2.1 Data collection and transformation

To address the limitations associated with the scarcity and imprecise annotation of existing GPR data, we constructed a proprietary dataset (subData) comprising 4,274 accurately annotated defect images. These images were collected from over 40 km of asphalt pavements across Zhejiang Province, China, encompassing diverse scenarios—including bridges, rural roads, and expressways—under varying seasonal conditions and pavement service lives. This variability ensures the inclusion of challenging detection conditions, characterized by fluctuating defect severity, varying spatial density, and diverse background noise levels, thereby substantially enhancing the dataset's generalizability.

The empirical field data were acquired using the two-dimensional ground-penetrating radar (GS8000, Proceq SA, Schwerzenbach, Switzerland). The system operated in free-scan mode with a sampling frequency of 1,500 Hz, an effective bandwidth of 3,200 MHz, offering a minimum detectable target size of 0.1 cm and a maximum penetration depth of 10 m. The sampling rate was set to 500 Hz, with a Global Navigation Satellite System (GNSS) initialization time of 15 s, and a spatial sampling interval of 100 scans m^-1. Detailed radar configuration parameters are provided in Table 1.

Field data were collected on roads in Zhejiang Province using the GS8000 system initialized with the aforementioned parameters. Considering an average road width of 12 m, the lateral spacing between adjacent survey lines was established at 2 m. The radar system traversed multiple 5 km road sections bidirectionally at this interval, yielding six continuous 5,000 m B-scan profiles. Subsequently, image segments with a depth range of 0–1.5 m and a horizontal length of 2.5 m were extracted for analysis. The data acquisition protocol is schematically illustrated in Figure 1A.

Figure 1 Experimental dataset construction process. (A) Field data acquisition protocol; (B) data preprocessing interface; (C) noise suppression procedure; (D) contrast enhancement procedure; (E) noise addition procedure

Raw GPR data inherently consist of electromagnetic reflections that capture both salient target information and substantial noise arising from various subsurface heterogeneities. To facilitate downstream analysis, specific signal processing filters were applied to suppress clutter and enhance target features. The data preprocessing workflow, implemented using Proceq software, is illustrated in Figure 1B.

Noise suppression mitigates background clutter through filtering or signal reconstruction. Specifically, redundant deep-layer textures are analyzed to isolate and eliminate statistically insignificant noise artifacts, and smoothing or interpolation is applied to ensure data continuity in processed regions. This process enhances signal coherence and accentuates reflections from subsurface targets (Figure 1C).

Contrast enhancement accentuates target features by optimizing the grayscale dynamic range. By analyzing the global gray-level distribution, the contrast is stretched or modified to amplify the intensity disparity in target regions. This clarifies subsurface reflection patterns and facilitates the interpretation of subtle structural variations (Figure 1D).

Noise injection augments the dataset by introducing simulated interference. The signal characteristics of the original GPR data are analyzed, and synthetic noise—generated according to a predefined noise model (e.g., Gaussian)— is superimposed onto the original B-scans. This simulates random interference typical of complex detection environments, thereby enhancing model robustness (Figure 1E).

The three operations described above were applied to expand the empirical dataset. In accordance with GPR technical guidelines and defect manifestations in actual engineering scenarios, four defect categories—void, cavity, loosening, and water-rich—were annotated. We implemented a semi-automated annotation workflow using X-AnyLabeling software [14]. Initially, a manually annotated subset was used to pretrain a detection model, which generated candidate bounding boxes for unlabelled images; these candidates were subsequently corrected and verified by domain experts. This process yielded 3,274 accurately annotated images. The distribution of defect types is summarized in Table 2.

To ensure robust feature learning and compensate for the scarcity of water-rich defects in empirical datasets, gprMax simulations were generated comprising 200 instances each of void, cavity, and loosening defects, alongside 400 instances of water-rich defects. In accordance with the Technical Standard for Integrated Detection and Risk Assessment of Urban Subsurface Defects (JGJ/T 437-2018) [15], the road profile was modelled as a four-layers stratified structure—consisting of air, pavement, composite, and subgrade layers—with thicknesses of 0.05, 0.15, 0.35, and 1.00 m, respectively.

