Enhanced Extraction of Rebar Parameters in Ground Penetrating Radar Data of Bridges Using YOLOv8 Detection Under Challenging Field Conditions

Abstract

Accurate detection of reinforcing bars (rebars) in concrete structures using ground penetrating radar (GPR) is crucial for effective structural evaluation but remains challenging, particularly when asphalt overlays compromise signal clarity. This study evaluates the performance of deep learning-based rebar detection using the You Only Look Once version 8 (YOLOv8) object detection model across three GPR datasets categorized as clear, interfering, and blurry. Models trained on each category were applied across varying conditions to assess generalization and robustness. A filtering algorithm was introduced to eliminate redundant and overlapping detections, thereby significantly improving the accuracy of YOLOv8-based predictions. The YOLOv8 approach outperforms traditional analytical techniques, especially under noisy or complex scenarios. In blurry GPR images where analytical methods fail, the filtered YOLOv8 model accurately detects rebar with a count that closely matches the ground truth. Across different datasets, the YOLOv8 approach demonstrates improved consistency in both location and quantity estimation, with filtered predictions correcting substantial over-detection seen in raw outputs. The study presents a practical framework for applying deep learning to GPR data, enhancing the reliability of rebar detection under diverse field testing and evaluation conditions. The findings highlight the importance of developing tailored training datasets and post-processing strategies when deploying AI tools for in-service bridge inspections.

1. Introduction

A broad spectrum of nondestructive evaluation (NDE) techniques has been developed to assess the condition of civil infrastructure without causing damage. These methods encompass visual inspection, ultrasonic and acoustic testing, magnetic particle and eddy current analysis, radiography, thermography, optical methods, microwave scanning, penetrant testing, acoustic emission monitoring, and ground penetrating radar (GPR) [1].

GPR continues to gain attention because of its effectiveness in locating reinforcing steel rebars and identifying subsurface anomalies in concrete structures. Such attention can be attributed to the GPR’s ability to provide rapid and non-invasive imaging, which makes it valuable for assessing structural integrity and informing maintenance strategies for transportation infrastructure. The increasing adoption of GPR in structural evaluations reflects its growing recognition as a reliable diagnostic tool in civil infrastructure. Recent guidelines from the American Concrete Institute, including ACI PRC-228.3-23 [2] and ACI 228.2R-13 [3], acknowledge the role of GPR in evaluating internal conditions and embedded elements in concrete structures.

Environmental conditions significantly influence GPR performance, particularly in detecting subsurface features such as rebars in concrete. Moisture from rainfall or snow can increase soil or overlay conductivity, resulting in greater signal attenuation and a reduced penetration depth. Similarly, freeze–thaw cycles alter the dielectric properties of the material, causing increased scattering and degradation of signal clarity [4,5]. Surface moisture after rainfall or during snow melt also introduces low-frequency noise and raises background signal levels, effectively masking weak reflections from embedded objects [6]. Zatar et al. [7] conducted experiments in an environmental chamber to investigate the impact of chloride content, rust from rebar corrosion, ambient temperature, and relative humidity on the GPR signal amplitude. The study found that both chloride contamination in concrete and rust formation on the steel surface reduced the amplitude of rebar reflections. Moreover, higher chloride concentrations led to greater signal attenuation. Under the experimental conditions, the reflection amplitudes from corroded rebars were approximately 1 dB lower than those from non-corroded rebars.

While prior research has introduced both analytical and deep learning-based methods to enhance rebar detection from radargrams, most approaches have been validated only under controlled or lightly obstructed conditions. For example, Zatar et al. [8,9] demonstrated the accurate localization of rebar using an analytical model for clean and moderately noisy radargrams.

Several researchers have reported on the use of a Convolutional Neural Network (CNN) for high-accuracy surface and deep detection [10,11,12,13]. Yang et al. [14] and Wang et al. [15] studied the use of deep learning to detect defects in manufacturing. Tian and Jia [16] developed a rapid detection method for steel surface defects. Several other researchers have examined on industrial and aviation systems [17,18,19,20].

Previous research conducted by the authors has demonstrated that deep learning models, especially those based on CNNs, can automate rebar detection from radargrams with high accuracy, even under moderate signal degradation. Among the many deep learning frameworks developed for object detection, the You Only Look Once (YOLO) series is widely recognized for its high-speed performance and real-time detection capabilities [21,22,23,24]. In 2020, Li et al. [25] implemented the YOLOv3 algorithm using Google’s TensorFlow framework to perform real-time pattern recognition in GPR images. A recent study proposed a two-stage strategy for detecting grouting defects in bridge tendon ducts by combining the impact-echo method with machine learning and spectrogram-based classification, achieving accuracies exceeding 90% in both laboratory and full-scale girder tests [26].

Building on this research, Li et al. [27] conducted a comparative study of YOLOv5 against YOLOv3 and YOLOv4, demonstrating that YOLOv5 delivered notable improvements, especially when trained on smaller datasets. It also showed enhanced robustness in detecting and distinguishing features within GPR data. Qiu et al. [28] applied YOLOv5 [29] for real-time target detection and coordinate localization in GPR imagery, highlighting the algorithm’s growing effectiveness in subsurface object identification. The YOLOv8 architecture has emerged as a fast and robust solution for real-time object detection [30,31]. The authors have trained separate YOLOv8 models using radargrams with varying levels of signal clarity, including clear hyperbolas from laboratory specimens and blurry hyperbolas from bridge beams with asphalt cover.

