Topographic correction for ground penetrating radar for enhanced detection of dike subsurface anomalies

Abstract

Dikes are critical components of flood control infrastructure in China. Ground-penetrating radar (GPR) is widely used for non-destructive detection of subsurface anomalies (e.g., cavities, seepage paths, and soil loosening) in dikes. Conventional GPR surveys are typically conducted along longitudinal lines on the dike crest, which limits comprehensive cross-sectional assessment. This study introduces a novel trapezoidal survey layout that integrates data from upstream/downstream slopes and the crest, coupled with a terrain-correction algorithm. The method projects slope-acquired radar signals onto a topographic model, enabling fused visualization of subsurface anomalies and dike morphology. Validated via borehole and cone penetration tests (CPT) at a Yellow River dike site, the corrected GPR profiles accurately delineated soil stratification and localized a loose-soil zone at the downstream slope foot. This approach significantly enhances interpretability for dike safety assessments.

Introduction

China has constructed a comprehensive network of river dikes spanning 325,000 km, serving as a critical barrier for flood protection¹. However, dikes often harbor hidden imperfections such as termite nests, cavities, seepage pathways, cracks, and differential settlements^2–4, which pose significant risks to their structural safety⁵. Proactively identifying such hidden anomalies is therefore paramount for preventive maintenance and disaster risk reduction.

Ground-penetrating radar (GPR) offers distinct advantages for dike inspection, including real-time data acquisition, high-resolution imaging, and minimal environmental interference^6,7. Furthermore, GPR excels at detecting unequal settlement⁸. Compared to invasive drilling methods⁹, GPR ensures non-destructive evaluation with enhanced efficiency and spatial coverage. It exerts minimal impact on the normal operation of the inspected hydraulic structures and provides more comprehensive detection data^10,11. The application of GPR and other geophysical methods for the specific purpose of dike and levee inspection has been established in numerous studies, forming a solid foundation for this work^12,13. Currently, GPR applications in dike detection primarily involve surveys along the dike axis^14,15. Such surveys on complex structures like dikes can be challenging due to topographic effects, and efforts to advance processing techniques for such settings, including the management of wavelet dispersion, are ongoing¹⁶. The detection depth of GPR is influenced by the moisture and clay content of the dike¹⁷. Specific applications to river embankments have further detailed the interplay between material properties, moisture, and GPR response¹⁸. Antenna systems now include both single-frequency and multi-frequency options, facilitating the detection of hidden danger in dikes^19,20. Concurrently, advancements are being made in the automated interpretation of GPR data, where deep learning approach is increasingly applied to identify and classify subsurface anomalies such as cavities and voids from radar images^21,22. A recent approach to vehicle-mounted GPR with prior map-based imaging has achieved speeds up to 100 km/h²³. Based on identified risk locations, a step-by-step measurement approach is employed, starting from the upstream slope, moving to the crest, and then to the downstream slope, with detailed measurements conducted perpendicular to the dike axis.

However, a significant challenge remains in the seamless integration of data from all three key segments of a dike—the upstream slope, crest, and downstream slope—into a single, coherent cross-sectional profile that is intuitively aligned with engineering models. Most existing topographic correction methods are applied to individual 2D lines, lacking a holistic view of the dike’s internal integrity.

This study addresses the limitations of conventional GPR surveys by introducing a topographic correction framework. By integrating slope and crest radar data with terrain modeling, the method enhances the interpretability of subsurface features, providing a more intuitive representation of dike conditions.

Methodology

Survey layout and data acquisition

Traditional GPR surveys (black dashed lines in Fig. 1) are typically aligned along the dike axis, while the proposed method employs a trapezoidal layout (red lines) spanning the upstream slope, crest, and downstream slope. The dike model (Fig. 2) is parameterized by crest width, slope lengths (upstream/downstream), slope angles, and depth (Z), discretized using a square mesh (∆l) for numerical processing. The principles and techniques of GPR are not elaborated here, readers are referred to the relevant literature for details^24–27.

Fig. 1.

Layout of the dike. Black dash line denotes the traditional survey line, the red line denotes the proposed survey line.