The specific material properties and corresponding electromagnetic parameters employed in the simulation model are summarized in Table 3.

To mitigate numerical dispersion, the spatial resolution ∆L must be rigorously defined. According to established empirical criteria, ∆L must be no larger than one-tenth of the minimum wavelength of the propagating electromagnetic field, as shown in Eqaution (1). Adherence to this constraint ensures the accurate representation of high-frequency waveform components, thereby effectively minimizing numerical dispersion and enhancing simulation fidelity.$\text{Δ}L\le \frac{\lambda }{10}$(1)

The minimum wavelength of the electromagnetic field is governed by the relative permittivity of the propagation medium and the signal frequency. To ensure a conservative estimate, the maximum frequency is defined as three times the central frequency, i.e., f_m = 2,700 MHz. The corresponding minimum wavelength is calculated via Equation (2):$\lambda =\frac{c}{f_{\text{m}}\sqrt{{\epsilon }_{\text{r}}}}$(2)where, c denotes the speed of light (3 × 10⁸ m s^-1), f_m represents the maximum frequency, and ε_r is the relative permittivity.

Based on the calculated minimum wavelength, the theoretical upper bound for the spatial resolution is defined. However, to rigorously suppress the numerical dispersion inherent in the Finite-Difference Time-Domain (FDTD) algorithm and minimize staircase approximation errors at curved material interfaces, we adopted a more stringent grid size of ∆L = 0.002 m. This high-resolution discretization is critical for capturing the high-frequency diffraction edges of early-stage voids and subtle cracks, ensuring that the simulated waveforms faithfully represent the complex geometric boundaries of real-world defects. The selection of the time window must account for the travel time of the electromagnetic wave from emission, through the medium, to the defect and back, as given by:$T=\frac{2\times h}{v}$(3)where, h denotes the model layer thickness (m) and v represents the wave velocity. v is calculated as follows:$v=\frac{c}{\sqrt{{\epsilon }_{\text{r}}{\mu }_{\text{r}}}}$(4)where, μ_r is the relative permeability.

The theoretical time window required for a round-trip travel to the maximum depth is calculated as 36 ns using Equation (3). However, given the heterogeneous nature of subsurface media, we extended the simulation window to 45 ns. This expansion is physically justified by two key factors. First, Water-rich defects possess a high relative permittivity, which significantly retards the electromagnetic wave velocity, increasing the travel time. Second, irregular defect geometries (e.g., cavity) often induce multiple scattering and resonance, generating prolonged signal tails. The adopted 45 ns window provides a sufficient physical margin to capture these delayed scattering signatures fully, preventing signal truncation at the bottom of the B-scan image. The specific parameters for road defect simulations are listed in Table 4.

Forward simulations of four defect types—void, cavity, loosening, and water-rich—were conducted using gprMax. Void typically occur in the subgrade, with small horizontal and large vertical extent, filled with air. Cavity occurs at the interface of the composite and subgrade layers, forming elongated rectangular shapes, also filled with air. Loosening appears in the subgrade with large horizontal and vertical dimensions, represented by a new medium with a dielectric constant of 3.5. Water-rich defects, generally in the subgrade, exhibit high moisture content, low strength, and high porosity, filled with water. Simulations were configured according to the dielectric constants in Table 4. Schematic representations of the four defect types are shown in Figure 2.

