How Small a Defect Can Radar See Through Rebar? Full-Wave Detection Limits for Metal-Shadowed Anomalies

Imaging methods that use coherent waves, such as ground-penetrating radar, microwave imaging and radio-frequency sensing, are often asked to find a small, weak defect sitting just behind a strong metal scatterer. Examples include a delamination beneath the steel reinforcing bars (rebar) in a bridge deck, a void behind a buried pipe, or tissue next to an implant. The field judges its methods by how clean the reconstructed image looks, using peak signal-to-noise ratio (PSNR) and structural similarity. Neither number answers the operational question: at a fixed false-alarm rate, what is the smallest defect that any processing of these measurements could reliably detect? We propose to compute that limit directly. We use an exact wave simulator as the forward model and borrow the detection-limit methods that particle physicists use to report sensitivity. The result is a pair of curves that bracket the best achievable detection performance, which any proposed method can then be measured against.

The core mechanism is finite-difference time-domain (FDTD) simulation, a grid solver that steps Maxwell's equations forward in time and captures the diffraction, multiple scattering and near-field effects that metal produces. The statistical lens comes from rare-event physics, where the standard question is how detection sensitivity scales with exposure and background. Metal interference sets the regime: in this regime the background is not white noise but structured clutter that looks like the signal. Our main claim is that metal-shadowed detection is clutter-limited rather than noise-limited. Past a threshold, more transmit power or more averaging gives no further gain, and only greater diversity in how the scene is illuminated (frequencies, antenna positions, polarizations) helps. Simplified scattering models that ignore multiple scattering cannot locate that threshold correctly once the rebar spacing drops below about half a wavelength in the host material.

Core Claim

Take a small defect producing a change $\Delta s$ in the measured radar traces (the signal). The background contains nuisance variation, such as rebar position jitter, heterogeneous concrete permittivity and moisture, with covariance $C$ across many simulated scenes, and the receiver adds noise of variance $\sigma^2$. The best linear detector, known in imaging-observer theory as the Hotelling observer, then achieves the detectability

$$d'^2 = \Delta s^\top \left(C + \sigma^2 I\right)^{-1} \Delta s .$$

When clutter is negligible, $d'^2$ grows as $1/\sigma^2$, so it improves like $\sqrt{N}$ with $N$ averaged acquisitions. This is the familiar counting-statistics scaling. When the defect's signature lies mostly inside the space spanned by clutter variation, $d'$ levels off as $\sigma \to 0$, no matter how many acquisitions are averaged. Three things in that expression depend on full-wave physics, and FDTD supplies all of them:

  • the defect signature, because the metal grid filters and re-radiates the wave before it reaches the defect;
  • the clutter covariance, because rebar jitter produces diffraction patterns, not just shifted reflections;
  • their overlap, which decides where detectability stops improving.

We make two falsifiable predictions. First, the minimum detectable delamination thickness rises steeply once rebar spacing falls below roughly half the wavelength in concrete. In that regime the rebar mesh acts like a reflecting screen and the field reaching the defect is mostly evanescent, meaning it decays with distance instead of propagating. Second, a first-order scattering (Born) model will badly misplace both this threshold and the point where detectability saturates.

Research Program

The application is ground-penetrating radar (GPR) inspection of reinforced concrete. This choice follows Quantitative Evaluation of Internal Pavement Distresses Based on 3D Ground Penetrating Radar. That paper used 3D FDTD to map radar echo amplitudes to crack size in asphalt, but it assumed a background with no metal and did not ask what crack size is detectable at a given false-alarm rate. We add the rebar and ask that question.

We will write a 2D transverse-magnetic FDTD solver in JAX that can be differentiated end to end. It uses absorbing boundary layers of the kind described in Subwavelength Semiconductor Nanocavity Laser. At roughly 1–2 GHz in a domain about 1 m by 0.5 m, a 2 mm grid gives scenes that run in seconds on a GPU, so 10^5 simulations are affordable. JaxLayerLumos shows that differentiable electromagnetic solvers in JAX are practical, but it handles only flat layered media and cannot represent rebar. At these frequencies rebar behaves almost as a perfect conductor, so we do not need the dispersive metal models of Metallic nano-structures for polarization-independent multi-spectral filters. Rebar is round, however, and a square grid approximates round objects with jagged steps. We will therefore confirm convergence by refining the grid and comparing against a subcell boundary treatment.

Each scene draws its nuisances from explicit priors:

  • rebar spacing, cover depth and positional jitter;
  • a random permittivity field for the concrete;
  • a moisture gradient;
  • surface roughness;
  • receiver noise.

The target is a thin air-filled delamination, or a void, placed beneath the rebar mesh. We sweep its thickness and lateral position.

We bracket the achievable performance from both sides.

  • Upper bound. If every nuisance were known exactly, the ideal detector would be a matched filter. Its detectability $\|\Delta s\|/\sigma$ is computed exactly from paired FDTD runs with and without the defect, and no detector facing unknown nuisances can beat it.
  • Lower bounds. The first is the Hotelling detectability above. We estimate $C$ from FDTD samples and use the solver's gradients to build a local Fisher-information approximation for very small defects. The second is a learned likelihood-ratio classifier trained on raw traces. This follows the precedent of Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset, where a neural discriminant raised signal efficiency from 80.8% to 96.4±3.0%. That result is evidence that learned detectors can recover sensitivity that hand-built statistics miss.

The gap between the upper and lower bounds measures how much detectability the nuisances themselves cost.

