Stable Nonlinear Imaging of Strong Scatterers from Ambient-Noise Correlations in Random Media

Passive imaging builds a picture of a hidden object without any transmitter. Sensors record background vibrations such as ocean microseisms or thermal noise. Cross-correlating the recordings at two sensors produces the signal one sensor would have received had the other emitted a pulse, so the array behaves like a set of "virtual sources." A surprising theoretical result says that when these waves first cross a randomly scattering medium, the clutter can help: scattering sends waves in from more directions, which widens the effective aperture and sharpens resolution beyond what the physical array alone allows. That result, and every practical passive imaging method built on it, assumes the target is a weak scatterer that the wave touches only once. Real targets are often not weak. Examples are a fluid-filled fault, a high-contrast inclusion in tissue, or a dense dielectric object. For such targets the single-bounce assumption is known to distort shape and underestimate contrast.

We propose an inversion that removes the weak-target assumption without losing the statistical stability that makes passive imaging in clutter work. The idea is to split the physics where each approximation is valid. Between the array and the target, the long path through the random medium is described only statistically, using closed-form moment formulas from a white-noise paraxial model. Paraxial means narrow-angle and forward-going, which is accurate for long propagation paths. The small region around the target is modeled exactly with the Lippmann–Schwinger equation, which includes all wide-angle multiple scattering. We then fit the target not to the raw correlations, which contain speckle that depends on the particular random medium, but to locally averaged products of correlations. The averaging windows are set by the medium's decoherence scales. This loss is stable across medium realizations because the moment theory predicts its expected value in closed form. The main question is whether scattering-enlarged aperture, predicted so far only for weak targets, can be converted into actual resolution and correct contrast for strong ones.

Core Claim

Role of Scattering in Virtual Source Array Imaging uses the Itô–Schrödinger equation to show that isotropic scattering in a slab of thickness $L$ enlarges the effective source aperture to $b_{eff}^2 = b^2 + \gamma^2 L^3/3$. Here $b$ is the physical aperture and $\gamma$ is the scattering strength. That paper then images with Kirchhoff migration under the Born approximation, which assumes the wave scatters off the target only once. Imaging in Random Media by Two-Point Coherent Interferometry shows that locally windowed cross-products of data preserve relative phases through clutter and allow near-ideal resolution. Its authors state explicitly that going beyond single-scattering Born targets is left open. On the other side, nonlinear solvers handle high-contrast targets well. SEAGLE and CISOR do so on phantoms from 0.2% to 32.2% contrast, but they fit an $\ell_2$ misfit to deterministic, actively measured fields.

We claim that a nonlinear inversion with a moment-matched interferometric loss can do three things at once. It can recover the geometry and contrast of targets above roughly 10% contrast from passive correlations. Its cross-range resolution should track the theoretical scale $\lambda L_t / b_{eff}$ rather than $\lambda L_t / b$, where $L_t$ is the target range. Its realization-to-realization variance should be far below that of the same nonlinear model fit to raw correlations. Neither prior line of work does all three.

Forward Model and Loss

Let $C(x_i, x_j, \omega)$ be the empirical cross-correlation between sensors $i$ and $j$ at frequency $\omega$. Let $\chi$ be the unknown target contrast. For a pair of sensor pairs $p, p'$ that are close in position, and frequencies $\omega \pm \Omega/2$, we form the windowed product $\widehat{Q}_{pp'}(\omega,\Omega)$. The spatial window is set by the paraxial decoherence length $X_d$ and the frequency window by the speckle coherence frequency $\Omega_d$. We predict its mean with a hybrid operator. The target's scattered field for each incident direction comes from a Lippmann–Schwinger solve inside a small homogeneous box, reusing the Fourier-domain solver of Three-Dimensional Optical Diffraction Tomography With Lippmann-Schwinger Model. The transport of these fields to and from the array through the random slab enters only through the closed-form second- and two-frequency moments of the paraxial Green's function. These are derived in High-order statistics for the random paraxial wave equation and Speckle Memory Effect in the Frequency Domain and Stability in Time-Reversal Experiments. The inversion minimizes

$$\mathcal{L}(\chi) = \sum_{p,p',\omega,\Omega} w_{pp'} \left| \widehat{Q}_{pp'}(\omega,\Omega) - \mathbb{E}\left[Q_{pp'}(\omega,\Omega;\chi)\right] \right|^2 + \tau\, \mathrm{TV}(\chi).$$

Here $\mathbb{E}[Q(\chi)]$ is the moment-model prediction, $w_{pp'}$ are inverse-variance weights obtained from the fourth-moment theory, $\tau$ weights the regularizer, and TV is total variation, a penalty that favors piecewise-constant images. Optimization uses CISOR's relaxed FISTA, an accelerated proximal gradient method, with gradients obtained by the adjoint-state method (the standard way to differentiate through a wave solver).

Paraxial theory is used only on the propagation path, where it is justified. The non-paraxial target zone is modeled exactly. This addresses the tension raised by High-fidelity intensity diffraction tomography with a non-paraxial multiple-scattering model: that paper shows paraxial marching fails inside dense objects, but it does not show failure over long forward paths.