Figure 2 Schematic models of four defect categories. (A) void; (B) cavity; (C) loosening; (D) water-rich

To ensure the fidelity and reliability of the virtual dataset, a rigorous quality control protocol was implemented. First, the simulation parameters (dielectric constants and conductivity) were strictly calibrated against standard engineering guidelines to reflect realistic subsurface properties. Second, to bridge the domain gap between idealized simulations and noisy field data, we applied a noise injection strategy (Figure 1E) to the generated B-scans, simulating complex environmental interference. Finally, the synthetic images underwent a strict manual validation process to confirm that the morphological features (e.g., hyperbolic diffraction tails) were physically consistent with empirical observations, ensuring that the generated samples faithfully represent real-world defect characteristics.

The empirical GS8000 dataset and the gprMax simulated data were combined and randomized to create the experimental subData dataset, ensuring a balanced representation across the four defect types and their diverse manifestations. To prevent insufficient feature learning for underrepresented defects, the dataset was re-annotated using the automated procedure and partitioned into training, validation, and test sets in a 7: 2: 1 ratio (2,992, 855, and 427 images, respectively), in accordance with standardized machine learning practices. Representative samples from the subData dataset are shown in Figure 3.

Figure 3 Representative samples from the subData dataset. (A) Real data acquired by GS8000; (B) synthetic data generated by gprMax

3.1 Overall architecture

The YOLOv11 architecture represents a high-accuracy, efficient evolution of the YOLO series, offering significant improvements over its predecessor, YOLOv8 [16]. Its structural design comprises three primary components: the backbone, neck, and detection head. In both the backbone and neck, YOLOv11 replaces the original C2f modules of YOLOv8 with C3k2 modules, which employ dual convolutional kernels to achieve finer feature-map partitioning, thereby optimizing both representation quality and extraction speed [17]. Furthermore, a C2PSA refinement layer is integrated following the fast spatial pyramid pooling (SPPF) module. This layer combines multiple C2f modules with polarized self-attention to map features into a high-dimensional space, facilitating the capture of complex nonlinear relationships and enhancing overall feature representation.

As illustrated in Figure 4, the proposed MARD-Net is built upon the YOLOv11n architecture and is specifically optimized for real-world subsurface road defect detection. It addresses critical challenges such as deployment on resource-constrained edge devices, low detection accuracy for visually similar defects, and the difficulty of extracting robust features from cluttered backgrounds. MARD-Net incorporates three key architectural innovations. (1) LADH: to mitigate the computational overhead and feature redundancy inherent in conventional YOLO detection heads, LADH decouples the regression and classification tasks via asymmetric branches. The regression branch stacks three depthwise separable convolutions (DSConv) to efficiently capture defect edges and fine spatial details, while the classification branch employs two standard convolutions to maintain global semantic consistency. This design optimizes the balance between computational complexity and detection accuracy, particularly enhancing robustness for small or irregularly shaped subsurface defects. (2) RepNCSPELAN-enhanced backbone: to accommodate the high variability in scale, texture, and morphology of subsurface defects, the backbone integrates the RepNCSPELAN4_CAA module. By leveraging re-parameterization and cross-stage fea-ture aggregation, this module overcomes the hierarchical limitations of conventional convolutions. RepNCSPELAN4_CAA partitions feature maps into parallel branches for independent refinement and subsequent fusion, thereby expanding the effective receptive field. This improves the modeling of complex subsurface structures, enabling the precise characterization of subtle cracks and Void. (3) MPCA: given that defect regions often occupy minimal portions of road images while background pixels dominate, standard feature extraction is prone to spatial blurring and class imbalance. MPCA addresses this by applying average pooling along global, horizontal, and vertical dimensions, combined with channel weighting and adaptive fusion to dynamically enhance multi-scale features. This mechanism selectively strengthens network responses to defect regions while suppressing redundant background noise, enabling precise attention and discrimination under complex pavement conditions.

Figure 4 Schematic diagram of the MARD-Net architecture. The framework is visually delineated into backbone (yellow background), neck (pink background), and head (cyan background) regions. In the LADH module, the regression and classification branches are distinguished by cool and warm color schemes, respectively

3.2 Multi-path coordinate attention

The MPCA module is designed to simultaneously capture global and local information from feature maps, thereby enhancing their representational capability [18]. To effectively embed positional information into channel attention, MPCA executes two operations: coordinate embedding and coordinate attention generation.

