Source-linked AI summary
Self-Supervised Representation Learning: From Spectral Foundation Models to Auroral Emission Spectra
Matthieu Le Lain, Gaël Cessateur, Sébastien Lefèvre
TL;DR
Auroral spectrographs collect many emission spectra, but only a small fraction receive expert labels, and transfer from astronomical spectral models had not been measured. The paper pretrains a masked autoencoder on auroral spectra and compares it with expert features and existing models. In-domain pretraining matches expert features under a linear probe and improves fine-tuning, while tested external models do not replace it.
Problem
ASIS has recorded nearly 330,000 spectra, but only 811 were labelled, and transfer of astronomical spectral models to atmospheric emission spectra had not been measured.
Source: §1 · p. 1Method
The authors pretrained a 1D Vision Transformer with a masked autoencoder on 223,000 unlabelled auroral spectra and evaluated its representations and fine-tuned classifier.
Source: §3.2 · p. 2, §5 · p. 4Results
Takeaways & Limitations
Among the tested models, transfer follows pretraining wavelength overlap rather than source type, a prediction the authors say remains to be tested across instruments.
Source: §1 · p. 1, §5 · p. 4The wavelength-overlap explanation is a testable prediction, not a proven law, and remains to be studied across instruments.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Auroral spectrographs such as the Auroral Spectrograph In Skibotn (ASIS) record hundreds of thousands of emission spectra, but only a few hundred can be labelled by an expert. To exploit the rest, we pretrain a 1D Vision Transformer with a masked autoencoder on 223,000 unlabelled spectra. Without labels, its representation recovers the emission-line intensity ratios that physicists use to diagnose the precipitating particles (R^2 0.91 vs. 0.77 for an untrained control) and, under one linear probe, classifies as well as 13 features designed by experts. Fine-tuned, the model outperforms the previous supervised auroral classifier on its own benchmark (macro-AP 88.5 vs. 77.8), reaches 0.870 mAP, and exceeds the same architecture trained from scratch by +0.159 with 10% of the labels; attribution shows that it uses both N2+ bands. Could an existing pretrained model replace it? Two astronomical spectral foundation models and a time-series model transfer according to their spectral window: SpectraFM, trained in the infrared, falls below the untrained control, whereas SpecFormer, trained in the optical, approaches in-domain pretraining without reaching it.
1. INTRODUCTION
Auroral spectra encode particle-precipitation signatures, but expert labelling is limited. This paper tests self-supervised pretraining and whether existing astronomical spectral models transfer to auroral emission spectra.
1. INTRODUCTION
- Only 811 of nearly 330,000 ASIS spectra had been labelled after three observing seasons, despite line intensities and ratios diagnosing particle precipitation.The 630/557.7 nm oxygen-line ratio traces precipitating-electron energy, while hydrogen Balmer lines signal proton precipitation.Source: §1 · p. 1
- Whether astronomical spectral foundation models transfer to atmospheric emission spectra had not been measured, leaving open whether a new domain must pretrain its own encoder.The answer bears on whether a new spectroscopic domain can reuse an existing encoder [as framed in the paper].Source: §1 · p. 1
- The paper pretrains an auroral encoder and compares it with two astronomical spectral foundation models and a time-series model under one evaluation protocol.The contribution reports that none replaces in-domain pretraining and that transfer follows overlap between pretraining and auroral wavelength windows rather than source type.Source: §1 · p. 1
- Prior auroral machine learning focused on all-sky image morphology, while spectra resolving emitting species received little attention.The cited prior ViT-1D used 8 coarse 35 nm tokens and did not outperform a single-patch MLP; earlier work had not evaluated transfer to atmospheric emission spectra.Source: §1 · p. 1
2. DATA
The ASIS corpus contains calibrated spectra spanning multiple observing nights, while expert labels cover a small subset. The data section describes preprocessing and the multilabel task.
2. DATA
- ASIS recorded 329,704 calibrated spectra over 220 nights between 2023 and 2026, with each spectrum covering 405–687 nm in 1,024 points.The corpus is split by night; 154 nights are used for pretraining, with 33 validation and 33 test nights held out.Source: §2 · pp. 1–2
- After filtering saturated, null, or non-finite pixels and removing labelled spectra, pretraining uses 223,065 spectra; 97.2% pass the pixel-quality filter.Source: §2 · p. 2
- Each spectrum is represented by a baseline-subtracted, transformed line channel and a separate downsampled continuum channel.The median baseline is preferred for sloped twilight continua, and the separate continuum channel retains scattered-light information used by two classes.Source: §2 · p. 2
- An expert labelled 811 spectra from six nights across seven multilabel classes, including precipitation signatures and spectra dominated by scattered light.The classes include Hα, low-energy electrons, N2 first-negative and first-positive bands, [N I], cloudy, and dawn_dusk; multiple labels may apply to one spectrum.Source: §2 · p. 2
- The 719 spectra labelled with five classes in the earlier study form a subset of the 811 labels and serve as the prior-classifier benchmark.Source: §2 · p. 2
3. METHOD
The method combines fine-grained spectral encoding with masked-autoencoder pretraining, then evaluates frozen representations and fine-tuned classifiers using shared data splits and baselines.
