⚛️ quantum physics

SAR and InSAR Change Detection with Quantum Generative Models

This paper demonstrates that integrating quantum machine learning, specifically a Quantum Circuit Born Machine (QCBM) executed on IonQ trapped-ion hardware, significantly improves SAR and InSAR change detection performance over classical methods, particularly in scenarios with sparse data and heavy-tailed statistics.

Original authors: Samwel K. Sekwao, Shaunak De, Alexis Hocken, Scott Staniewicz, Evgeny Epifanovsky, Craig Stringham, Gordon Farquharson, Martin Roetteler, Panagiotis Kl. Barkoutsos, Jason Iaconis

Published 2026-09-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Samwel K. Sekwao, Shaunak De, Alexis Hocken, Scott Staniewicz, Evgeny Epifanovsky, Craig Stringham, Gordon Farquharson, Martin Roetteler, Panagiotis Kl. Barkoutsos, Jason Iaconis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Satellite radar offers a unique way to watch the Earth, seeing through clouds and darkness to map the ground with precision. Unlike cameras that rely on sunlight, these sensors send out their own radio waves and listen for the echo, creating a picture of the landscape based on how rough or smooth the surface is. This technology is vital for spotting changes: a new building, a flooded street, or a shifting volcano. However, the radar signal is notoriously noisy, filled with a grainy static called speckle that makes it hard to tell if a pixel has changed because the ground moved or simply because the signal fluctuated. To solve this, scientists usually compare a new image to an old one, trying to predict what the new image should look like if nothing had changed. When the data is plentiful, this prediction works well. But when the images are extremely detailed, the data becomes so sparse and irregular that traditional computer methods struggle to make a reliable guess, often missing real changes or flagging harmless noise as a disaster.

A team of researchers has now tested a different approach, using a new type of computer that operates on the principles of quantum mechanics to build a better prediction model. Instead of relying on standard statistical tricks to smooth out the noisy radar data, they trained a quantum machine learning model to understand the complex relationship between the old and new images. They tested this system on real satellite data from Capella Space, looking at two very different scenarios: a busy airport in San Diego and a volcanic eruption on a French island. In the case of the airport, where the radar data was messy and difficult to interpret, the quantum model significantly outperformed the best existing classical methods. It was better at ignoring the background noise and highlighting the actual changes, such as aircraft movement or construction. When they moved the trained model onto actual quantum hardware to run the test, the results held up, proving that these quantum computers can handle real-world tasks even while they are still in their early stages of development.

The researchers focused on a specific challenge found in high-resolution radar images. When you zoom in close enough to see individual cars or small buildings, the statistical patterns of the signal become extremely skewed, with most values clustered in one area and a few extreme outliers stretching far away. Traditional methods try to fix this by smoothing the data, but that blurs the fine details that make high-resolution imaging useful in the first place. The quantum team avoided this trade-off by using a generative model, which is a type of artificial intelligence that learns the underlying rules of the data and can create new, realistic examples of what the image should look like. They used a specific architecture called a quantum circuit Born machine, which uses the unique properties of quantum bits to learn the complex connections between the pixels of the before and after images. By training this model on the joint distribution of the data, it learned to predict the background with high accuracy, even in the regions where the data was too sparse for conventional computers to make a confident call.

The experiments were conducted on two distinct datasets to ensure the method was robust. The first was a pair of images of Marine Corps Air Station Miramar in San Diego, taken at a resolution of 1.2 meters. The second dataset captured the aftermath of a volcanic eruption at Piton de la Fournaise, where the researchers looked for changes in the coherence of the radar signal caused by fresh lava flows. For the airport data, the quantum model achieved a score of 0.32 in identifying changes, which was notably higher than the 0.16 and 0.24 scores achieved by the two leading classical methods. This improvement was most pronounced when the data was difficult and non-standard, suggesting the quantum model excels where traditional statistics fail. For the volcanic lava flow, all three methods reached a similar peak performance of about 0.66, indicating that the quantum approach is at least as good as the best classical tools in easier scenarios, but superior when the data is messy.

To ensure these results were not just a fluke of the simulation, the team ran the trained model on actual quantum hardware provided by IonQ. They tested the system under different conditions, running the training on the quantum processor, the inference on the processor, or both. While the hardware introduced some noise that slightly lowered the scores compared to a perfect simulation, the quantum model still outperformed the classical baselines in every configuration. This demonstrated that the advantage is real and not dependent on a perfect, error-free environment. The researchers also tested whether a model trained on one part of the airport could be applied to a different part without retraining. The quantum model successfully transferred its knowledge, maintaining a higher performance than the classical methods, which suggests it has learned general rules about the environment rather than just memorizing specific pixels.

The study also carefully ruled out alternative explanations for the success. The researchers tested whether the quantum advantage came simply from using a more sophisticated way to smooth the data. They replaced the standard smoothing method used by the classical models with a more advanced statistical technique called kernel density estimation. This change did not improve the classical models; in fact, it made their performance worse. This confirmed that the quantum model's success was not due to a simple flaw in the classical smoothing process, but rather a genuine capability of the quantum approach to learn from sparse, complex data. The results suggest that the quantum model is particularly effective when the data is sparse, filling in the gaps where traditional methods leave the prediction undefined or noisy.

Looking ahead, the researchers see potential for this technology to scale. The current experiments used a quantum circuit with 20 qubits, a size that is manageable for today's hardware but small compared to the complexity of full satellite imagery. The study showed that as the resolution of the model increased, its performance improved, hinting that larger quantum computers could handle even more detailed data. The team also noted that while their current method works well for standard radar intensity, future work could extend these techniques to handle the circular nature of phase data used in interferometry, which would allow for even more precise measurements of ground movement. For now, the work stands as a demonstration that quantum machine learning can move beyond theory and solve practical problems in remote sensing, offering a new tool for monitoring our changing planet.