This study aims to evaluate the effectiveness of deep learning models for rebar detection in GPR radargrams under challenging field conditions. Unlike previous YOLO-based rebar detection studies, which were primarily conducted under controlled laboratory conditions, this research applies YOLOv8 to GPR data from in-service bridge beams with asphalt overlays, addressing real-world challenges such as signal attenuation, noise, and variable field conditions. Three YOLOv8 models were trained using distinct datasets: clear, interfering, and blurry. The clear-trained model was applied to field scans targeting stirrups beneath asphalt overlays, demonstrating strong performance in moderately attenuated conditions. The interfering-trained model was utilized for high-noise transverse scans of concrete beams in the laboratory, where voids and nearby rebars heavily distorted signal reflections. The blurry-trained model was used on radargrams collected from in-service bridge beams with thick asphalt and concrete cover, where signal quality was severely degraded. A post-processing filtering algorithm is needed to remove overlapping and potentially inaccurate hyperbolas.

2. Research Significance

This study represents one of the first applications of YOLOv8 for detecting reinforcing bars in GPR data collected from in-service bridge beams with asphalt overlays. In this context, signal attenuation and noise present substantial challenges. While previous research has primarily focused on controlled laboratory conditions or earlier versions of YOLO, this study advances the field by addressing the practical constraints of real-world bridge inspections.

The significance of this study is fourfold: (1) Targeted dataset design: Separate YOLOv8 models were trained on clear, interfering, and blurry datasets to capture the variability encountered in field-acquired GPR data. (2) Improved detection reliability: A custom filtering algorithm was introduced to remove duplicate and spurious detections, improving robustness in noisy environments. (3) Adaptability in challenging conditions: The proposed method demonstrates consistent performance in low-visibility scenarios caused by asphalt overlays and other field conditions. (4) Practical engineering integration: The developed workflow maps detected reinforcement into engineering application, supporting decision-making for maintenance and rehabilitation planning. By directly addressing conditions typical of in-service bridges, the study bridges the gap between laboratory research and deployable field-ready solutions for structural evaluation and health monitoring.

The proposed workflow is summarized in a technical roadmap illustrated in Figure 1. This figure visually illustrates the logical progression of the study, highlighting how each stage contributes to the accurate and reliable detection of rebar in real-world conditions. The roadmap outlines the entire process, from GPR (ground penetrating radar) data collection to engineering application. It consists of six key stages: (1) data collection from laboratory and field bridge beams under varying conditions, (2) preprocessing to prepare YOLOv8-compatible training data, (3) model training for clear, interfering, and blurred datasets, (4) detection and filtering to refine predictions, (5) performance evaluation using quantitative metrics, and (6) engineering application of detection results for bridge inspection and maintenance planning

3. Research Methodology

GPR is a geophysical method that employs high-frequency electromagnetic (EM) pulses to generate images of subsurface features. This technology operates by emitting EM waves from a transmitting antenna, which then travel through the ground or structure and reflect off buried objects or material interfaces. Frequencies typically range from 200 MHz to 2.6 GHz for most civil engineering applications. The dielectric constant and electrical conductivity influence the propagation behavior of these waves.

As EM energy penetrates a medium, it undergoes attenuation, meaning the signal gradually weakens. In low-conductivity (i.e., resistive) materials such as dry concrete, ice, or sand, the radar signal maintains strength over longer distances, enabling deeper penetration [8,9]. Conversely, in highly conductive environments, such as saltwater or saturated concrete, the energy is quickly absorbed, significantly limiting the depth of penetration. This makes GPR particularly effective for evaluating materials such as concrete, asphalt, sand, and timber in construction contexts [8].

The dielectric constant of a material governs how fast radar energy moves through it. By transmitting EM waves into a material like concrete and measuring the time it takes for the reflections to return from subsurface features, GPR systems can calculate depth. This is achieved by multiplying the round-trip travel time by the wave velocity. In air, radar energy travels close to the speed of light, while in water, it slows to about one-ninth of that speed. Dielectric constants range from 1 for air to approximately 81 for water. For typical construction materials, the dielectric constant usually falls between 3 and 12, corresponding to wave velocities between 0.178 and 0.089 m per nanosecond [32].

When the transmitted signal encounters subsurface objects or interfaces between materials with contrasting dielectric constants or conductivities, such as steel reinforcement, voids, or transitions between different media, it reflects toward the surface and is detected by the receiving antenna. The greater the contrast in these electrical properties, the stronger the reflection. For instance, steel rebars embedded in concrete yield prominent reflections due to their high electrical conductivity.

The strength of the reflected signal, known as amplitude, is captured and stored in a two-dimensional data matrix, where the columns represent horizontal positions and rows correspond to depth. These matrices are visualized as radargrams using proprietary software that accompanies most commercial GPR systems. A typical GPR antenna system includes both a transmitter (T) and a receiver (R). The transmitter emits the radar pulses, while the receiver collects the signals that bounce back. In this study, a 1.6 GHz antenna with a transmitter–receiver offset of 58 mm was used.

4. Analytical Method

The shape of the hyperbola is a critical factor in GPR data processing. It is influenced by two parameters: (1) scan spacing, where shorter scan spacings (in/scan) result in wider parabolas, and (2) wave velocity, with lower dielectric constants (higher electromagnetic wave velocity) producing wider parabolas, and vice versa. Additionally, larger targets generate brighter reflections; however, the change in the hyperbola shape for different target sizes is nearly negligible for any diameter under two inches, as target sizes are a fraction of the electromagnetic wavelength. The theoretical equations of the hyperbola are derived based on the following assumptions [8]:

•

The positive peaks of the reflected waves from the rebar correspond to the negative peaks of the transmitted wave, indicating a phase reversal.