3.2.1 Coordinate embedding

As illustrated in Figure 5, coordinate information is embedded within the MPCA module via horizontal and vertical average pooling, which facilitates the capture of precise positional cues. Simultaneously, global average pooling compresses spatial information into channel descriptors, to preserve global feature context. The MPCA module transforms the global average pooling result into z_c:$z_c=\frac{1}{H\times W}{\displaystyle \sum _{i=1}^H}{\displaystyle \sum _{j=1}^W}{x}_c(i,j)$(5)where x_c(i, j) represents the input feature map at channel c, and H and W denote the height and width of the feature map, respectively.

Figure 5 Schematic diagram of the MPCA architecture

A one-to-one feature encoding scheme is established. Given input x_c(i, j), pooling kernels of size 1 × W and H × 1 are employed to encode each channel along horizontal and vertical coordinates, respectively. By applying these pooling operations, the output for channel c at height h is expressed as:$z_c^h\left(h\right)=\frac{1}{W}{\displaystyle \sum _{0\le j\le W}}{x}_c\left(h,i\right)$(6)

Similarly, the output of channel c with width W is expressed as:$z_c^w\left(w\right)=\frac{1}{H}{\displaystyle \sum _{0\le j\le H}}{x}_{\text{c}}\left(j,w\right)$(7)The three pooling operations correspond to Figure 5: horizontal average pooling(X_avg), vertical average pooling(Y_avg), and global average pooling(P_avg). X_avg and Y_avg pooling operations aggregate features along the spatial dimensions, yielding two feature maps with distinct awareness. Meanwhile, P_avg extracts global statistics for each channel. Together, these operations preserve precise positional information while capturing global features, enabling the network to localize targets more accurately.

3.3 Lightweight asymmetric detection head

Given the complexity of subsurface road environments and the high computational demands of real-time processing, a LADH was developed to facilitate edge deployment. Empirical experiments demonstrate that LADH not only improves defect detection accuracy but also significantly reduces computational cost [19]. As illustrated in Figure 6, the architecture employs an asymmetric design: the regression branch replaces two standard 3 × 3 convolutions with three consecutive depthwise separable convolutions followed by an intermediate 1 × 1 standard convolution; conversely, the classification branch substitutes two depthwise separable convolutions with two 1 × 1 standard convolutions to streamline feature integration.

Figure 6 Schematic diagram of LADH

The regression branch employs a stack of three consecutive DSConv layers, effectively enhancing local feature modeling. This design expands the original 3 × 3 receptive field to 7 × 7, thereby improving fine-grained feature representation and enabling precise localization of large-scale defects such as water-rich and void. As illustrated, each DSConv unit consists of a depthwise convolution (DWConv) followed by a pointwise convolution (PWConv). To accurately reflect the sequential nature of this operation, the mathematical formulation is defined as:$F_{out}={\text{PWConv}}_{1\times 1}\left({\text{DWConv}}_{3\times 3}\left(F_{\text{in}}\right)\right)$(12)where, F_in and F_out denote the input and output feature maps, respectively; DWConv represents depthwise convolution, and PWConv denotes pointwise convolution.

Equation (12) formalizes the core operation of the regression branch. By stacking three consecutive DSConv layers, the effective receptive field (ERF) is mathematically expanded from 3 × 3 to 7 × 7. This expansion is critical for the physical interpretation of GPR B-scans. Subsurface defects like void manifest as hyperbolic diffraction curves with extended tails. The expanded 7 × 7 receptive field enables the LADH to cover these complete geometric signatures in a single forward pass, preventing the misinterpretation of partial hyperbolic arcs as noise, which frequently occurs with standard limited-view kernels.