Fig. 2.

Schematic diagram of the topographic correction process. (a) Parameters of dike cross-section, (b) schematic diagram of square mesh.

Topographic correction algorithm

The algorithm resolves slope-induced signal distortion through terrain-constrained projection (Fig. 2). The process is defined within a global coordinate system where the X-axis is horizontal and the Z-axis is vertical downward, with the origin set at the intersection of the extension line at the top of the dike and the vertical line at the bottom of the upstream slope, as shown in Fig. 2a.

Step 1: Crest data discretization.

The crest section (width , depth Z) is discretized into grid matrix with spacing :

1

where v is the propagation velocity of radar wave in dike soil (generally 0.1 m/ns), and dt represents the sampling time interval of GPR signal. GPR amplitudes are assigned to the grid via lateral resampling of the original radar traces.:

2

where m denotes the longitudinal mesh count, , and n denotes the horizontal mesh count, .

Step 2: Downstream slope data projection.

For the downstream slope (length , downstream slope angle ):

Resample radar data to match slope grid density, and map amplitudes to terrain-aligned grid:

3

where i denotes the longitudinal mesh count, j denotes the horizontal mesh count, horizontal distance from toe. The acquired radar data is projected onto the global (X, Z) grid. For each grid point (i, j) on the slope-defined mesh, its global coordinates are calculated as:

4

5

Linear interpolation was used for all resampling operations in this study.

Step 3: Upstream slope data projection.

For upstream slope (length , downstream slope angle ):

Resample radar data to match slope grid density, and map amplitudes to terrain-aligned grid:

6

where k denotes the longitudinal mesh count, q denotes the horizontal mesh count, horizontal distance from toe. The acquired radar data is projected onto the global (X, Z) grid. For each grid point (i, j) on the slope-defined mesh, its global coordinates are calculated as:

7

8

The amplitude is assigned from the upstream slope radar data via linear interpolation.

Step 4: Data fusion.

The three individual matrices are integrated into a unified cross-section matrix by concatenating them along the hozizontal (X) axis, ensuring surface continuity at the junctions between the slope and crest segments:

9

where denotes concatenation along the horizontal axis constrained by surface continuity.

Parameter selection and computational efficiency:

The mesh size is a critical parameter. It was set to 0.026 m in this study. This value was chosen to match the highest lateral resolution of our GPR data (50 scans/meter) while maintaining a manageable computational load. A smaller would not yield more information, while a larger one would degrade resolution. The algorithm, implemented in MATLAB and leveraging matrix operations, is computationally lightweight. The computational load scales approximately linearly with the total length of the survey lines, as the processing is applied independently to each trace and subsequently fused. To validate the proposed method, a case study was conducted at a representative dike site along the Yellow River.

Case study

Site description

The Fengqiu dike section along the Yellow River, modified through heightening and impermeable wall installations in 1951 and 1976. It features a crest width of 10 m with 1:3 slope ratios (slope angle 18.4°). The site exhibits elevation variations (85.9–86.3 m) and complex stratigraphy due to historical construction practices, making it susceptible to hidden defects. The study area’s location is shown in Fig. 3.

Fig. 3.

Overview of the study area. (a) Google map of study area. The base map in this figure is from Google Maps. ©2024 Google. (b) Data acquisition along the dike crest. (c) GPR equipment and antenna.

GPR implementation

A dual-frequency GPR system (70 MHz and 300 MHz antennas, https://impulseradargpr.com/crossover/) was employed. This dual-frequency strategy optimizes the trade-off between penetration depth and spatial resolution for dike assessment. The 70 MHz antenna ensures sufficient depth penetration to evaluate the dike’s global structure. Parameters were optimized for depth-resolution balance: dielectric constant = 8, time windows = 500 ns, 512 samples, and 50 scans/m. Data acquisition covered upstream slope, crest, and downstream slope, with an odometer-calibrated wheel ensuring profile length accuracy. Given the selected antenna frequencies and a dielectric constant of 8, the theoretical vertical resolution (λ/4) is approximately 0.35 m for the 70 MHz antenna²⁸. In this study, the method successfully delineated a subsurface anomaly (the loose-soil zone) with lateral dimensions of approximately 3.0 m, demonstrating its practical detection capability for significant features.

