Multi-label Classification of Ground-based Auroral Emission Spectra Using 1D Vision Transformers
Jul 6, 2026·
·
1 min read
Matthieu Le Lain
Abstract
Can Vision Transformer attention act as a built-in interpretability mechanism on ordered 1D scientific signals? We investigate this question through the multi-label classification of ground-based auroral emission spectra acquired in Skibotn, Norway. A 1D Vision Transformer is compared with an MLP baseline on 719 expert-annotated spectra spanning five emission classes. Both models perform strongly, and the attention maps of the ViT-1D recover known emission-line locations without any spectroscopic prior.
Date
Jul 6, 2026 — Jul 8, 2026
Location
Montpellier, France
Poster presented at the RFIAP 2026 conference (Reconnaissance des Formes, Image, Apprentissage et Perception), held jointly with CAp 2026 in Montpellier, France, July 6-8, 2026.
This work, in collaboration with the Royal Belgian Institute for Space Aeronomy and IRISA, introduces the multi-label classification of ground-based auroral emission spectra as a novel task, and shows that the attention maps of a 1D Vision Transformer align with known auroral emission lines. See the companion paper and the blog post for details, or try the trained model directly in the interactive demo.
Authors
AI for astronomy & astrophysics
PhD student at IRISA, Université Bretagne Sud (expected 2026), and lecturer in computer science,
working on foundation models for astronomy and astrophysics, with a broader interest
in deep learning applied to scientific data.
Also contributing to the UniverseTBD collaboration on vision-language models for astronomy.