Stacking multiple DSConv layers enhances the model's nonlinear representation and spatial feature extraction at a relatively low computational cost. To further improve cross-channel interaction, a 1 × 1 standard convolution is integrated after the DSConv cascade, effectively aggregating channel information and suppressing background noise.

The rationale behind extending the receptive field to 7 × 7 via stacked DSConvs is deeply rooted in the physical propagation characteristics of GPR signals. Unlike optical images where objects possess explicit shapes, discrete underground defects (e.g., void) manifest as distinctive hyperbolic diffraction patterns in B-scan images due to the wide-beam nature of electromagnetic waves. A standard 3 × 3 kernel is often insufficient to encompass the spatial span of these hyperbolas, potentially leading to the misinterpretation of partial arcs as noise. By expanding the receptive field to 7 × 7, the regression branch can capture the complete geometric structure of the diffraction signature—encompassing both the apex and the extending tails. This holistic spatial modeling is critical for accurately regressing the bounding box coordinates of the physical defect source, rather than merely detecting localized signal anomalies.

In the original YOLOv11n detection head, the classification branch employs two cascaded DSConv layers. Although this design reduces computational complexity, it limits feature aggregation, potentially weakening high-level semantic representation. To address this, the classification subnetwork was redesigned by replacing the DSConv layers with two consecutive 1 × 1 standard convolutions. This modification preserves computational efficiency while enhancing inter-channel interactions, ensuring robust semantic representation without additional overhead.

3.4 RepNCSPELAN4_CAA

The RepNCSPELAN4 module constitutes a core innovation of this study, integrating re-parameterization with cross-scale feature aggregation via a multi-path cooperative mechanism to facilitate efficient feature extraction and fusion. As illustrated in Figure 7, the input channels are initially expanded via a 1 × 1 convolution and subsequently partitioned into two distinct processing paths. The primary path employs re-parameterized Cross-Stage Partial networks (RepNCSP). During training, these networks enhance feature representation through multi-branch convolutions; during inference, they are structurally re-parameterized into a single-path architecture, thereby maintaining high accuracy while minimizing computational latency. The secondary path cascades RepNCSP submodules to refine multi-level features, progressively augmenting local details retention through channel compression, re-parameterized bottlenecks, and residual connections. Shallow and deep features are densely concatenated along the channel dimension, and the resulting output is compressed via a 1 × 1 convolution. This process yields a multi-scale feature map that synthesizes rich spatial details with robust semantic representations. By leveraging re-parameterization and dynamic channel splitting, the module optimizes the balance between efficiency and feature diversity. This design effectively captures multi-scale spatial relationships amidst complex geological clutter, establishing a robust foundation for real-time, high-precision detection.

Figure 7 Schematic diagram of the RepNCSPELAN4. (A) Schematic diagram of the RepNCSPELAN4; (B) schematic diagram of the RepNCSP; (C) schematic diagram of the RepNBottleneck

The RepNCSPELAN4 module integrates re-parameterization, cross-stage channel optimization, and multi-scale feature aggregation to achieve an optimal balance between processing speed, detection accuracy, and architectural lightness for subsurface road defect detection. During the training phase, multi-branch convolutions are employed to enrich feature representation; conversely, during inference, these branches are structurally re-parameterized into a single-path architecture. This mechanism yields significant performance gains with minimal computational overhead. A cross-stage channel splitting strategy partitions input features into two distinct paths: a primary path that deepens feature abstraction via stacked re-parameterized bottlenecks, and a lightweight secondary path that preserves raw feature information. This dual-path design effectively mitigates gradient redundancy and reduces computational load, rendering the module highly suitable for edge deployment.

Furthermore, an ELAN-style multi-scale dense feature concatenation strategy fuses shallow detail with deep semantic information, enhancing discrimination for small defects, dense defect clusters, and targets within complex geological backgrounds. This design maintains a lightweight structure while effectively addressing conventional limitations—such as the prevalence of false negatives and false positives in multi-scale target detection—providing a robust framework for efficient, high-resolution detection. To further augment this structure, we incorporate the context anchor attention (CAA) module, originally proposed by Cai et al. [20] and illustrated in Figure 8.