Two further borrowings from rare-event statistics shape the analysis. Limits on non-Gaussianities from WMAP data added linear correction terms to its estimators to cancel spurious signal created by sky masks. Our Hotelling detector does the same job here, removing clutter directions that would otherwise imitate a defect. Indications of intermediate-scale anisotropy of cosmic rays showed that apparent hotspots weaken once the analysis accounts for having searched many sky locations. A GPR scan likewise checks for the defect at every position along the scan, so we report false-alarm rates corrected for the number of positions searched.

Baselines

The naive reference is what field practice does. We remove the mean trace to suppress the horizontal rebar reflections, then apply a matched filter tuned to the defect's signature in a homogeneous medium, which ignores the metal. The detection threshold is set so the false-alarm rate holds on simulated scenes with no defect. Beating it means a lower minimum detectable thickness at a probability of detection of 0.9 and a corrected false-alarm rate of 10^-3.

The second baseline is a generalized likelihood-ratio detector built on a Born (single-scattering) forward model. It is the cheap physics-based alternative a reviewer would propose. It receives the same nuisance priors and the same data splits as our method, and it is fitted by the same profile-over-nuisance optimization.

The third baseline is reconstruct-then-detect, which mirrors how the metal-artifact-reduction literature works. We train a U-Net to map metal-corrupted images to metal-free images on FDTD pairs. This is supervised artifact reduction in the spirit of DuDoNet and Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction. The same detector is then run on its outputs. By the data-processing inequality, this pipeline cannot beat a detector that sees the raw traces. Our question is how much it loses, and whether it creates false defects in scenes that contain none.

The strongest competitor is the learned raw-trace classifier itself, trained with the same number of simulations and the same hyperparameter search budget as every other learned component. It shows how close a strong learned detector gets to the full-wave bound.

There is no reported result on Lacuna for defect detection beneath rebar with GPR, so the file provides no number to beat on this task. The closest reported wave-imaging figure is from A sparse coding approach to inverse problems with application to microwave tomography imaging: 2.9% reconstruction error for an iterative wave-field method, against 9.4% for back-propagation and 3.2% for a U-Net, on soft-tissue targets with no metal. That paper motivates our design, since it shows linearized physics losing to full-wave physics. It does not constrain our numbers.

The key ablation rebuilds the entire bounding pipeline with the Born model in place of FDTD, holding the nuisance priors fixed. This isolates what full-wave physics contributes.

Why not just…?

Why not just use Born or ray models, which are far faster? This is a real threat in one regime. When the rebar is sparse relative to the wavelength, scattering is close to single-bounce and Born-based bounds may agree with FDTD. The Born ablation settles the question by plotting the minimum detectable thickness from both models against the ratio of rebar spacing to wavelength. If the two curves agree everywhere within simulation error, the full-wave claim fails and we will say so. We expect them to diverge below about half a wavelength, because that is where the mesh's response stops being a sum of individual bar reflections.

Why not just train a CNN detector on simulated data and report its ROC curve? The receiver operating characteristic (ROC) curve is useful, and we report it as one of our lower bounds. On its own, though, it cannot tell a weak detector from an inherently undetectable defect. Only the gap to the known-nuisance upper bound separates "the method is weak" from "the information is not in the data." Without that bound, a plateau in CNN performance is uninterpretable.

Why not just reconstruct a clean image and judge it by PSNR? CT artifact-reduction methods already reach high PSNR, for example 37.57 dB for Unsupervised Polychromatic Neural Representation for CT Metal Artifact Reduction and 33.51 dB for DuDoNet on DeepLesion. Yet PSNR averages over the whole image and can rise even while a small defect next to the metal is erased. Generative inpainting, such as 3D CBCT Artefact Removal Using Perpendicular Score-Based Diffusion Models, can also hallucinate defects that are not there. We test this directly by correlating per-scene PSNR with per-scene detectability loss, and by measuring false-alarm rates of the reconstruct-then-detect pipeline on defect-free scenes.

Why not just average more traces or transmit more power? If the clutter-limited claim is correct, this stops helping. The saturation experiment below tests exactly that. If detectability keeps growing like $\sqrt{N}$, the clutter-limited regime is irrelevant for realistic nuisance levels, and the practical advice reduces to "average more."

Proposed Experiments

The first decisive check maps the minimum detectable delamination thickness against rebar spacing over wavelength, cover depth and center frequency, under both FDTD and Born models. It reports the upper bound, the Hotelling bound and the learned-detector curve, and tests the predicted steep rise below about half a wavelength.

The second varies noise variance over four orders of magnitude at fixed clutter. The prediction is that detectability first grows like $\sqrt{N}$ and then plateaus. We then show that adding illumination diversity (more frequencies, both polarizations, wider antenna apertures) lifts the plateau while added power does not. Because the solver is differentiable, we can rank which added measurement most increases Fisher information about the defect. This is a physics-based answer to the acquisition question that MASC addressed with reinforcement learning in MRI. MASC optimized MSE and SSIM (reporting a 20.7% MSE reduction and a 4.0% SSIM gain at 10× acceleration), not detection limits.

The third quantifies the detectability that reconstruct-then-detect loses relative to raw-trace detection, along with its rate of false defects on clean scenes.

As a cheap cross-modality sanity check, we insert synthetic low-contrast lesions next to the simulated metal in DeepLesion. We then test whether PSNR rankings of DuDoNet-style and Polyner-style reconstructions predict how detectable those lesions remain. This uses straight-ray physics and does not involve FDTD.

Every FDTD conclusion must survive grid refinement. If a physical concrete slab with embedded rebar becomes available, a small validation scan would test whether the simulated clutter covariance matches measured clutter. The file contains no public benchmark for this, and we make no claim beyond simulation until such data exist.