment strategy (ReduceLROnPlateau) was implemented, automatically reducing the learning rate by 50% when validation loss plateaued for three consecutive epochs. Training spanned 100 epochs with a fixed batch size of 3. To ensure robustness, intermediate model weights were saved at every 202nd iteration, enabling recovery from interruptions and optimal checkpoint selection.

To comprehensively evaluate the performance of the object detection model, four key metrics were adopted, covering both precision and recall aspects across varying Intersection over Union (IoU) thresholds: mean average precision (mAP) comprehensively reflects the overall detection performance averaged over IoU thresholds from 0.5 to 0.95 (step size 0.05), mAP50 represents the average accuracy at an IoU threshold of 0.5, mAP75 corresponds to the detection accuracy at an IoU threshold of 0.75, and AR (Average Recall) quantifies the mean maximum recall rate of each target category within the IoU range of 0.5 to 0.95.

4.2. Performance Evaluation

The EfficientDet was trained and tested on the object detection dataset. Experimental results, as illustrated in Figure 9, indicate accurate target identification on both datasets. For simulation data, we restored the original coordinate system. The horizontal axis corresponds to 125 measurement positions along a 2.5 m scan line with 20 mm intervals, indexed from 1 to 125. The vertical axis represents the time range of 0–15 ns divided into three segments: 0–5 ns, 5–10 ns, and 10–15 ns. During CycleGAN training, we intentionally removed axis labels. This was done to help the model focus on learning the intrinsic features of the B-scans without being distracted by coordinate information. In the revised manuscript, we have reinstated full coordinate annotations to all synthetic B-scans.

Figure 10 further validates the model’s adaptability in real-world environments. Measured data present challenges like background clutter and random noise interference. Despite this, the model accurately identifies low-contrast hyperbolic features. It achieves this without a miss or false alarm. The results demonstrate that EfficientDet delivers consistent multi-object detection performance across cross-domain data (simulated and measured), proving its effectiveness and stability in hyperbolic feature detection tasks for GPR B-scan images. It should be noted that the measured dataset is an open-source collection that inherently lacked corresponding ground truth bounding box coordinate information upon acquisition. Consequently, the prediction results for measured data in the final figures were not supplemented with this information. Regarding the simulation data, to prevent coordinate axis interference during CycleGAN’s image-to-image translation, we intentionally removed axis labels during training. For the final result figures, we have added explicit time-scale annotations to all synthetic B-scan images. These annotations directly reflect the two-way travel time of EM waves, which correlates with target depth (i.e., greater depth corresponds to longer return time).

During training, both classification and regression losses exhibit stable convergence trends as iterations progress (Figure 11). As shown in Figure 11a, the classification loss decreases steadily to low values, indicating progressive optimization of target category recognition. As illustrated in Figure 11b, the regression loss initially fluctuates but eventually converges, reflecting effective learning of bounding box localization. Neither loss shows significant oscillations or secondary increases, suggesting that the training process remains unaffected by gradient anomalies or overfitting. The synchronous convergence of both losses underscores the model’s balanced optimization of classification and localization tasks, validating the efficacy of the EfficientDet framework in detecting hyperbolic features in B-scan images.

4.3. Comparison with Different Object Detection Models

To prove our model works better, we compared it with two popular detection models—RetinaNet and YOLOv3. To keep things fair, all models were trained with exactly the same data for 100 cycles and tested under the exact same settings. Experimental results on the simulated dataset, as summarized in

Table 4. Detection method evaluation results.

Model	mAP	mAP50	mAP75	AR
RetinaNet	0.469	0.752	0.516	0.528
YOLOv3	0.499	0.796	0.563	0.551
EfficientDet	0.579	0.986	0.578	0.667

, demonstrate the comprehensive advantages of the EfficientDet in detection performance. Its overall detection capability (mAP = 0.579) surpasses YOLOv3 and RetinaNet by 16.0% and 23.5%, respectively, validating the effectiveness of the proposed architecture. Under lenient localization requirements (mAP50 = 0.986), the detection accuracy approaches theoretical limits, indicating extremely high confidence in target presence determination. However, in stringent localization scenarios (mAP75 = 0.578), the performance improvement narrows, which may be attributed to target edge ambiguity in complex environments. Notably, the model achieves a significantly higher average recall rate (AR = 0.667) compared to baseline models (YOLOv3: 0.551; RetinaNet: 0.528), proving its enhanced capability to suppress missed detections for multi-scale targets.