Figure 8 Schematic diagram of the CAA

The CAA module models context via multi-stage collaborative processing. Input features X are first compressed through global average pooling (P_avg), followed by a 1 × 1 convolution adjusts channel dimensions to produce the initial global context representation F^pool. This provides a semantic basis for subsequent long-range modeling, as expressed in Equation (13) [20].$F^{\text{pool}}={\text{Conv}}_{1\times 1}\left(P_{\text{avg}}\left(X\right)\right)$(13)

Subsequently, depthwise separable one-dimensional strip convolutions are applied along the horizontal (1 × k_b) and vertical (k_b × 1) axes to model spatial dependencies across width and height, respectively. The kernel size is dynamically set as k_b = 11 + 2n (where n denotes network depth), allowing the receptive field to expand linearly while better accommodating elongated target structures, as described in Equations (14) and (15) [20].$F^w={\text{DWConv}}_{1\times k_b}\left(F^{\text{pool}}\right)$(14)$F^h{\text{ = DWConv}}_{k_b\times 1}\left(F^w\right)$(15)

Finally, the direction-sensitive feature F^h is mapped into a spatial attention map A. A lightweight convolution followed by a sigmoid activation generates spatial attention weights, which are dynamically fused with local features P. A residual connection is then applied to enhance responses in target regions, yielding the output enhanced feature F^attn, as defined in Equations (16) and (17)[20]:$A=\sigma \left({\text{Conv}}_{1\times 1}\left(F^h\right)\right)$(16)where σ denotes the sigmoid function.$F^{\text{attn}}=\left(A\odot P\right)\oplus P$(17)

The fusion strategy defined in Equation (17) represents a residual attention mechanism, where "$\odot$" denotes element-wise multiplication, and "⊕" denotes element-wise addition. The term A$\odot$P injects global contextual guidance (captured by the large-kernel strip convolutions) into the local features, while the term ⊕P preserves the original high-resolution spatial details. This design ensures robustness against scale variation, allowing the network to simultaneously model extensive defects (e.g., water-rich) via global context and small anomalies (e.g., cavity) via local features. This dynamic balance between semantic abstraction and textural preservation is critical for multi-scale detection.

By decomposing large-kernel convolutions into lightweight one-dimensional operations, this process captures cross-region dependencies of elongated subsurface defects while reducing computational cost. It also suppresses background noise, balancing local detail preservation with global semantic representation.

The CAA module is integrated between the multi-scale feature concatenation stage and the final convolution within the RepNCSPELAN4 module. This integration employs a dynamic context anchoring strategy to enhance semantic awareness during feature fusion, forming the RepNCSPELAN4_CAA module within the YOLOv11n architecture. The design preserves the inherent advantages of re-parameterization and cross-stage channel optimization while explicitly addressing extreme target-scale variations and complex background interference in underground scenarios. The CAA module's 1D strip convolutions adaptively capture long-range dependencies, while its spatial attention mechanism softly filters the concatenated multi-scale features, enhancing target-region responses and suppressing background clutter. This approach is particularly effective for extensive defects such as voids and loosened zones, providing efficient feature enhancement for complex detection scenarios.

