QuantFormer: A Hybrid Quantum Classical Transformer for Hyperspectral Image Classification

Saad Ahmed, Jay Lunia
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:103-114, 2026.

Abstract

Hyperspectral image (HSI) classification is challenging because each pixel has hundreds of spectral bands while only a small number of labelled samples are available. This paper presents QuantFormer, a hybrid quantum–classical transformer that embeds a small variational quantum circuit as a spectral token encoder inside a vision transformer backbone for pixel-wise land-cover mapping. A unified patch-based pipeline with band-wise normalization, principal component analysis, and quantum token encoding is evaluated on four benchmarks: Indian Pines, Pavia University, a 7-class Houston 2013 subset, and EuroSAT_MS. With roughly 35k trainable parameters, QuantFormer attains overall accuracy above 99% on the three airborne hyperspectral datasets and about 89.8% on EuroSAT_MS, competitive with deep 3D CNNs while using substantially fewer weights. Beyond full-data experiments, we also study limited-label regimes and provide practical guidance on when quantum token encoders are a viable alternative to classical projections, without claiming quantum advantage over the strongest classical baselines.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-ahmed26a, title = {QuantFormer: A Hybrid Quantum Classical Transformer for Hyperspectral Image Classification}, author = {Ahmed, Saad and Lunia, Jay}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {103--114}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/ahmed26a/ahmed26a.pdf}, url = {https://proceedings.mlr.press/v318/ahmed26a.html}, abstract = {Hyperspectral image (HSI) classification is challenging because each pixel has hundreds of spectral bands while only a small number of labelled samples are available. This paper presents QuantFormer, a hybrid quantum–classical transformer that embeds a small variational quantum circuit as a spectral token encoder inside a vision transformer backbone for pixel-wise land-cover mapping. A unified patch-based pipeline with band-wise normalization, principal component analysis, and quantum token encoding is evaluated on four benchmarks: Indian Pines, Pavia University, a 7-class Houston 2013 subset, and EuroSAT_MS. With roughly 35k trainable parameters, QuantFormer attains overall accuracy above 99% on the three airborne hyperspectral datasets and about 89.8% on EuroSAT_MS, competitive with deep 3D CNNs while using substantially fewer weights. Beyond full-data experiments, we also study limited-label regimes and provide practical guidance on when quantum token encoders are a viable alternative to classical projections, without claiming quantum advantage over the strongest classical baselines.} }
Endnote
%0 Conference Paper %T QuantFormer: A Hybrid Quantum Classical Transformer for Hyperspectral Image Classification %A Saad Ahmed %A Jay Lunia %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-ahmed26a %I PMLR %P 103--114 %U https://proceedings.mlr.press/v318/ahmed26a.html %V 318 %X Hyperspectral image (HSI) classification is challenging because each pixel has hundreds of spectral bands while only a small number of labelled samples are available. This paper presents QuantFormer, a hybrid quantum–classical transformer that embeds a small variational quantum circuit as a spectral token encoder inside a vision transformer backbone for pixel-wise land-cover mapping. A unified patch-based pipeline with band-wise normalization, principal component analysis, and quantum token encoding is evaluated on four benchmarks: Indian Pines, Pavia University, a 7-class Houston 2013 subset, and EuroSAT_MS. With roughly 35k trainable parameters, QuantFormer attains overall accuracy above 99% on the three airborne hyperspectral datasets and about 89.8% on EuroSAT_MS, competitive with deep 3D CNNs while using substantially fewer weights. Beyond full-data experiments, we also study limited-label regimes and provide practical guidance on when quantum token encoders are a viable alternative to classical projections, without claiming quantum advantage over the strongest classical baselines.
APA
Ahmed, S. & Lunia, J.. (2026). QuantFormer: A Hybrid Quantum Classical Transformer for Hyperspectral Image Classification. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:103-114 Available from https://proceedings.mlr.press/v318/ahmed26a.html.

Related Material