The performance of Ground Penetrating Radar (GPR) is greatly influenced by the cross-coupling between the transmitter and the receiver, and the response from the background. Their combined effect often masks the weaker target signals, especially in cases where shallow buried targets are present. Moreover, errors in velocity estimation, results to over/under migrated images, which further compromises the reliability of GPR, especially in case of non-homogeneous media. Therefore, background clutter suppression and velocity estimation are both pivotal for effectively locating targets. For this purpose, a novel deep learning scheme for background clutter prediction was developed, where a two joint artificial neural networks architecture (ANNs) combined with principal component analysis (PCA) is implemented. In the suggested scheme, the first network predicts the background response, which is subsequently subtracted, while the second network estimates the background permittivity and conductivity. Subsequently, the permittivity profile along the measurement line is used as input in Reverse-Time Migration to focus the signal without the need of hyperbola fitting and homogeneity assumptions. The training data were generated synthetically using the Finite-difference time-domain (FDTD) method. A model of a real GPR antenna is used in the simulations, making the scheme applicable to real data. The efficiency of the proposed method is validated using both numerical and real data, with successful predictions in all cases, demonstrating its ability to perform well even when tested with previously unseen real complex scenarios. Via a series of examples, the proposed scheme was proven superior to commonly used background removal techniques and conventional migration.

Numerical modelling of Ground Penetrating Radar (GPR), such as the finite-difference time-domain (FDTD) method, has been extensively used to enhance the interpretation of GPR data and as a key component of full-waveform inversion (FWI). A major drawback of numerical solvers, especially within the context of FWI, is that they are still computationally expensive requiring often unattainable computational resources and access to high performance computing (HPC). In this work, we present a near real-time deep learning forward solver for GPR data that can generate entire B-scans, given certain model parameters as inputs. The machine learning (ML) model is tuned for reinforced concrete slab scenarios, but the same rational can be applied in a straightforward manner to other applications as well. Training was performed using entirely synthetic data, where a three-dimensional (3D) digital twin based on the 2000 MHz “palm” antenna from Geophysical Survey Systems, Inc. (GSSI) was included in FDTD simulations for the training set. The accuracy of the deep learning solver is demonstrated with both synthetic and real data from reinforced concrete slabs. The predicted ML responses were in a very good agreement with FDTD, showing a high degree of accuracy. The ML solver is then used as part of a FWI algorithm to characterize the concrete slab and estimate the depth and radius of the buried rebars. Coupled FWI with an ML-based forward solver results in significantly less execution times compared to conventional FWI using numerical solvers. The high accuracy of the proposed FWI, combined with the efficiency and speed of the ML-based forward solver, make the proposed scheme an ideal tool for characterizing concrete structures in non-destructive testing.

Hyperbola fitting is a mainstream interpretation technique used in ground penetrating radar (GPR) due to its simplicity and relatively low computational requirements. Conventional hyperbola fitting is based on the assumption that the investigated medium is a homogeneous half-space, and that the target is an ideal reflector with zero radius. However, the zero-radius assumption can be easily removed by formulating the problem in a more generalized way that considers targets with arbitrary size. Such approaches were recently investigated in the literature, suggesting that hyperbola fitting can be used not only for estimating the velocity of the medium, but also for estimating the radius of subsurface cylinders, a very challenging problem with no conclusive solution to this day. In this paper, through a series of synthetic and laboratory experiments, we demonstrate that for practical GPR survey, hyperbola fitting is not suitable for simultaneously estimating both the velocity of the medium and the size of the target, due to its inherent non-uniqueness, making the results unreliable and sensitive to noise.

Full-waveform inversion (FWI) is considered one of the most promising interpretation tools for hydrogeological applications using ground-penetrating radar. However, FWI has had limited practical uptake for several reasons: large computational requirements, an inability to reconstruct loss mechanisms of soil, and the need for a good initial starting model. We aim to address these issues via a novel FWI subject to a fractally correlated distribution of water. Initially, the dispersive properties of the soil are expressed as a function of the water fraction using a semiempirical model. This approach means that the permittivity, conductivity, and relaxation mechanisms are all correlated, and therefore, sensitivity problems between the permittivity and loss mechanisms no longer affect the performance of FWI. Subsequently, the distribution of the water fraction is constrained to follow a fractal geometry. Fractal-correlated noise is then compressed using a principal component analysis (PCA) in order to further reduce the number of the system’s unknowns and accelerate FWI. PCA reduces the volume and dimensions of the optimization space, and thus, initialization is no longer necessary. Finally, a novel measurement configuration is suggested that uses superposition with all the individual measurements in order to reduce the number of forward models that need to be executed for every iteration of FWI. These enhancements substantially reduce the computational requirements of FWI and therefore eliminate the need for high-performance computers and time-consuming algorithms. The proposed scheme has been successfully tested with several numerical case-studies, which indicates the potential of this approach to become a commercially appealing interpretation tool for hydrogeology.

Ground penetrating radar (GPR) is a well-established tool for detecting and locating reinforcing bars (rebars) in concrete structures. However, using GPR to quantify the diameter of rebars is a challenging problem that current processing approaches fail to tackle. To that extent, we have developed a novel machine learning framework that can estimate the diameter of the investigated rebar within the resolution range of the employed antenna. The suggested approach combines neural networks and a random forest regression and has been trained entirely using synthetic data. Although the training process relied only on numerical training sets, nonetheless, the suggested scheme is successfully evaluated with real data indicating the generalization capabilities of the resulting regression. The only required input of the proposed technique is a single A-scan, avoiding laborious measurement configurations and multisensor approaches. In addition, the results are provided in real time and making this method practical and commercially appealing.