•

The transmitted waves reflect off the surface of the rebar along the shortest two-way travel path.

The length of the traveling path of the signal reflected from a rebar can be determined using the equations illustrated in Figure 2:

$$L_1=\sqrt{{\left(z\right)}^2+{\left(X-{\displaystyle \frac{S}{2}}\right)}^2}$$

(1a)

$$L_2=\sqrt{{\left(z\right)}^2+{\left(X+{\displaystyle \frac{S}{2}}\right)}^2}$$

(1b)

where X = x_p − x is the distance in the horizontal direction between the rebar and the center line of the transmitter and receiver (T-R); z is the depth of the rebar center; S is the spacing of the transmitter and receiver (T-R offset); x_p is the horizontal coordinate of the rebar; and x is horizontal coordinate of the antenna.

Considering the peak point of the hyperbola, the two-wave travel time is expressed by Equation (2) as:

$$t_p-t_0={\displaystyle \frac{\sqrt{{4z}^2+S^2}}{V_s}}$$

(2)

where t₀ is time zero and t_p is the time of the peak of the hyperbola. Considering a point on the hyperbola (t_i), the two-wave travel time can be obtained by employing Equation (3):

$$t_i{-t}_0={\displaystyle \frac{L_1+L_2}{V_s}}$$

(3)

Substituting Equations (1) and (2) into Equation (3) results in Equation (4):

$$t_i=\left(t_p-t_0\right){\displaystyle \frac{\sqrt{{\left(z\right)}^2+{\left(X-\frac{S}{2}\right)}^2}+\sqrt{{\left(z\right)}^2+{\left(X+\frac{S}{2}\right)}^2}}{\sqrt{{4z}^2+S^2}}}+t_0$$

(4)

Equation (4) represents the theoretical hyperbolic curve produced by a rebar embedded in reinforced concrete structures. However, the variations in hyperbolic curves corresponding to different rebar sizes are minimal.

The algorithm developed for this study processes GPR data to identify hyperbolic shapes in the radargram and extracts amplitude data for each hyperbolic peak. By comparing the amplitudes of these hyperbolic shapes with theoretical hyperbolas derived from Equation (4) and calculating the sum of squared differences (R²), the program can determine whether the hyperbolic shapes represent rebar. Users can adjust the acceptable sum of squared differences to enhance the accuracy of rebar detection by setting a predefined threshold (Q). This study limited the allowable difference between positive amplitude hyperbolic shapes from the GPR data and the theoretical curves to 80%. Once the algorithm locates the rebar, it can also determine the rebar’s depth and location.

5. Introduction to YOLOv8

The YOLOv8 model is an advanced object detection framework optimized for speed, accuracy, and adaptability. Unlike earlier YOLO versions that relied on anchor-based detection, YOLOv8 employs an anchor-free detection strategy, eliminating the need for predefined bounding boxes. This simplification enhances localization accuracy, reduces computational complexity, and enables the model to generalize more effectively to objects of varying shapes and sizes, such as hyperbolic patterns in GPR radargrams.

5.1. Backbone

The backbone of YOLOv8 is based on a Cross Stage Partial (CSPDarknet) network, which enhances feature extraction while reducing redundancy in gradient flow. It consists of multiple convolutional layers, residual blocks, and spatial pyramid pooling, enabling efficient learning of both low-level and high-level image features. For GPR imagery, this ensures effective extraction of hyperbolic signatures from noisy and cluttered radargrams [33,34].

5.2. Neck

The neck integrates features from multiple scales using a Path Aggregation Feature Pyramid Network (PAFPN). This structure enables the model to integrate semantic-rich, deeper features with high-resolution, shallow features, thereby enhancing the detection of both small and large hyperbolas representing rebars at various depths. PAFPN enhances contextual learning and robustness to interference from voids and overlapping reflections [33,34].

5.3. Head

The head of YOLOv8 handles both classification and regression tasks in a unified way, directly predicting object class probabilities and bounding box offsets for each detected feature region. Its anchor-free approach simplifies the prediction process by regressing object centers and sizes without the need for predefined anchor templates. This architecture is especially effective for rebar detection, where the shapes of hyperbolas can vary in scale and curvature based on depth and material properties. Once predictions are made, non-maximum suppression (NMS) is applied to eliminate redundant bounding boxes, ensuring that only the highest-confidence detection for each target is retained. This process guarantees that each rebar hyperbola is counted only once.

5.4. Loss Function

YOLOv8 uses a combination of advanced loss functions, including Complete Intersection over Union (CIoU) loss for bounding box regression and binary cross-entropy for classification. The YOLOv8 loss function is composed of three terms: (1) Bounding box regression loss, which uses Complete Intersection over Union (CIoU) to measure similarity between predicted and ground-truth boxes. (2) Objectness loss, which uses binary cross-entropy to determine if a grid cell contains an object. (3) Classification loss, which also uses binary cross-entropy to classify the object class.

5.5. Optimization

The model leverages Stochastic Gradient Descent (SGD) or AdamW optimizers with Automatic Mixed Precision (AMP), which accelerates training and reduces GPU memory requirements while maintaining numerical stability. The overall design ensures: (1) faster convergence with fewer parameters; (2) stable gradient flow due to CSPDarknet architecture; and (3) high detection accuracy under variable confidence thresholds, as confirmed by training curves (loss, mAP, F1–confidence).