Experimental results demonstrate that EfficientDet exhibits significant advantages in both simulated and real-world scenarios (Figure 12 and Figure 13). Specifically, its detection confidence remains stably distributed within the 0.9–1.0 range, and the IoU values of predicted bounding boxes relative to hyperbolic vertices (mAP75 = 0.578) significantly outperform baseline models, validating the robustness of BiFPN multi-scale feature fusion against geometric distortions and noise interference. In contrast, RetinaNet shows an abnormal concentration of detection confidence in high-value intervals (Figure 12b), indicating an overfitting tendency that leads to a higher false negative rate (AR = 0.528) in low-contrast scenarios. Meanwhile, YOLOv3’s positioning accuracy (mAP75 = 0.563) is constrained by the preset anchor box mechanism’s insufficient adaptability to the dynamic morphology of hyperbolic features.

In measured data (Figure 13), EfficientDet further validates its engineering applicability: it maintains high detection accuracy without missed or false alarms under strong background clutter and noise interference, while RetinaNet and YOLOv3 exhibit significantly weaker performance in complex scenarios. The consistency between simulated and measured results confirms that CycleGAN data augmentation and BiFPN dynamic fusion strategies effectively enhance the model’s cross-domain generalization capability.

5. Conclusions

This study addresses the critical challenge of automating hyperbolic target detection in GPR B-scan images, hindered by scarce annotated real-world data and complex signatures under noise. We propose a novel framework combining CycleGAN for unsupervised data augmentation and EfficientDet for robust detection. CycleGAN generates diverse synthetic images by translating limited real data across geological domains. EfficientDet, trained on this augmented data, demonstrates substantial improvements in handling distortions and noise. Experimental validation confirms effectiveness, achieving a 0.579 mAP on synthetic datasets—surpassing YOLOv3 and RetinaNet by 16.0% and 23.5%, respectively. Analysis reveals persistent challenges:

(i).

Recall-Precision Trade-off: While precision under lenient localization thresholds is exceptional (mAP50 = 0.986), the average recall (AR = 0.667), though superior to baselines, indicates potential missed detections, particularly for low-contrast targets or in high-clutter environments.

(ii).

Localization Precision under Strict Criteria: The significant performance gap between mAP50 (0.986) and mAP75 (0.578) highlights difficulties in precisely localizing hyperbolic vertices under stringent IoU thresholds, likely due to residual sensitivity to geometric distortions and noise affecting vertex accuracy.

Future research will pursue several pivotal objectives, with these efforts targeting current methodological limitations to advance non-destructive testing technologies toward higher precision and broader applicability in complex environments:

(i)

Replacing binary permittivity maps to enable the generation of B-scans depicting more complex subsurface scenarios, such as soil-concrete interfaces and gradual dielectric transitions, thereby increasing the visual complexity and diversity of the synthetic training data.

(ii)

Developing detection methodologies for closely spaced or overlapping subsurface targets to address current limitations in resolving adjacent objects.

(iii)

Establishing systematic C-scan processing workflows to shift focus beyond current B-scan hyperbolic signature detection toward exploring 3D subsurface interpretation.

(iv)

Exploring technical pathways for integrating EM wave propagation laws (Maxwell’s equations) into neural networks to enhance the physical plausibility of synthetic B-scans. This direction aims to combine physics-guided DL frameworks with generative models, improving the realism of reflection/attenuation behavior while maintaining data generation efficiency.

These efforts collectively target current methodological limitations. The goal is to enhance non-destructive testing technologies, achieving higher precision and practical utility in complex environments.