Finite-difference time-domain forward modeling of ground-penetrating radar (GPR) is becoming regularly used in model-based interpretation methods, such as full-waveform inversion (FWI) and machine learning schemes. Oversimplifications in such forward models can compromise the accuracy and realism with which real GPR responses can be simulated, which degrades the overall performance of interpretation techniques. A forward model must be able to accurately simulate every part of the GPR problem that affects the resulting scattered field. A key element, especially for near-field applications, is the antenna system. Therefore, the model must contain a complete description of the antenna, including the excitation source and waveform, the geometry, and the dielectric properties of any materials in the antenna. The challenge is that some of these parameters are not known or cannot be easily measured, especially for commercial GPR antennas that are used in practice. We present a novel hybrid linear/nonlinear FWI approach that can be used, with only knowledge of the basic antenna geometry, to simultaneously optimize the dielectric properties and excitation waveform of the antenna and minimize the error between real and synthetic data. The accuracy and stability of our proposed methodology are demonstrated by successfully modeling a 1.5-GHz commercial antenna from Geophysical Survey Systems, Inc. Our framework allows accurate models of GPR antennas to be developed without requiring detailed knowledge of every component of the antenna. This is significant because it allows commercial GPR antennas, regularly used in GPR surveys, to be more readily simulated.

The simulation, or forward modeling, of ground penetrating radar (GPR) is becoming a more frequently used approach to facilitate the interpretation of complex real GPR data, and as an essential component of full-waveform inversion (FWI). However, general full-wave 3-D electromagnetic (EM) solvers, such as the ones based on the finite-difference time-domain (FDTD) method, are still computationally demanding for simulating realistic GPR problems. We have developed a novel near-real-time, forward modeling approach for GPR that is based on a machine learning (ML) architecture. The ML framework uses an innovative training method that combines a predictive principal component analysis technique, a detailed model of the GPR transducer, and a large data set of modeled GPR responses from our FDTD simulation software. The ML-based forward solver is parameterized for a specific GPR application, but the framework can be applied to many different classes of GPR problems. To demonstrate the novelty and computational efficiency of our ML-based GPR forward solver, we used it to carry out FWI for a common infrastructure assessment application-determining the location and diameter of reinforcement bars in concrete. We tested our FWI with synthetic and real data and found a good level of accuracy in determining the rebar location, size, and surrounding material properties from both data sets. The combination of the near-real-time computation, which is orders of magnitude less than what is achievable by traditional full-wave 3-D EM solvers, and the accuracy of our ML-based forward model is a significant step toward commercially viable applications of FWI of GPR.

Directly measuring the radiation characteristics of Ground Penetrating Radar (GPR) antennas in environments typically encountered in GPR surveys, presents many practical difficulties. However it is very important to understand how energy is being transmitted and received by the antenna, especially for areas of research such as antenna design, signal processing, and inversion methodologies. To overcome the difficulties of experimental measurements, we used an advanced modelling toolset to simulate detailed three-dimensional Finite-Difference Time-Domain (FDTD) models of GPR antennas in realistic environments. A semi-empirical soil model was utilised, which relates the relative permittivity of the soil to the bulk density, sand particle density, sand fraction, clay fraction and volumetric fraction of water. The radiated energy from the antenna was studied in lossless homogeneous dielectrics as well as, for the first time, in lossy heterogeneous environments. Significant variations in the magnitude and pattern shape were observed between the lossless homogeneous and lossy heterogeneous environments. Also, despite clear differences in time domain responses from simulations that included only an infinitesimal dipole source model and those that used the full antenna model, there were strong similarities in the radiated energy distributions.

A regression model is developed in order to estimate in real time the signal-to-clutter ratio (SCR) for landmine detection using ground-penetrating radar. Artificial neural networks are employed in order to express SCR with respect to the soil's properties, the depth of the target, and the central frequency of the pulse. The SCR is synthetically evaluated for a wide range of diverse and controlled scenarios using the finite-difference time-domain method. Fractals are used to describe the geometry of the soil's heterogeneities as well as the roughness of the surface. The dispersive dielectric properties of the soil are expressed with respect to traditionally used soil parameters, namely, sand fraction, clay fraction, water fraction, bulk density, and particle density. Through this approach, a coherent and uniformly distributed training set is created. The overall performance of the resulting nonlinear function is evaluated using scenarios which are not included in the training process. The calculated and the predicted SCR are in good agreement, indicating the validity and the generalization capabilities of the suggested framework.

We use ground-penetrating radars (GPRs), firn cores, and electromagnetic finite-difference time-domain (FDTD) numerical modeling to characterize the GPR response to a frozen high-arctic firn pack. As a result of extensive summertime percolation, the firn pack comprises a high fraction of ice layers, lenses, and vertical glands. We show that the GPR response on the firn pack mainly depends on the following: (1) the thickness of the ice layers; (2) the distance between layers; (3) the layer roughness; and (4) the presence or absence of elliptical ice lenses. Using 3-D FDTD modeling, we show that the GPR is not sensitive to typical ice glands, which implies that the GPR underestimates firn heterogeneity, such that firn stratigraphy in percolation and wet-snow zones could be incorrectly interpreted as being better preserved than it actually is. We find that thin ice layers (< 0.05 m) or multiple thin ice layers give a strong response. Thicker ice layers typically give a weaker backscatter per unit area, mainly due to the lack of interference of the reflections from the upper and lower interfaces, but are, due to their continuity, easily trackable. Ice layers with a thickness comparable to the GPR wavelength give 180deg phase-shifted upper and lower reflections and are, in general, separated by a band of low GPR response, due to the lack of permittivity contrast within the ice layers. Despite the ice lenses' relatively short horizontal correlation length, as inferred from cores, bands of high-amplitude clutter caused by these features can be traced over several kilometers in GPR profiles.