5.6. Anchor-Free Detection

Traditional object detection networks use pre-defined bounding boxes (anchors) and learn to adjust them to fit objects. YOLOv8 employs an anchor-free approach, directly predicting object center points and sizes rather than relying on anchor templates. For each grid cell, YOLOv8 predicts: (1) objectness score; (2) class probabilities; (3) bounding box center coordinates; and (4) bounding box width and height. This approach reduces hyperparameters (by eliminating anchor tuning), simplifies training, and improves generalization for objects of irregular shape, such as hyperbolic reflections in GPR data.

5.7. Advantages for GPR Applications

The YOLOv8 architecture’s efficiency and accuracy enable real-time detection of rebar signatures in GPR images with minimal pre-processing. The integration of CSPDarknet, PAFPN, and anchor-free detection ensures strong generalization even when radargrams are affected by asphalt overlays or internal voids. Additionally, YOLOv8 supports easy model scaling and fast inference speeds, making it suitable for field deployment during structural inspections.

The complete architecture of YOLOv8, showing the CSPDarknet backbone, PAFPN neck, anchor-free detection head, and non-maximum suppression stage, is presented in Figure 3.

6. Experiment Program

6.1. Precast Prestressed Concrete Box Beam

The West Virginia Division of Highways (WVDOH) provided precast, prestressed concrete box beams, each measuring 32.5 feet in length, 3.0 feet in width, and 17 inches in depth, as illustrated in Figure 4. Originally part of a bridge in Cabell County, West Virginia, these beams were repurposed after the bridge was replaced, offering practical value for current research. Each beam is reinforced with four #4 top longitudinal bars and stirrups spaced 12 inches apart, along with two groups of five prestressed tendons located at the bottom. Additionally, a single prestressed tendon is positioned along the centerline of each beam. To reduce weight, each beam was designed with two 10-inch-diameter air voids. Figure 4 and Figure 5 depict the typical cross-sectional and plan views of these box beams.is also positioned along the centerline. To reduce weight, each beam includes two 10-inch-diameter air voids. Figure 4 and Figure 5 illustrate the typical cross-sectional and plan views of these box beams.

6.2. In-Service Prestressed Concrete Bridge Case Study

GPR testing was conducted on a bridge case study located in West Virginia, spanning a total length of 94 feet. The bridge is supported by twelve precast box beams, each measuring 36 inches in width and 42 inches in height. Within the hollow rectangular cross-section of each beam, four #4 longitudinal rebars are placed near the top, spaced at 8-inch intervals. Two layers of thirty prestressed tendons are positioned near the bottom face of the beam, with a minimum spacing of 2 inches between adjacent tendons.

The bridge deck is covered by an asphalt overlay ranging in thickness from 1.5 to 2.5 inches. However, the asphalt is in a state of deterioration, which may compromise the overall structural integrity and performance. Figure 6 highlights the surface defects observed on the bridge, underscoring the urgent need for maintenance to ensure continued safety and functionality.

7. Data Collection and Acquisition

The study involved conducting a comprehensive GPR survey, focusing on a precast prestressed concrete box beam. The study was structured to yield a detailed mapping of the internal reinforcement in box beams. The top surface of the box beam was scanned in both transverse and longitudinal directions, effectively capturing critical structural data.

Zatar et al. [9] examined how scan spacing affects the accuracy of measuring the depth and location of rebar. They found that transverse scan line spacings ranging from 4 inches to 24 inches yielded nearly identical results. Similarly, changes in the number of longitudinal scan lines had minimal impact on measuring stirrup depth and spacing.

In the present study, ninety-seven transverse scans were conducted at 4-inch intervals to ensure high spatial resolution and to minimize the risk of missing closely spaced reinforcement. Additionally, three longitudinal scans were performed, one along the beam’s centerline and two positioned 11 inches to either side, to provide comprehensive coverage of the reinforcement layout.

Data collection focused on the top surface because this location allowed unobstructed access to the GPR antenna, reduced safety risks compared to scanning the underside, and enabled consistent coupling between the antenna and the surface. Scans of the underside, intended for detailed prestressed tendon sizing, were postponed for subsequent analysis, as these measurements require higher-resolution imaging.

Analysis of the top scans yielded information about the dimensions and configurations of the top longitudinal bars and the shear reinforcement stirrups. These findings helped to understand the structural integrity and load-bearing capabilities of the beam. Figure 7 illustrates the detailed scan grid layout utilized for the box beam, providing a visual representation of the scanning strategy.

In tandem with the beam analysis, the actual bridge underwent a thorough survey using 95 transverse profile lines and 36 longitudinal lines, spaced 12 inches apart. The focus of the transverse profiles was primarily on the top longitudinal bars, while the longitudinal profiles were designed to capture detailed data concerning the stirrups.

The key scanning parameters utilized throughout this study, both during the training phase and the evaluation phase, are summarized in

Table 1. Testing parameters.

Parameters	Unit	Value	Description
SPN	-	2048	Samples per scan
SPD	-	Varies	Scan per second
SPM (Ns)	-	244–480	Scan per foot
Top position	in	3.3	Level of the antenna
Range	in	33	Scan depth
Scan spacing	in	0.025	Scan spacing in a scan line
Antenna frequency	GHz	1.6
T-R offset	in	2.28	Distance between transmitter and receiver

. These parameters are critically important as they directly influence model accuracy and the overall interpretation of the structural conditions observed. This methodical approach exemplifies the importance of precise data collection and analysis in ensuring the safety and reliability of concrete structures.