Technical Summary: SAR and InSAR Change Detection with Quantum Generative Models

Problem Statement
Synthetic Aperture Radar (SAR) and Interferometric SAR (InSAR) change detection are critical for disaster response, infrastructure monitoring, and land-use enforcement. The core challenge lies in the "background estimator," which must predict what a scene would look like in a second acquisition if no change had occurred. Conventional methods estimate this conditional expectation, $E[y|x]$ , empirically by accumulating joint histograms of coregistered pixel pairs.

This approach degrades significantly in high-resolution regimes (e.g., sub-meter X-band imagery) where pixel statistics are sparse, heavy-tailed, and strongly skewed. In these conditions, the joint histogram consists of a narrow, sparsely sampled ridge rather than a dense cloud, leaving many conditioning bins empty or populated by only a few samples. Consequently, the empirical conditional means are noisy or undefined, particularly in the distribution tails where low false-alarm detection operates. While classical remedies involve regularizing statistics via variance-stabilizing transformations and spatial filtering, these processes often sacrifice the fine spatial detail that high-resolution imagery provides.

Methodology
The authors propose replacing the empirical conditional estimator with a generative model trained on the joint distribution of pixel intensities in Copula space. The specific implementation utilizes a Quantum Circuit Born Machine (QCBM) executed on IonQ's trapped-ion quantum processors.

Copula Transformation: The method transforms pixel intensities into Copula space using marginal cumulative distribution functions (CDFs). This separates the dependence structure between the "before" and "after" images from their individual marginal distributions, allowing the model to focus on statistical dependencies.
QCBM Architecture: The QCBM employs two quantum registers (one for the before-image intensity, one for the after-image), totaling $2n_{bits}$ qubits. The circuit initializes maximally entangled states to ensure uniform marginals, followed by parameterized unitary gates ( $R_z, R_x, R_{ZZ}$ ) to learn the joint dependence structure.
Training and Inference: The model is trained by minimizing the Kullback–Leibler (KL) divergence between the target distribution (derived from data) and the circuit-generated distribution using a classical optimizer (COBYLA for simulation, SPSA for hardware). Post-training, the model generates samples from the learned joint distribution. Conditional sampling is performed via rejection sampling to estimate the expected background intensity for observed reference pixels.
Experimental Setup: The study evaluates the method on two Capella Space datasets:
- SAR Amplitude: High-resolution (1.2 m) X-band imagery of Marine Corps Air Station Miramar.
- InSAR Coherence: Interferometric coherence maps of the 2026 Piton de la Fournaise volcanic eruption.
- Baselines: The QCBM is compared against a Non-Linear Background Estimator (NLBE) using lookup tables and a classical Copula-based estimator.

Key Results
The experiments demonstrate that the QCBM-based estimator matches or outperforms classical baselines, with the magnitude of improvement dependent on data sparsity and distribution characteristics.

Non-Gaussian Data (Miramar SAR): On datasets with strongly skewed, heavy-tailed distributions (Original, Original-Med3, NG-Med5, NG-Med7), the QCBM substantially outperformed classical methods.
- On the "Original" Miramar dataset, the QCBM achieved a filtered F1 score of 0.30, compared to 0.14 for NLBE and 0.20 for the Copula baseline.
- On the "Original-Med3" dataset, QCBM reached 0.41, versus 0.16 (NLBE) and 0.24 (Copula).
- The advantage persisted across varying discretization resolutions, with QCBM performance increasing as resolution (qubit count) increased, peaking at 22 qubits (11 bits per variable) with a score of ~0.36.
Gaussian Data: When preprocessing (Yeo-Johnson transformation) rendered the data approximately Gaussian, the performance gap narrowed. In these regimes, QCBM, Copula, and NLBE achieved comparable results (F1 scores ~0.54–0.56).
InSAR Volcanic Data: On the lava flow dataset, all three methods reached a maximum filtered F1 of approximately 0.66. However, the QCBM maintained high performance over a broader range of detection thresholds, suggesting greater robustness to threshold selection.
Hardware Execution: The study successfully executed the trained model on IonQ's Forte-Enterprise QPU.
- Inference on QPU hardware (with ideal training) resulted in a modest F1 reduction (0.40 to 0.37).
- Training on QPU hardware resulted in a larger reduction (to 0.32), attributed to a combination of hardware noise and the lower shot count (5,000 shots vs. 100,000 in simulation) used to manage execution time.
- Despite hardware noise, the QPU-executed QCBM (F1 0.32) still significantly outperformed the classical baselines (0.16 and 0.24).
Generalization: A cross-chip experiment showed that a model trained on one geographic chip (Miramar2) could be applied to a different chip (Miramar1) without retraining, achieving an F1 of 0.27, which remained superior to the classical baselines (0.14 and 0.20).

Significance and Claims
The paper claims to demonstrate the feasibility of executing a QCBM-based background estimator on trapped-ion hardware for real-world remote sensing tasks. The primary significance lies in addressing the "sparse data" regime inherent to high-resolution SAR, where classical empirical estimators fail due to insufficient statistics.

The authors posit that the QCBM's advantage stems from its ability to learn and sample from the joint distribution without requiring dense bin occupancy, effectively treating the sparsity of the conditional histogram as a resource where quantum generative models can outperform classical smoothing techniques (as confirmed by an ablation study where Kernel Density Estimation failed to close the performance gap). The work establishes that quantum generative models can provide competitive or superior change detection performance, particularly in non-Gaussian, sparse regimes, and that this performance is robust enough to survive the transition from ideal simulation to noisy intermediate-scale quantum (NISQ) hardware.

More like this