4.1 Evaluation metrics

To validate the performance of the proposed method, we employ precision (P), recall (R), F1-score, mean Average Precision (mAP), floating point operations (FLOPs), and parameter count (Params) as evaluation metrics. Precision measures the proportion of true positives among all positive predictions, reflecting the reliability of the model's detections. Recall quantifies the proportion of actual positive instances correctly identified, indicating the model's coverage capability. mAP, a standard benchmark in object detection, represents the mean of the Average Precision (AP) calculated across all defect classes. Furthermore, FLOPs and Params serve as indicators of computational complexity and model size, respectively. Lower values for these metrics denote reduced resource consumption, which is critical for facilitating deployment on resource-constrained edge devices. The corresponding mathematical formulations are provided below:$P=\frac{\text{TP}}{\text{TP}+\text{FP}}$(18)$R=\frac{\text{TP}}{\text{TP}+\text{FN}}$(19)$P_{\text{A}}={{\displaystyle \int }}_0^{1}P\left(R\right)\text{d}R$(20)$\text{mAP}=\frac{1}{N}{\displaystyle \sum _{i=1}^N}{P}_{\text{A}i .}$(21)In Equations (18)–(21), TP denotes the number of true positives (correctly identified targets); FP represents false positives (negative samples misclassified as positive); FN indicates false negatives (positive samples misclassified as negative); N is the total number of classes; P_Ai denotes the AP for class i, and mAP represents the mean of the average precision values calculated across all classes.

4.2 Experimental configuration

All deep learning experiments were conducted on a server running the Ubuntu operating system, configured with Python 3.10.14, PyTorch 2.0.1, and CUDA 12.1. The hardware infrastructure comprised an Xeon Gold 6133 CPU with 128 GiB RAM and an NVIDIA GeForce RTX 4090 GPU with 24 GiB VRAM. Key hyperparameters are listed in Table 5. Input images were resized to 640 × 640 pixels via bilinear interpolation, and the batch size was set to 64. The model was trained for 300 epochs using stochastic gradient descent (SGD) with an initial learning rate of 0.01 and a momentum of 0.9. To stabilize early training, a 1,000-iteration warm-up phase was implemented, during which the bias learning rate was gradually increased from 0.1 and momentum from 0.8 to 0.9. A weight decay of 0.0001 was applied throughout to prevent overfitting. An early stopping mechanism was employed, halting training if no performance improvement was observed over 15 consecutive epochs, thereby preserving optimal generalization. The loss function used is defined as follows:$L_{\text{aux}}={\lambda }_{\text{box}}{L}_{\text{box}}+{\lambda }_{\text{cls}}{L}_{\text{cls}}+{\lambda }_{\text{dfl}}{L}_{\text{dfl}}$(22)where L_box denotes the bounding box regression loss, L_cls represents the classification loss, and L_dfl corresponds to the distribution focal loss. The constants λ_box, λ_cls and λ_dfl were set to 7.5, 0.5, and 1.5, respectively.

To ensure a strictly fair comparison, all benchmark models were retrained using the "subData" dataset constructed in this study. All models adhered to an identical data partitioning scheme (7 : 2 : 1 for training, validation, and testing) and were executed under uniform hardware configurations.

4.3 Results and analysis

4.3.1 Benchmark model comparison

As detailed in Table 6, two-stage detectors such as Faster R-CNN [21] and Cascade-RCNN [22] achieved mAP@50 of 0.553 and 0.574, respectively. However, their substantial computational costs preclude real-time deployment for road defect detection. Similarly, alternative single-stage detectors—DINO [23], TOOD [24], and RetinaNet [25]—yielded lower mAP@50 compared to YOLO variants while exhibiting excessive parameter counts and reduced computational efficiency, rendering them unsuitable for lightweight applications. Within the YOLO family, models such as YOLOv8s [26], YOLOv10s [27], YOLOv11n [28], YOLOv11s [28], and YOLOv12s [29] demonstrated competitive mAP@50 performance. Nevertheless, the larger variants (YOLOv8s, YOLOv10s, and YOLOv11s) incurred high computational penalties, limiting their utility for rapid on-site detection. Our comparative analysis reveals distinct trade-offs inherent in existing architectures. Two-stage detectors like Faster R-CNN [21], while structurally robust, suffer from excessive computational redundancy in region proposal generation, leading to low inference speeds (< 45 frames per second (FPS)) unsuitable for real-time GPR scanning. Conversely, advanced transformer-based models like DINO [23], despite their global receptive fields, introduce high parameter overheads (Params >100M), complicating edge deployment. Among the YOLO series, although YOLOv8s [26] and YOLOv12s [29] achieve competitive accuracy, they struggle to balance fine-grained localization with computational cost. MARD-Net addresses these specific limitations: the LADH module significantly reduces the regression burden via asymmetric branches, while MPCA effectively suppresses the high-frequency clutter that confuses standard CNN backbones. Consequently, MARD-Net achieves the highest mAP@50 (0.6926) among all compared models, outperforming both the baseline YOLOv11n and the advanced YOLOv12s. Furthermore, it reduces the parameter count by approximately 0.25 M compared to YOLOv11n. Crucially, regarding empirical inference speed on an NVIDIA RTX 4090, MARD-Net attains 168 FPS (Batch = 1) and 548 FPS (Batch = 32), surpassing the baseline YOLOv11n (152 FPS) and significantly outperforming the heavier YOLOv12s (99 FPS). MARD-Net thus offers an optimal equilibrium between detection accuracy and computational efficiency, validating its suitability for high-precision, lightweight underground defect detection in resource-constrained environments.