To accurately align GPR profiles with the dike topography, marker points were established at the toe and crest of both upstream and downstream slopes. The GPR system’s odometer was calibrated at these markers to ensure precise horizontal positioning. The starting and ending points of each slope segment were explicitly recorded and used to constrain the data projection in the topographic correction algorithm. Preprocessing via Reflexw2D included time-zero correction, DC removal, automatic gain control, band-pass filtering, and moving average to mitigate noise. Terrain data were integrated using the proposed algorithm to generate topographically corrected GPR profiles.

Survey implementation and logistics

The conventional approach provides information only for the crest, potentially missing critical anomalies developing on the slopes—the most vulnerable parts of a dike. The trapezoidal method, while slightly more resource-intensive per unit length, yields a comprehensive cross-sectional dataset that is invaluable for a conclusive safety assessment. This integrated view can prevent costly failures by enabling the early detection and precise localization of defects, thereby guiding more effective and targeted remediation efforts.

Results

GPR imaging outcomes

Uncorrected GPR profiles (Fig. 4) showed disrupted reflections on slope surfaces, complicating anomaly interpretation. After topographic correction and data fusion (Fig. 5), a distinct clutter zone emerged at the leeward slope foot, indicative of a loose soil zone. This reflection pattern is characteristic of material heterogeneities such as loose soil. In general, different anomaly types exhibit distinct signatures in corrected GPR profiles; for instance, cavities typically produce hyperbolic diffractions, while seepage paths often cause signal attenuation zones⁴.

Fig. 4.

Processed GPR profiles with annotated reflection features before topographic correction. The concrete crest (Fig. 3) produces a stronger surface reflection than the grass-covered slopes. The crest profile is extended 1.5 m upstream and downstream to bridge the slopes, ensuring signal continuity for data fusion.

Fig. 5.

GPR profile after topographic correction and data fusion. The red circle represents the identified hidden defects, and the white dashed line represents the inferred stratification boundary.

Validation by in-situ tests

Borehole drilling (20 m depth) revealed stratified artificial fill composed of silt and silty clay (Fig. 6a), confirming the general soil composition in the shallow section. To provide a quantitative assessment of the GPR’s accuracy, two types of comparisons were made. First, the static cone penetration tests (CPT), which provide continuous data to a greater depth, identified clear soil interfaces. These CPT-measured interfaces were compared with the corresponding reflectors in the topographically corrected GPR profile, primarily from the crest and upstream slope sections where the signal penetration was sufficient. As summarized in Table 1, the depth agreement for four major interfaces is excellent. The calculated MAE for depth estimation is 0.275 m, with an RMSE of 0.328 m. The larger RMSE stems mainly from the errors at deeper interfaces, where the limiting vertical resolution of the GPR signal and the expanding Fresnel zone make precise interface picking more challenging compared to the direct mechanical measurement of the CPT. The MAE, however, more accurately reflects the method’s overall reliable performance. This level of accuracy is deemed excellent for GPR applications in geotechnical contexts. Second, for the anomaly detection capability, the ‘loose-soil zone’ identified in the corrected GPR profile (Fig. 5) as a distinct clutter zone was found to correlate with a zone of reduced CPT tip resistance (Fig. 6b,c). The lateral extent of this anomaly, as interpreted from the GPR profile, is approximately 3.0 m, demonstrating the method’s capability to not only locate but also delineate the scale of subsurface defects.

Fig. 6.

Geological borehole bar diagram and static cone penetration cone curve. (a) Drill hole Log, (b) plot of the cone penetration test (CPT) tip resistance versus depth, (c) plot of the lateral frictional resistance versus depth.

Table 1.

Quantitative comparison of interface depths between GPR and CPT.

Interface description	CPT depth (m)	GPR depth (m)	Absolute error (m)
Interface 1	3.1	3.2	0.1
Interface 2	6.8	7.2	0.4
Interface 3	10.5	10.4	0.1
Interface 4	11.3	11.8	0.5

Mean Absolute Error (MAE): 0.275 m; Root Mean Square Error (RMSE): 0.328 m.