8. AI Training Models

The dataset used to train and evaluate the YOLOv8 model was obtained from three laboratory-scale bridge box beams and one in-service bridge structure. These specimens featured asphalt overlays with thicknesses ranging from 1.5 to 2.5 inches, which are typical of standard construction practices and environmental protection requirements. The inclusion of varying asphalt conditions allowed for a comprehensive assessment of how overlay thickness affects the visibility and detectability of embedded rebar in GPR radargrams.

Meeting the input requirements of YOLOv8 and enhancing model performance involved preprocessing radargrams into image segments with a resolution of 640 × 640 pixels, a size that strikes a balance between computational efficiency and sufficient visual detail to detect hyperbolic signatures accurately. Overlapping image patches were extracted from each radargram to ensure complete coverage and to prevent truncation of hyperbolas located near image edges.

The rebar detection methodology developed by Zatar et al. [9] was essential in accurately identifying multiple hyperbolas within the radargrams. Their approach demonstrated high accuracy in controlled laboratory environments without asphalt overlays, providing a foundational reference for this study. However, detecting rebars beneath asphalt layers presents additional challenges due to signal attenuation and the reduced clarity of hyperbola signatures. To tackle these challenges, a hybrid approach was adopted, combining automated detection with manual verification to ensure greater reliability under in-service bridge conditions. The refined detection results were subsequently used to define bounding box locations for preparing the training dataset.

A hybrid Delphi–Python workflow was implemented to execute the process efficiently. The Delphi program automated the generation of bounding boxes for each annotated hyperbola, ensuring that they were enclosed entirely while producing overlapping cropped image patches (default size: 640 × 640 pixels) to prevent any features from being cut off at the edges. In this context, the sliding window approach involves moving a fixed-size rectangular window step by step across the radargram, cropping each region into an image patch for YOLOv8 processing. Overlaps between adjacent windows are incorporated so that hyperbolas spanning window boundaries are fully captured in at least one patch. Additionally, the program generated YOLO-compatible label files, where the bounding box center coordinates (x_c, y_c) and dimensions (w, h) were normalized relative to the image dimensions (W, H) according to:

$$x_n={\displaystyle \frac{x_c}{W}};y_n={\displaystyle \frac{y_c}{H}};w_n={\displaystyle \frac{w}{W}};h_n={\displaystyle \frac{h}{H}};$$

(5)

The processed dataset was then passed to a Python 3.9 training script utilizing the Ultralytics YOLOv8 framework. This script handled model initialization, hyperparameter configuration (epochs, batch size, learning rate, confidence threshold, and IoU threshold), and Automatic Mixed Precision (AMP) to speed up training and reduce GPU memory usage. It trained the model on the prepared dataset, validated performance on a separate validation set, and output key metrics, including precision, recall, and mean Average Precision (mAP). The training process also generated performance curves, loss vs. epoch, mAP vs. epoch, and F1–confidence curves, to monitor convergence and detection quality.

Improving the effectiveness and reliability of hyperbola detection in GPR imagery requires categorizing the data into three distinct groups based on quality: clear, interfering, and blurry. These categories reflect the visibility and interpretability of hyperbolic signatures in the radargrams, which is crucial for tailoring model training strategies to different data conditions.

Clear data refers to radargrams with well-defined hyperbolas, typically collected from stirrups under laboratory conditions with minimal interference. Understanding these classifications is crucial for addressing detection challenges and improving model robustness across varying data qualities. Interfering data arises when hyperbolas from stirrups are distorted by reflections from nearby main rebars, often leading to overlapping or noisy patterns. This issue is further complicated by internal voids within the box beam cross-section, which disrupt signal propagation and contribute to reflection artifacts. The clear and interfering datasets exhibited similar signal attenuation because they were both collected from beams without asphalt overlays, with the distinction arising from interference caused by internal voids in the interfering class. Blurry data is characterized by weak or poorly defined hyperbolas, primarily due to signal attenuation from asphalt overlays. These conditions are common in field-collected radargrams from in-service bridges with asphalt covers ranging from 1.5 to 2.5 inches. The blurred dataset originated from a single asphalt-covered bridge beam, and since no additional asphalt-covered datasets were available, no further sub-categorization of blur severity was performed.

To ensure balanced and representative learning, the authors produced a dataset composition and training data selection where training data for each category was selectively obtained:

-

Precise data was collected from Beam #1 and Beam #2.

-

Interfering data was obtained from Beam #2 and Beam #3.

-

Blurry data was obtained from field radargrams captured on the in-service bridge.

Table 2. Training data.

Data Category	Number of Training Images	Epoch	conf	IoU	Batch Size	Learning Rate	Optimizer
Clear data	85	50	0.3	0.5	8	0.01	SGD
Interfering data	371	150	0.3	0.5	2	0.01	SGD
Blurry data	1036	300	0.25	0.45	2	0.01	SGD

provides a summary of the number of images and key training parameters used for each data group. In all cases, images were standardized to a resolution of 640 × 640 pixels, and the YOLOv8s model was utilized with AMP enabled to optimize training performance. To improve model generalization, standard data augmentation techniques were applied, including color jittering, horizontal flipping, scaling, and random erasing.

