What the study found
The study found a clear hierarchy in how design choices affect performance in quantum and hybrid convolutional neural networks for satellite image classification. In hybrid architectures, data encoding had the largest effect on validation accuracy, while variational ansätze and measurement basis had much smaller effects. In purely quantum models, performance depended most on the measurement protocol and the data-to-amplitude mapping.
Why the authors say this matters
The authors present these findings as practical guidance for designing parameterized quantum circuits in quantum and hybrid convolutional neural networks. They suggest that knowing which choices matter most can help prioritize model design decisions for satellite image classification.
What the researchers tested
The researchers systematically evaluated about 500 model configurations applied to the EuroSAT satellite image dataset. They compared different data encoding techniques, variational ansätze, and measurement choices in quantum and hybrid convolutional neural network architectures, and benchmarked hybrid models against direct classical equivalents with the parameterized quantum circuits removed.
What worked and what didn't
For hybrid architectures, the data encoding strategy was the dominant factor, with validation accuracy varying by more than 30% across different embeddings. By contrast, variational ansätze and measurement basis changed validation accuracy by less than 5%.
For purely quantum models, which were limited to amplitude encoding, the measurement strategy changed validation accuracy by up to 30%, and the data-to-amplitude mapping changed it by around 8 percentage points.
What to keep in mind
The abstract does not describe limitations beyond the scope of the tested models and dataset. The results are specific to the EuroSAT classification task and the model configurations examined in this study.
Key points
- In hybrid quantum and hybrid convolutional neural networks, data encoding had the biggest effect on validation accuracy.
- Across hybrid models, variational ansätze and measurement basis changed validation accuracy by less than 5%.
- In purely quantum models, measurement protocol had the strongest effect, changing validation accuracy by up to 30%.
- The study tested about 500 configurations on the EuroSAT satellite image classification task.
- Hybrid models were also compared with classical equivalents that removed the parameterized quantum circuits.
Disclosure
- Research title:
- Encoding choices dominate performance in hybrid quantum neural networks
- Authors:
- Jesús Lozano-Cruz, Albert Nieto-Morales, Oriol Balló-Gimbernat, Adán Garriga, Antón Rodríguez-Otero, Alejandro Borrallo-Rentero
- Institutions:
- Universidad de Oviedo, Fundación Centro Tecnológico de la Información y la Comunicación, Instituto Tecnológico de Materiales de Asturias, Universidad Internacional De La Rioja, Universitat Autònoma de Barcelona, Centre Tecnologic de Telecomunicacions de Catalunya, Computer Vision Center, Fujitsu (China)
- Publication date:
- 2026-04-22
- OpenAlex record:
- View
- Image credit:
- Photo by Synth Mind on Unsplash · Unsplash License
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.