Table 6 Comparative performance of MARD-Net, YOLO series models, and other state-of-the-art frameworks. Note: to ensure a strictly fair comparison, all models were retrained on the subData dataset using identical data splits. Inference speeds were measured on a single NVIDIA GeForce RTX 4090 GPU

	Model	mAP@50	mAP@50: 95	Param (M)	FLOPs (G)	FPS(Batch=1)	FPS(Batch=32)
others	Faster-RCNN [21]	0.5530	0.2860	41.374	192	24	43
	Cascade-RCNN [22]	0.5740	0.2630	69.229	219	18	30
	DINO [23]	0.5500	0.2740	47.55	256	22	41
	TOOD [24]	0.5690	0.2930	32.03	180	26	49
	RetinaNet [25]	0.5340	0.2720	36.43	190	25	46
YOLO series	YOLOv8n [26]	0.6639	0.4145	2.6855	6.8	143	485
	YOLOv8s [26]	0.6821	0.4411	9.8300	23.4	96	267
	YOLOv10n [27]	0.6471	0.4083	2.2663	6.5	148	483
	YOLOv10s [27]	0.6810	0.4402	7.2203	21.4	102	284
	YOLOv11n [28]	0.6719	0.4270	2.5833	6.3	152	514
	YOLOv11s [28]	0.6821	0.4411	9.4151	21.3	98	269
	YOLOv12n [29]	0.6552	0.4109	2.5395	6.3	150	508
	YOLOv12s [29]	0.6817	0.4409	9.1963	21.2	99	278
MARD-Net(Ours)		0.6926	0.4507	2.3301	4.3	168	548

Compared with the baseline YOLOv11n [28], MARD-Net reduced parameters by 0.2532 M and computation by 2 GFLOPs, while substantially improving mAP@50. To visually evaluate the localization performance, Figure 9 presents a qualitative comparison where model predictions (solid red boxes) are superimposed onto ground truth annotations (solid green boxes). As illustrated in Table 6, the benchmark models (YOLOv8s [26], YOLOv11n [28], and YOLOv12s [29]) frequently exhibit spatial deviations or generate false positives, particularly in regions with dense interference. In contrast, MARD-Net demonstrates a high degree of overlap (IoU) between predictions and ground truth across all defect types—void, cavity, loosening, and water-rich. Notably, in complex backgrounds, MARD-Net effectively suppresses false positives where other models produce erroneous bounding boxes, while maintaining precise boundary delineation and higher confidence scores. Collectively, MARD-Net exhibits superior robustness and discriminative capability, validating its effectiveness for high-precision underground detection.

Figure 9 Visualization and presentation of results. (A) The result of void; (B) the result of cavity; (C) the result of loosening; (D) the result of water-rich