Discussion

The strong empirical validation provided in this study, with a quantitative mean absolute error of 0.275 m for stratigraphic interfaces, confirms that resolving the geometric distortion is the most critical step for achieving an interpretable subsurface profile of dikes. The higher RMSE is largely attributable to increased depth estimation uncertainties at deeper interfaces, where GPR signal resolution naturally diminishes. The method also successfully delineated a loose-soil anomaly with a lateral extent of approximately 3.0 m, demonstrating its practical capability for defect detection and characterization. However, it is important to acknowledge that the primary validation was performed at a single site. Consequently, the generalizability of the proposed method across dikes with significantly different geological conditions and construction histories warrants further investigation through systematic multi-site validation in future work. Furthermore, while the algorithm effectively corrects for geometric distortion, it does not explicitly model other physical effects induced by topography, such as variations in antenna-ground coupling. The empirical success suggests these are secondary for initial interpretation, but the dual-frequency strategy (70/300 MHz) served as a pragmatic mitigation, ensuring both deep penetration and high-resolution shallow imaging. The 300 MHz data is highlighted for resolving shallow, critical zones where coupling and resolution are paramount, while the 70 MHz data ensures sufficient penetration for a holistic integrity assessment.

The novelty of the proposed approach is most evident when compared to conventional topographic correction methods, which are typically applied to individual, isolated 2D lines. Our method’s core innovation lies in the integrated trapezoidal survey layout and the subsequent data fusion process, which synthesizes discrete profiles into a single, continuous cross-sectional image. This provides a holistic view that is directly analogous to the engineering cross-sections used in dike assessment, offering a clear advantage for interpreting structures that span the entire dike profile. This integrated methodology is inherently compatible with modern surveying platforms. The fusion of our processing framework with drone-based GPR systems, for instance, presents a significant opportunity to automate the trapezoidal survey, enhance safety, and provide co-registered, high-resolution topographic data, thereby transforming the method from a spot-check tool into an efficient system for large-scale dike monitoring²⁹.

For practical application, the influence of environmental conditions, particularly soil moisture content, must be considered. Variations in moisture alter the soil’s dielectric constant and the radar wave velocity, which can affect depth estimation accuracy between surveys¹³. Therefore, we recommend conducting baseline surveys during stable, dry periods to ensure a consistent velocity model. The algorithm itself is designed for practical utility; its computational load scales approximately linearly with survey length. The primary output—a georeferenced, topographically corrected cross-section—is readily integrable with Geographic Information Systems (GIS) and digital twin platforms. This compatibility, combined with the method’s capacity for clear anomaly localization, paves the way for its use in automated defect recognition and long-term performance monitoring within comprehensive digital dike management systems.

Conclusion

This study presents a topographic correction framework for GPR in dike inspection, addressing the limitations of conventional longitudinal surveys. By integrating slope and crest data with terrain modeling, the method clearly displays radar profiles with terrain and enables accurate localization of hidden defects, as validated by field tests at the Fengqiu dike section and quantitative error metrics. The spatial accuracy was quantified, showing a mean absolute error of 0.275 m in depth for stratigraphic interfaces when compared to cone penetration tests.

A recognized limitation of this work is that the primary field validation was conducted at a single site. Consequently, the generalizability of the proposed method across dikes with significantly different geological conditions, construction histories, and topographic designs should be further established. Future work will therefore focus on systematic multi-site validation to rigorously test the method’s applicability and develop guidelines for its reliable deployment across diverse flood defense infrastructure.

Author contributions

X.Z. designed the study, conducted experiments, and wrote the manuscript. C.Z.L. contributed to algorithm development and data interpretation. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Joint Funds of the National Natural Science Foundation of China (No. U2243223), the Joint Funds of the National Natural Science Foundation of China (No. U2443229).

Data availability

Correspondence and requests for materials should be addressed to X.Z. and CZ.L.

Declarations

Competing interests

The authors declare no competing interests.