The training strategy employed was tailored to the specific characteristics of each data group, considering the number of images, potential model challenges, and the application of data augmentation. Each category serves to enhance the model’s ability to generalize across different image qualities, which is crucial for real-world civil and transportation infrastructure applications where image clarity can significantly vary.

The model was trained on a relatively small dataset of clear images for 50 epochs. With a confidence threshold (conf) of 0.3 and an Intersection over Union (IoU) threshold of 0.5, the model aimed to strike a balance between sensitivity and specificity. A batch size of 8 and a learning rate of 0.01 were employed with the SGD optimizer. Despite the limited number of training images, the use of standard data augmentation techniques can help enhance the model’s robustness, enabling it to generalize beyond the training data.

This dataset, comprising a larger number of 371 interfering datasets, was trained over 150 epochs. The model maintained the same confidence and IoU thresholds as in the clear data category. With a batch size of 2, the complexity of the data likely required more careful tuning during training. The considerable size of this dataset suggests that the model would encounter diverse scenarios of interference, thus providing it with a broader exposure to variations in the object detection task.

The training on blurry images, with the highest count of 1036, was conducted over the longest duration of 300 epochs. Here, the model’s confidence threshold was slightly reduced to 0.25 and IoU to 0.45. The smaller batch size of 2 could be attributed to the additional effort required by the model to learn from the more challenging blurry images. The extensive training duration suggests an emphasis on adapting to this data’s unique challenges, with augmentations further assisting the model in identifying blurred objects.

To evaluate the accuracy of hyperbola detection, the F1 score was employed. The F1 score is defined as the harmonic mean of precision and recall, providing a balanced measure of a model’s ability to accurately identify targets while minimizing both false positives and false negatives. This makes it particularly suitable for GPR hyperbola detection tasks, where both missed detections and false alarms can impact the reliability of the results. The F1 score is calculated as:

$$F1=2\times (Precision\times Recall)/(Precision+Recall)$$

(6)

Figure 8 presents the F1 score curves for all training datasets. For clear hyperbolas, peak accuracy is achieved at a confidence threshold of 0.515, indicating that well-defined signals require higher confidence levels for optimal detection. Interfering hyperbolas achieve balanced performance at a confidence threshold of 0.351, indicating that moderate filtering is most effective in mitigating overlapping reflections. For blurry hyperbolas, maximum performance occurs within a confidence threshold range of 0 to 0.25, implying that a lower threshold is necessary to preserve weaker detections affected by asphalt-induced attenuation.

Figure 9 presents the YOLOv8 training and validation curves for clear, interfering, and blurry datasets, showing stable convergence in loss, precision, recall, and mAP@0.5 values. Clear data achieved optimal performance fastest (~50 epochs), while interfering and blurry data required more epochs due to increased complexity. Figure 10 shows the corresponding precision–recall curves, where the clear dataset achieved the highest AUC-PR, followed by interfering and blurry data. These results confirm the model’s robustness while illustrating the greater detection challenge in asphalt-covered conditions.

The trained YOLOv8s model was evaluated for its inference speed on a workstation equipped with an NVIDIA GeForce GTX 1050 Ti GPU and an Intel Core i7 CPU. The model achieved an average inference time of approximately 40 milliseconds per image, resulting in a throughput of about 25 frames per second (FPS). This performance meets the practical requirements for near-real-time deployment in the field, allowing for rapid interpretation of GPR radargrams during inspections. Such speed enables inspectors to make informed decisions promptly without delaying the inspection process. Additionally, further optimization, such as deploying the model on a higher-performance GPU or utilizing acceleration frameworks like NVIDIA TensorRT, could enhance processing speed even more, making it suitable for large-scale bridge assessments.

The results in

Table 3. Performance and computational complexity comparison between YOLOv5 and YOLOv8 across the three GPR datasets.

Model	YOLOv5			YOLOv8
Dataset	Clear	Interfering	Blurry	Clear	Interfering	Blurry
Precision	0.958	0.963	0.998	0.955	0.933	0.909
Recall	0.988	1	1	1	1	0.955
mAP@0.5	0.992	0.994	0.995	0.992	0.981	0.984
mAP@0.5:0.95	0.659	0.796	0.977	0.749	0.809	0.895
IoU	0.992	0.994	0.995	0.992	0.981	0.984
Params (M)	7.01	7.01	7.01	3.01	11.13	11.13
MACs (G)	7.88	7.88	7.88	4.04	14.22	14.22
GFLOPs	15.76	15.76	15.76	8.09	28.44	28.44
Inference Time (ms/img)	18.22	18.18	18.15	12.1	28.19	28.19
Peak GPU Mem (MB)	79.6	79.6	79.6	35.8	80.9	80.9

show that both YOLOv5 and YOLOv8 achieve high detection accuracy across the three datasets, with precision, recall, and mAP@0.5 values exceeding 0.93 in all cases. YOLOv5 generally achieves slightly higher precision, while YOLOv8 occasionally attains better mAP@0.5–0.95, particularly for the clear and interfering datasets. In terms of computational complexity, YOLOv8 in its small configuration (3.01M parameters for the clear dataset) demonstrates lower GFLOPs, faster inference times, and reduced GPU memory usage compared to YOLOv5, making it more suitable for resource-constrained deployments. However, the larger YOLOv8 configurations (used for interfering and blurry datasets) increase computational cost, leading to slower inference and higher memory requirements compared to YOLOv5. Overall, the comparison confirms that YOLOv8 can match or exceed YOLOv5 in accuracy while offering potential efficiency advantages depending on the chosen model size, enabling flexible trade-offs between accuracy and computational demands for different field conditions.

