SAR and InSAR Change Detection with Quantum Generative Models

AI summary

Overview

Research area: Quantum machine learning applied to remote sensing — specifically, using a quantum circuit Born machine (QCBM) as the background estimator inside synthetic aperture radar (SAR) and interferometric SAR (InSAR) change-detection pipelines, executed on IonQ trapped-ion hardware.

Technical level: Intermediate. Readers should be comfortable with basic SAR/InSAR concepts (amplitude versus phase/coherence), elementary probability (conditional expectation, copulas, KL divergence), and the general idea of parameterized quantum circuits.

Scope: The paper compares a QCBM-based generative background estimator against two classical baselines (NLCD and Copula) on one X-band SAR amplitude pair and one InSAR coherence pair from Capella Space, in ideal simulation, noisy simulation, and on IonQ QPUs.

What This Paper Is About

Change detection asks which pixels in a scene changed between two radar acquisitions. The standard approach estimates what the "after" image should have looked like had nothing changed, by reading a conditional average out of a joint histogram of before/after pixel values — and that histogram becomes sparse and unreliable at sub-meter resolution, where pixel statistics are heavy-tailed. The paper's goal is to replace that empirical conditional with one drawn from a trained quantum generative model, so that well-defined background estimates exist even in histogram bins the data never populated.

Key Contributions

1. A QCBM-based background estimator in copula space. The joint distribution of before- and after-image intensities is modeled through its copula, represented by a quantum circuit with two $n_{bits}$-qubit registers (2 $n_{bits}$ qubits total), trained by minimizing KL divergence with a classical optimizer. Conditional expectations are then read from samples generated by the trained circuit rather than from observed data.

2. End-to-end execution on trapped-ion hardware. Training and inference are demonstrated on IonQ's Forte and Forte-Enterprise QPUs, with a systematic decomposition of the four ideal/QPU combinations for training and inference, isolating the cost of hardware execution at each stage.

3. Benchmarking against two classical baselines across preprocessing regimes. The QCBM is compared with the nonlinear background estimator (NLBE, renamed NLCD in the paper) and a classical copula variant on the Miramar1 SAR dataset under six preprocessing configurations designed to span non-Gaussian and approximately Gaussian pixel distributions, plus an InSAR lava-flow dataset.

4. Diagnostic experiments isolating the source of the quantum advantage. These include a kernel density estimation (KDE) ablation for the classical baselines, a discretization-resolution sweep from 4 to 24 qubits, a cross-chip transfer test without retraining, and a correlation of performance gain with measured joint-histogram occupancy.

Main Findings

• QCBM leads on non-Gaussian SAR data (ideal simulator). On Miramar1: Original 0.30 versus 0.14 (NLCD) and 0.20 (Copula); Original-Med3 0.41 versus 0.16 and 0.24; NG-Med5 0.36 versus 0.21 and 0.26; NG-Med7 0.32 versus 0.25 and 0.29. All values are best filtered F1 after uniform filtering.

• The advantage disappears when the distributions are made Gaussian. On YJ-Med3, QCBM and Copula both reached 0.56 and NLCD 0.55; on YJ-Med7, QCBM and Copula both reached 0.54 and NLCD 0.53.

• Fully hardware-executed results on the airport dataset. For QPU training and inference on Miramar1-Med3, the maximized filtered F1 was 0.32, versus 0.16 and 0.24 for the two classical baselines (as stated in the abstract). A separate QPU-inference run with a 10-bit-per-variable (20-qubit) model reached 0.37 versus 0.16 and 0.24.

• Inference is more robust to hardware noise than training. Starting from 0.40 with ideal training and ideal inference, moving inference to the QPU cost about 0.03, while moving training to the QPU cost about 0.08. The paper attributes the training degradation partly to shot count: ideal training used 100,000 shots per iteration, QPU training 5,000, so finite-sampling noise in the KL objective is entangled with hardware noise.

• InSAR lava flow: parity at peak, wider operating range for QCBM. QCBM (11 bits per variable, 22 qubits, ideal training with QPU inference), NLCD, and Copula all reached a maximum filtered F1 of approximately 0.66. The QCBM curves held near-maximum F1 over a broader interval of detection thresholds, and the QPU-inference curve was shifted relative to the ideal-simulator curve, indicating hardware affected threshold calibration more than peak score.

• Performance scales with circuit size. Mean QCBM filtered F1 rose from about 0.09 at 2 bits per variable (4 qubits) to a maximum of about 0.36 at 11 bits per variable (22 qubits), then decreased modestly at 12 bits. QCBM overtook NLCD (about 0.16) at around 6 bits per variable (12 qubits), and led Copula at every circuit size tested; Copula rose from about 0.03 to about 0.24 at 11 bits per variable.

• The KDE ablation does not rescue the classical methods. Replacing the lookup-table conditional estimator with gaussian_kde reduced the best filtered F1 on Miramar1-Med3 to 0.05 for both NLCD and Copula, ruling out an overly simple classical estimator as the explanation for the gap.

• Cross-chip transfer without retraining. A 10-bit-per-variable (20-qubit) model trained only on Miramar2 and applied directly to Miramar1 (Original) achieved 0.27, above NLCD at 0.14 and Copula at 0.20 on that dataset.

• Gain tracks histogram sparsity. Occupied bins in the joint histogram of real-space pixel intensities were 1.0% (Original), 0.80% (Original-Med3), 4.6% (NG-Med5), and 4.14% (NG-Med7), versus 17.1% (YJ-Med3) and 12.7% (YJ-Med7). The QCBM advantage appears precisely where occupancy is lowest.