3.1. Encoder and pretraining
- The 1D Vision Transformer converts 1,024 spectral points into 64 tokens spanning 4.4 nm each, with two input channels and eight Transformer blocks.Pretraining on 223,065 unlabelled spectra addresses overfitting observed with these fine tokens under 719 labels.Source: §3.1 · p. 2
- Masked-autoencoder pretraining masks 35% of tokens and reconstructs them with a decoder, using equal-weight mean squared error across the two channels.The 811 labelled spectra are excluded from the objective; they are used to select checkpoints through a linear probe on frozen embeddings.Source: §3.1 · p. 2
- The informed masking prior weights token positions by line-channel variability measured over 20,000 spectra, without using a catalogue or labels.Its highest-variability positions include bright auroral lines; masking probabilities are normalized to a constant global rate.Source: §3.1 · p. 2
3.2. Evaluation protocol
- Representations are evaluated with stratified multilabel five-fold cross-validation repeated three times, using identical folds, seeds, and label subsets.The linear probe uses one logistic regression per class; macro mAP averages precision across the seven classes, weighting dawn_dusk equally.Source: §3.2 · p. 2
- Scores are framed by 13 hand-crafted expert features and a same-architecture encoder with random, untrained weights.The expert features encode catalogue line intensities, diagnostic ratios, and continuum level; the untrained encoder is the control for pretraining.Source: §3.2 · p. 2
3.3. Fine-tuning
- Fine-tuning feeds the class token and mean token representation to a seven-output linear layer; Table 1 reports frozen-representation mAP, including at 10% labels.Training uses class-weighted binary crossentropy, Gaussian-noise augmentation, and AdamW; the table compares representations and identifies their pretraining wavelength windows.Source: §3.3 · p. 2, Table 1 · p. 3
4. RESULTS
In-domain pretraining improves auroral classification, especially with scarce labels, and frozen representations recover expert spectral diagnostics. External-model transfer varies with spectral-window overlap, while attribution identifies physically coherent emission features used by the classifier.
4.1. Transfer of frozen representations
- 0.863 macro mAP: in-domain pretraining raises the frozen linear probe above the untrained control’s 0.816 and ties expert features at 0.861.Concatenating the embedding and expert features yields no gain (0.864, p = 0.30).Source: §4.1 · p. 3
- 0.844: optical-window SpecFormer exceeds the untrained floor but remains below in-domain pretraining; infrared SpectraFM falls below the floor at 0.789.Mantis, pretrained on generic time series, does not rise above the floor (0.824, p = 0.56).Source: §4.1 · p. 3
4.2. Five-class benchmark against the supervised ViT-1D
- 88.5 macro-AP: the fine-tuned model beats the previous supervised ViT-1D’s 77.8 on its five-class benchmark.The same pipeline’s scratch model reaches 83.0, accounting for part of the gain; the pretrained model improves all five classes.Source: §4.2 · p. 3
4.3. Seven-class multilabel task
- 0.870 ± 0.032 mAP: the pretrained model matches the expert-feature score of 0.861 within fold dispersion and exceeds the scratch model’s 0.840.The frozen encoder with a linear head also reaches 0.870 ± 0.034; no per-class difference from expert features exceeds fold dispersion.Source: §4.3 · pp. 3–4, Table 2 · p. 3
- +0.159 mAP over scratch with 10% of labels (0.819 vs. 0.661); the pretrained model remains within dispersion of expert features (0.819 vs. 0.823).With all labels, the gain over scratch is +0.030.Source: §4.3 · p. 4
4.4. Masking ablation
- Masking rates from 0.35 to 0.75 score within [0.856, 0.864] on the frozen probe; informed 0.75 masking cuts pretraining time by 23% at the same quality.The compared runs take 70 minutes for informed 0.75 masking versus 91 minutes for uniform 0.35 masking.Source: §4.4 · p. 4
4.5. Physical content of the representation
- R² 0.91 vs. 0.77 for the untrained control: the frozen embedding recovers five expert line ratios, with the largest gap on the nearly constant 636.4/630 control ratio.For that ratio, R² is 0.72 vs. 0.41.Source: §4.5 · p. 4
- For elow, 56% of attribution falls on the red [O I] doublet; n2_471 and molecular decisions use both N2+ bands rather than only the bands annotators inspect.For n2_471, attribution is 18.2% at 427.8 nm and 16.7% at 470.9 nm; for molecular, the two bands receive 31% versus 4% for the inspected N2 first-positive band.Source: §4.5 · p. 4, Figure 1 · p. 4
5. CONCLUSION
The pretrained auroral encoder captures physics without labels and, when fine-tuned, outperforms the prior supervised classifier and the same architecture trained from scratch. Its strongest label-scarce gain is +0.159 at 10%, while transfer from tested foundation models remains limited and spectral-window overlap is a testable prediction.
5. CONCLUSION
- Fine-tuned, the pretrained encoder outperforms the previous supervised auroral classifier and the same architecture trained from scratch, with attribution showing use of the whole N2 emission system.The reported N2 attribution contrasts with reliance on only the single band an expert checks.Source: §5 · p. 4
- The pretrained representation captures underlying physics without labels and, under one linear probe, classifies as well as 13 expert-designed features.Source: §5 · p. 4
- None of the three tested foundation models replaces in-domain pretraining; their transfer tracks spectral-window overlap, a testable prediction left for study across instruments.Source: §5 · p. 4
- +0.159 at 10% labels is the gain over the same model trained from scratch, and masking rate and an informed static prior change only pretraining time.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The study used only instrument data and required no ethical approval; it received no funding and reports no conflict of interest.