9. Hyperbola Detection Method

A custom Python-based detection framework was developed to automatically identify hyperbolic signatures in GPR radargrams using the YOLOv8 deep learning model. This method integrates modern object detection techniques with domain-specific post-processing heuristics to enhance robustness in complex subsurface imaging scenarios. The overall pipeline consists of several key stages: image preprocessing, sliding window detection, confidence-based filtering, non-maximum suppression, geometric refinement, and result annotation. Each component is designed to address specific challenges, such as large image sizes, overlapping detections, and noisy radar reflections.

9.1. Image Preprocessing and Sliding Window Detection

Given the high resolution and extended dimensions of typical radargrams, direct full-image detection is computationally inefficient and prone to errors related to boundaries. To address this, each input radargram is divided into overlapping patches using a sliding window approach. This ensures full spatial coverage and minimizes the risk of cutting through hyperbolas located near image edges. Each image section, typically sized 640 × 1280 pixels, is passed through the trained YOLOv8 model to identify potential hyperbolic patterns. Overlapping is essential to preserve contextual integrity between adjacent windows and to ensure that no significant feature is missed due to artificial segmentation.

9.2. Bounding Box Extraction and Confidence Filtering

The YOLOv8 model returns a set of bounding boxes for each window, each of which is associated with a predicted confidence score. Detections with confidence values below a predefined threshold (e.g., 0.18) are filtered out to suppress weak or spurious predictions. To preserve spatial accuracy, the bounding box coordinates from each cropped section are recalculated relative to the original complete radargram coordinate system. This filtering step reduces noise in the detection output and helps prioritize more reliable predictions during subsequent processing.

9.3. Non-Maximum Suppression (IoU Filtering)

Due to overlapping regions in the sliding window approach, the same hyperbola may be detected multiple times. To eliminate these duplicates, the algorithm applies non-maximum suppression (NMS) based on the IoU metric. For every newly proposed bounding box, the IoU is calculated against all retained boxes. If the IoU exceeds a user-defined threshold (e.g., 0.4), only the bounding box with the highest confidence score is preserved. This mechanism ensures that each hyperbola is represented by a single, most confident bounding box, minimizing redundancy in the detection output.

9.4. Geometric Consistency Filtering

To further enhance precision, especially in high-clutter or low-signal environments, a secondary geometric consistency rule is applied. If the center-top point of a bounding box lies within the vertical extent of another detected region, the overlapping box is flagged for removal. This heuristic targets cases of partial or fragmented detections, often caused by weak reflections or multipath effects, and promotes spatial coherence in the results. By incorporating domain-specific geometric logic, the system achieves better discrimination between genuine hyperbolas and noise artifacts that may share similar visual characteristics.

9.5. Result Annotation and Output

The final set of filtered bounding boxes is superimposed on the original radargram for visual inspection and verification. Detection metadata, including coordinates, dimensions, and confidence scores, is exported in structured text files for quantitative evaluation. Additionally, the number of hyperbolas detected per image is recorded to support statistical performance analysis and algorithm benchmarking.

This detection framework effectively balances the strengths of deep learning with GPR-specific interpretive rules, providing a robust and scalable solution for automated rebar detection and subsurface anomaly identification. It is beneficial in large-scale structural health monitoring, where manual annotation is time-consuming, error-prone, and impractical.

10. Results and Discussions

Figure 11 shows the detected hyperbolas using an interfering trained model for no interference (left) and high interference (right) of the voids. Despite the presence of strong horizontal reflections and diffractions caused by voids near the surface, the model successfully detects all four hyperbolic signatures representing rebars. This indicates YOLOv8’s ability to distinguish structural hyperbolas from background noise and non-rebar reflections, which often confuse traditional analytical approaches. Each bounding box is well-aligned with the apex of the hyperbolas, demonstrating high localization precision, even where signal distortion is evident, especially in the second and third hyperbolas.

Figure 12 presents a comparison between the YOLOv8 detection method and the analytical method for detecting rebars in 97 transverse radargrams collected from prestressed concrete beam #1. Each radargram contains four rebars at known locations, serving as ground truth for evaluation. The YOLOv8 detection method demonstrated significantly more consistent and accurate performance. It correctly detected exactly four rebars in 75% of the scans while over-predicting in 7% and under-predicting in 18% of the cases. In contrast, the analytical method correctly identified four rebars in only 41% of scans, overpredicting in 32% and underpredicting in 27% of the cases. The concentration of errors at the beginning of the beam can be attributed to the dense arrangement of stirrups in this region, with a spacing of only 2 inches to resist high shear forces. This close spacing produces overlapping reflections from both the stirrups and the main longitudinal rebars, which distorts the hyperbolic signatures in the radargrams. As a result, the reflections are complex for the detection model to resolve, leading to reduced accuracy in this localized area.

The YOLOv8’s lower false positive rate and higher consistency indicate that it is more robust against interference and signal distortion, particularly in regions affected by overlapping hyperbolas or noise. By learning from a diverse training set, the YOLOv8 model generalizes well to complex patterns that often confound traditional curve-fitting approaches. These results further support the superiority of the AI-based method for reliable automated rebar detection in real-world concrete structures.