Research Program

We first derive $\mathbb{E}[Q]$ and its variance for a target illuminated through an isotropic white-noise slab. We check that in the Born limit the expression reduces to the two-point interferometric imaging function of Borcea and Garnier. We then implement the hybrid operator in a differentiable code. Data are generated by full-wave simulation that does not use the paraxial approximation anywhere: a 2D, and later 3D, finite-difference or spectral solver with a random slab, distributed noise sources, a passive array, and the target. This avoids fitting a model to data it generated itself. We also construct anisotropic, partly layered slabs in which the paraxial isotropic model is deliberately wrong. These test where the method breaks down, which Role of Scattering in Virtual Source Array Imaging predicts through an aperture that shrinks under layering.

Baselines

The naive reference is Kirchhoff migration of the raw empirical correlations, exactly as in Role of Scattering in Virtual Source Array Imaging. It uses the same data and needs no tuning beyond the choice of frequency band. It shows what the enlarged aperture gives with no target modeling.

The strongest existing passive method is two-point interferometric imaging with its optimization-based reconstruction, from Imaging in Random Media by Two-Point Coherent Interferometry. It is the strongest competitor because it already achieves near-ideal resolution and signed reflectivity through clutter, and differs from our method only in the Born target model. We give it the same windows, frequency band, recording duration, and a regularization weight tuned on the same held-out realizations. Beating it means lower relative permittivity error and correct shape at contrasts at or above 10%, with no loss of resolution at low contrast.

The strongest nonlinear competitor is CISOR (Accelerated Image Reconstruction for Nonlinear Diffractive Imaging) with a plain $\ell_2$ loss on raw correlations. It gets the same Lippmann–Schwinger solver, iteration budget, and TV tuning grid. That paper reports that its SNR stays high across 0.2–32.2% contrast in deterministic settings, but the file gives no numerical value. We found no reported result on Lacuna for any nonlinear method on passive correlation data in random media. The closest quantitative anchors are forward-model errors from Three-Dimensional Optical Diffraction Tomography With Lippmann-Schwinger Model: a relative error of $9.01\times10^{-6}$ for the Lippmann–Schwinger model, against $1.82\times10^{-4}$ for Rytov and $2.45\times10^{-5}$ for the beam propagation method, on a red blood cell of index 1.05. These concern the forward model, not passive imaging.

Two ablations isolate the proposal's ingredients. The first keeps the moment loss but uses a Born target. The second keeps everything but replaces the Itô–Schrödinger transport kernel with a homogeneous-medium kernel; this tests whether knowing $\gamma$ matters. A third check deliberately mis-sizes the averaging windows by factors of 0.3 and 3 around the predicted $X_d$ and $\Omega_d$.

Why not just…?

Why not just keep Born and two-point interferometry, which are cheaper? For weak targets this is the right choice, and we expect it to match our method below a few percent contrast. For stronger targets, linear inversions degrade sharply with contrast. Accelerated Image Reconstruction for Nonlinear Diffractive Imaging shows this on Shepp–Logan phantoms, and Full wave 3D inverse scattering transmission ultrasound tomography in the presence of high contrast shows geometric distortion near bone. The contrast sweep below measures the crossover point directly.

Why not just average the correlations over a longer recording and use CISOR with an $\ell_2$ loss? This is the real threat. Longer recording reduces noise from the source distribution, but it does not remove speckle caused by the medium. The random slab is fixed during recording, so its phase scrambling is a bias in the data, not noise that averages away. Averaging over time therefore does not substitute for averaging over position and frequency. The experiment that settles this sweeps recording duration over two orders of magnitude and holds the medium fixed. If the $\ell_2$ variant converges to our error as duration grows, the moment loss is unnecessary and we will say so.

Why not just train a neural operator to invert the correlations? Ambient Noise Full Waveform Inversion with Neural Operators inverts ambient-noise data 285x faster per iteration than a spectral-element solver. It needed 27,000 training simulations, stays in 2D, and does not extrapolate beyond its training frequencies or domain sizes. Our moment model carries $\gamma$ and $L$ as explicit parameters. We will test generalization to unseen $\gamma$ and $L$ against the operator's reported training scope.

Why not just invert the random medium and the target jointly with full-wave inversion? A single realization of the medium has far more unknowns than the passive data can constrain, and inverting it is severely non-convex. The moment approach replaces those unknowns with two statistical parameters, $\gamma$ and the correlation length, which we estimate from the frequency decorrelation of the correlations themselves.

Proposed Experiments

The main experiment sweeps target contrast from 0.2% to 32.2% on the Shepp–Logan phantom, mirroring the CISOR protocol. Each setting uses 50 independent realizations of the random slab, and we report mean and standard deviation of relative permittivity error for every baseline.

The resolution experiment places two cylinders at decreasing separations and finds the separation at which they can still be distinguished. It repeats this at contrasts of 1%, 10%, and 25% while varying $L$ and $\gamma$. We then check whether that separation follows $\lambda L_t / b_{eff}$ with $b_{eff}^2 = b^2+\gamma^2L^3/3$. The Born Sparrow limit of $0.344\lambda$ for point scatterers, from Quantitative Analysis of Super Resolution in Electromagnetic Inverse Scattering for Microwave Medical Sensing and Imaging, serves as the homogeneous-medium reference.

To check the target model against real physics, we use measured target scattering from the Fresnel FoamDielExtTM data (the same data used by SEAGLE) and propagate it synthetically through simulated random slabs. Two scenarios would falsify the claim. In the first, the moment loss gives no variance reduction over time-averaged $\ell_2$. In the second, resolution in the isotropic regime shows no dependence on $\gamma$. The layered-slab runs define the expected failure region. There, both our method and Kirchhoff should lose resolution as the theory predicts, and the proposal succeeds only if our method degrades no worse than the baselines.