Methodology in Plain English

The detector needs a "background" image: for each pixel value $x$ in the before image, what value $y$ would the after image most likely take if nothing changed? The classical answer is to bin all coregistered pixel pairs into a 2D histogram and average the $y$ values in each $x$ bin. When the histogram is sparse, many bins are empty or contain a handful of samples, and the averages become noisy or undefined.

The authors instead learn the joint distribution of $x$ and $y$ with a quantum generative model. Pixel intensities are mapped into copula space through their marginal cumulative distribution functions and discretized, which separates the dependence structure from the individual marginals. A parameterized quantum circuit with two entangled registers is trained so that its measurement statistics match that discretized target distribution, using KL divergence as the loss and a classical optimizer in the loop. After training, measuring the circuit produces unlimited samples; these are mapped back to intensity space, filtered by the observed $x$, and averaged to give the background estimate for each pixel. Comparison of background to observation then yields a change mask, which is smoothed with scipy.ndimage's uniform_filter and thresholded, and scored by filtered F1.

Two real datasets anchor the evaluation. The SAR pair is X-band Stripmap, HH polarization, 1.2 m nominal resolution, from Capella-2 ("Sequoia") over Marine Corps Air Station Miramar on 22 and 29 October 2020 at incidence angles of 32.89° and 36.60°, cropped to 500 m × 500 m, despeckled with a 3 × 3 refined Lee filter and coregistered with GeFolki. The InSAR pair is coherence magnitude from Capella-13 (53° inclination, 2.95-day repeating ground track) over Piton de la Fournaise, using a pre-event 3–6 February 2026 pair and a 6 February–8 March 2026 pair that spans the eruption, with coherence estimated over a 4 × 6-pixel window.

Why This Matters

Research impact. The paper reframes where quantum generative models might be useful: not as a drop-in replacement for classical estimators in well-sampled regimes, but as a way to supply well-defined conditional expectations where classical statistics run out of data. The correlation between measured histogram occupancy and the size of the QCBM's advantage is a testable hypothesis about the conditions for quantum utility, and the ideal-versus-QPU decomposition gives a clean accounting of where hardware noise actually costs performance.

Real-world applications (as framed in the paper):

• Disaster response and damage assessment after events such as the Piton de la Fournaise lava flow.
• Monitoring of the built environment and settlement growth.
• Land-use enforcement.
• Defense and intelligence uses: wide-area activity monitoring, treaty verification, and search for objects or infrastructure of interest, independent of weather and illumination.

Industry relevance. The work is a collaboration between IonQ and Capella Space, using commercially acquired satellite imagery and commercial trapped-ion QPUs. Demonstrating that a trained model can transfer to an unseen image chip without retraining (0.27 versus 0.14 and 0.20 for the classical baselines) and can run inference on hardware while retaining an advantage speaks directly to operability rather than purely idealized benchmarking.

Future Directions

• Conditioning on multiple reference acquisitions. The authors note this raises the dimensionality of the joint distribution and empties the histogram far faster than finer discretization does, while costing only linearly more qubits — potentially a stronger test of the sparsity argument.

• A circular copula for interferometric phase. Phase is wrapped onto a circle, and the copula construction used here is not defined for such variables; a circular formulation would allow phase and coherence to be modeled together rather than through intensity alone.

• More efficient conditional sampling. Conditioning is currently done by rejection, whose cost grows as the conditioning becomes rarer — exactly the sparse regime where the method works best.

• Broader characterization. The authors call for a wider survey of classical density estimators (including structurally constrained generative models comparable to the QCBM), matched shot counts between ideal and QPU training to separate hardware noise from finite-sampling noise, and a larger image collection to establish how far cross-chip transfer extends across regions and acquisition conditions.

Target Audience

Quantum machine learning researchers looking for a concrete, hardware-executed application with classical baselines; remote sensing and SAR/InSAR practitioners interested in what quantum generative models can and cannot add to change detection; and applied scientists in defense, disaster response, or environmental monitoring who need to judge whether the reported gains are relevant to their operating regimes. Readers focused on large-scale quantum advantage claims should note that the demonstrations use up to 24 qubits and that the benefit is concentrated in sparse, non-Gaussian settings rather than being universal.

Authors’ abstract

Change detection in synthetic aperture radar (SAR) and interferometric synthetic aperture radar (InSAR) underpins disaster response, infrastructure monitoring and land-use enforcement. Detection is limited by the background estimator, which conventionally forms a conditional expectation directly from observed pixel statistics and degrades where those statistics are sparse, including the regime produced by the heavy-tailed marginals of sub-meter-resolution radars. In this work, we integrate state-of-the-art satellite imagery with quantum machine learning on IonQ trapped-ion-based quantum processors. By replacing the empirical conditional with a quantum circuit Born machine (QCBM)-sampled generative model in Copula space, we substantially improve change detection on sparse real-world images. On Capella Space satellite image acquisitions, the generative estimator matches conventional methods when the observed statistics are adequate, and substantially outperforms them when they are not. Executing the trained model on IonQ trapped-ion based hardware reproduces the results of the ideal and noisy simulations and demonstrates up to par, or even better, performance with the classical state-of-the-art methods. For a SAR dataset of an airport, QPU circuit evaluations for both training and inference achieved a maximized filtered F1 score of 0.32, compared with 0.16 and 0.24 for the two classical baselines. For an InSAR dataset of a volcanic lava flow, all three methods reached a maximum filtered F1 of approximately 0.66. These experiments demonstrate the feasibility of executing a QCBM-based background estimator on trapped-ion hardware. We further demonstrate that the QCBM method successfully extends to interferometric coherence data, achieving performance comparable to classical approaches.