Project Description

Abstract

QUAI4EO 1 was a joint research project between CERN, ESA Φ-lab, and EPFL, running from January 2021 to September 2024 and carried out by PhD candidate Su Yeon Chang. Its objective was to establish whether quantum machine learning (QML) can be applied in practice to the analysis of Earth observation imagery on today's noisy intermediate-scale quantum (NISQ) hardware.

Three challenges framed the work. Images are intrinsically multi-dimensional data, which makes them difficult to represent on devices offering a limited number of qubits. The practical usability of quantum algorithms and quantum machine learning models had to be demonstrated on real tasks rather than assumed. And the data embedding step, together with hardware-level limitations, acts as a bottleneck, constraining what can realistically be achieved on current processors.

The project addressed these challenges through two complementary workstreams within a hybrid computer-vision pipeline. The first developed quantum generative models for image generation, resulting in a latent style-based quantum GAN in which a compact quantum generator operates in the latent space of a pretrained autoencoder and is trained against a classical discriminator; the architecture produces high-quality Earth observation and benchmark imagery. The second developed a hybrid classical–quantum convolutional neural network for multi-class classification, evaluated on EuroSAT land-cover data, achieving per-class accuracies between 0.77 and 0.99 across the AnnualCrop, Forest, Herbaceous Vegetation, and Industrial classes.

The main result of the collaboration is the demonstration of the practical usability of quantum neural networks as generators, classifiers, and regressors for image data on NISQ hardware. The work was published as a peer-reviewed conference contribution at IGARSS 2022 and as a preprint on image generation in 2024.

Detailed description

Context and motivation

Earth observation generates large volumes of multidimensional imagery, and the analysis of that imagery is increasingly dominated by machine learning. Quantum computing has been proposed as a route to new model families for such data, but most proposals have been assessed analytically or in idealized simulation rather than on hardware. QUAI4EO 1 was set up to test the practical side of that proposition: whether quantum neural networks can be built, trained, and evaluated on realistic image data using devices available today.

The project was hosted by CERN in collaboration with ESA Φ-lab and EPFL, and was carried out as a doctoral research project by Su Yeon Chang between January 2021 and September 2024.

Challenges addressed

• Images are intrinsically multidimensional data, so mapping them onto a quantum register is not straightforward and dominates the resource budget.
• The practical usability of quantum algorithms and quantum machine learning had to be demonstrated, not assumed — including training stability and the quality of the outputs obtained.
• Limitations in the data embedding step and at the hardware level (qubit count, noise, circuit depth) bound what can be executed on NISQ devices.

Workstream 1 — Quantum generative models for image generation

The first workstream focused on image generation using quantum generative models. The developed architecture is a latent style-based quantum GAN: rather than generating pixels directly, a quantum generator produces features in the latent space of a pretrained autoencoder, and the autoencoder's decoder reconstructs the image. A classical discriminator is trained against the quantum generator, and the generator parameters are updated through classical optimization.

This design keeps the quantum circuit compact — it only needs to cover the latent representation rather than the full image, which makes the approach compatible with the qubit counts and circuit depths available on NISQ hardware. The resulting model produces high-quality imagery on both Earth observation and standard benchmark datasets, and was published as “Latent Style-based Quantum GAN for High-quality Image Generation” (arXiv:2406.02668, 2024).

Workstream 2 — Hybrid quantum CNN for multi-class classification

The second workstream addressed supervised classification of Earth observation imagery with a hybrid classical–quantum convolutional neural network, in which quantum layers are embedded in an otherwise classical CNN. The model was evaluated on the EuroSAT land-cover dataset in a four-class setting: AnnualCrop, Forest, Herbaceous Vegetation, and Industrial.

The confusion matrices obtained show per-class accuracies ranging from 0.77 to 0.99, with the strongest performance on the Forest and Industrial classes and the main residual confusion between Herbaceous Vegetation and the other vegetated classes. The work was presented as “Multi-class classification with hybrid classical-quantum CNN” at IGARSS 2022.

Outcome and significance

Taken together, the two workstreams demonstrate the practical usability of quantum neural networks as generators, classifiers, and regressors for image data on NISQ hardware. The contribution of the collaboration is therefore twofold: concrete hybrid architectures that make image-scale problems tractable on current quantum devices, and evidence-based trained models, accuracy figures, and generated imagery, for assessing where quantum machine learning stands today with respect to Earth observation tasks.

Project Outcomes

• Latent style-based quantum GAN — implementation of the quantum generator and pretrained-autoencoder pipeline for image generation (Chang, S. Y. et al., “Latent Style-based Quantum GAN for High-quality Image Generation”, arXiv:2406.02668 (2024))
• Hybrid classical–quantum CNN — implementation and EuroSAT multi-class classification experiments (repository link to be added) (Chang, S. Y. et al., “Multi-class classification with hybrid classical-quantum CNN”, IGARSS 2022)
• PhD thesis of Su Yeon Chang (EPFL), completed September 2024.

Outcome of the collaboration

The collaboration delivered quantum neural network models that are usable in practice on NISQ hardware for Earth observation imagery, covering generation, classification, and regression. The two flagship results are a latent style-based quantum GAN that produces high-quality images from a compact quantum generator, and a hybrid classical–quantum CNN that achieves 0.77–0.99 per-class accuracy on EuroSAT land-cover classification.