Figure 13 presents a radargram of the in-service bridge from the longitudinal direction, with rebar locations marked. The hyperbolas shown represent rebar detections obtained using both the analytical and YOLOv8 detection methods. A comparison between the YOLOv8 detection and the analytical method for identifying hyperbolas corresponding to stirrups in a field scan is shown in Figure 14. The data were collected from a reinforced concrete bridge with a 2- to 3-inch asphalt cover, using a clear-data-trained model for YOLOv8 inference. Despite the attenuation and dispersion caused by the asphalt overlay, the YOLOv8 model was able to closely follow the rebar reflection pattern with a high degree of consistency.

The analytical method, based on curve-fitting of hyperbolic signatures, demonstrated reasonable performance but exhibited greater fluctuation, particularly at lower amplitudes or in the presence of distorted reflections. Quantitatively, the YOLOv8 model detected all but one stirrup, while the analytical method missed four. Additionally, the time-of-flight (TOF) estimates from the YOLOv8 predictions align more tightly along the trendline, indicating greater stability and robustness across varying scan traces. This comparison highlights the effectiveness of the YOLOv8 model in detecting subsurface objects, even when signal quality is degraded. Its performance suggests strong potential for field deployment without requiring retraining or tuning, provided the model has been exposed to diverse training conditions.

Figure 15 illustrates the detected hyperbolas using the blurry-trained model applied to transverse GPR scans of the in-service bridge. The reflection signals from the main rebars in these radargrams were highly distorted and significantly attenuated due to the presence of a thick asphalt overlay and concrete cover. A filtering algorithm was applied to the initial YOLOv8 hyperbola detections to reduce false positives and improve detection accuracy.

This algorithm removed redundant or inaccurate bounding boxes based on spatial overlap and position criteria. Before filtering, the model detected approximately 40 hyperbolas per image, significantly overestimating the actual count. After applying the filtering logic, which eliminates boxes whose top-center point lies within a more confident detection, only 24 hyperbolas remained, closely aligning with the expected number of rebars. This post-processing step proved essential for refining detection results, especially in cases with overlapping or noisy signals.

Figure 16 illustrates the performance of the YOLOv8 model in detecting rebars from blurry radargrams obtained from the field sections, which exhibit weak signal quality due to a thick asphalt layer. The analytical method was excluded from this evaluation due to its inability to detect hyperbolas under such degraded conditions. Two curves are shown: the initial detection (solid thin line) and the filtered detection (dashed line), with the actual number of rebars (22 per section) shown as a reference (bold horizontal line).

The initial detection was consistently overestimated, with an average of 30 detections per section, primarily due to overlapping or duplicate bounding boxes around low-contrast hyperbolas. After applying the filtering algorithm, which removes redundant detections based on geometric and confidence criteria, the average detection was reduced to 21, closely approximating the actual number of rebars. This result underscores the significance of post-processing in refining YOLOv8 outputs and minimizing false positives, particularly in low-visibility conditions. The filtered YOLOv8 predictions provide a reliable estimate of rebar count in blurry GPR images, where traditional methods fail.

The study results show that the filtering method significantly reduces false positives while preserving accurate rebar detections. It highlights the influence of tailored model training and robust post-processing in achieving reliable performances. The findings will inform recommendations for selecting or combining training datasets when preparing models for field deployment. Ultimately, this work contributes to improving the reliability and flexibility of AI-assisted rebar detection in diverse inspection scenarios.

The output from the YOLOv8 detection framework provides precise spatial coordinates of the detected rebars. These measurements are directly translated into engineering drawings that map both longitudinal and transverse reinforcement layouts. By accurately determining rebar spacing and location, the generated drawings not only document existing reinforcement configurations but also serve as a basis for evaluating structural capacity. This information enables engineers to assess the load-carrying performance of the beam, identify potential deficiencies, and plan targeted maintenance or rehabilitation measures, ensuring long-term structural integrity.

11. Conclusions

This study demonstrates the effectiveness of using deep learning, specifically YOLOv8, for detecting hyperbolic signatures of reinforcing bars in GPR images under varying field conditions. Three distinct models were trained on clear, interfering, and blurry datasets, and their cross-performance was evaluated. While initial detections often resulted in redundant bounding boxes, a filtering algorithm based on the spatial overlap and confidence level effectively reduced false positives and brought the number of predicted rebars closer to the ground truth.

Comparison with traditional analytical methods confirmed that the YOLOv8 detection approach is more robust, especially in low-visibility or noisy scenarios where analytical techniques failed to identify weak hyperbolic patterns. In transverse scans of bridge decks with 1.5–2.5-inch asphalt cover in West Virginia, the YOLOv8 model achieved superior accuracy, detecting nearly all embedded stirrups while minimizing over-detection after filtering. The findings emphasize that while YOLO-based models are powerful, their reliability heavily depends on the quality and context of the training data. Tailored datasets and post-processing steps are essential for achieving high performance in practical structural assessment. This study offers a scalable framework for deploying AI-enhanced GPR analysis in several real-world civil and transportation infrastructure applications.

While the proposed YOLOv8-based approach achieved high accuracy across clear, interfering, and blurry datasets, the current evaluation did not include scenarios with double-layer reinforcement or extremely dense rebar arrangements. Such configurations can produce overlapping or merged hyperbolic signatures, increasing detection difficulty. Additionally, the dataset was limited to bridge beams with asphalt overlays up to 2.5 inches in thickness; results under thicker overlays or different pavement materials may vary. Future work will extend the dataset to include multi-layer and high-density rebar arrangements, as well as diverse environmental and structural conditions, to further evaluate and enhance model robustness.