Self-Supervised Representation Learning: From Spectral Foundation Models to Auroral Emission Spectra

Matthieu Le Lain; Gaël Cessateur; Sébastien Lefèvre

cs.LGSep 25, 2026

Self-Supervised Representation Learning: From Spectral Foundation Models to Auroral Emission Spectra

Authors: Matthieu Le Lain, Gaël Cessateur, Sébastien Lefèvre

Organizations: IRISA, UMR CNRS 6074, Université Bretagne Sud, Vannes, France · Royal Belgian Institute for Space Aeronomy, Brussels, Belgium · UiT - The Arctic University of Norway, Tromsø, Norway

Abstract

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.

Figures & tables

Figure 1: (a) An ASIS spectrum, the expert’s classes above their criterion. (b) Ridge-regression R 2 of five line ratios read from the frozen embedding, pretrained encoder vs. untrained control (5-fold; ratios clipped at the 1st/99th percentiles; legend: mean of the five). (c) Share of positive IG attribution per window ( $\pm 2$ nm around nine lines; N 2 first-positive (1PG) band: 654–687 nm minus the H $\alpha$ window), per class of the fine-tuned classifier on its positives; rest outside the windows; red frame = annotation criterion; $\star$ = absent from the expert features.

Representation	Window (nm)	mAP	@10 %
Untrained control	—	0.816 $\pm$ 0.034	0.683 $\pm$ 0.036
Expert features	—	0.861 $\pm$ 0.032	0.823 $\pm$ 0.041
Ours (auroral MAE)	405–687	0.863 $\pm$ 0.037	0.829 $\pm$ 0.036
Expert feat. + Ours	405–687	0.864 $\pm$ 0.038	0.831 $\pm$ 0.037
SpecFormer (galaxies)	360–980	0.844 $\pm$ 0.028	0.805 $\pm$ 0.037
Mantis (time series)	—	0.824 $\pm$ 0.025	0.758 $\pm$ 0.047

Table 1: Frozen representations under one linear probe (logistic regression, regularisation selected by inner cross-validation on the training folds), 7 classes,

5{\times}3

cross-validation,

\pm

standard deviation across folds. Window: wavelengths of the pretraining data. Last column: mAP with 10 % of the annotations (subsets of Table 2 ). +: concatenation of the two. Bold: best single representation.

Configuration	mAP	F1	AUROC
Expert features (probe)	0.861 $\pm$ 0.032	0.789	0.910
SSL frozen, head only	0.870 $\pm$ 0.034	0.812	0.908
SSL fine-tuned, 15 ep.	0.870 $\pm$ 0.032	0.804	0.909
SSL fine-tuned, 100 ep.	0.839 $\pm$ 0.030	0.731	0.890
From scratch, 15 ep.	0.840 $\pm$ 0.033	0.776	0.877
From scratch, 100 ep.	0.800 $\pm$ 0.037	0.730	0.856

Table 2: Seven-class fine-tuning (811 spectra,

5{\times}3

cross-validation). Top: training strategies (mAP, F1, AUROC), all with the same linear head (frozen, head only: encoder not updated); the expert-feature row is the probe of Table 1 . Bottom: label efficiency (mAP) with a fraction of each fold’s training labels, same subsets for the three columns; last column, gain of pretraining over the same model from scratch. Bold: best per column (top) and row (bottom).

Class	Expert feat.	SSL frozen	SSL fine-tuned
H $\alpha$	94.6 $\pm$ 1.7	94.9 $\pm$ 2.0	95.7 $\pm$ 2.3
molecular	92.0 $\pm$ 2.5	92.8 $\pm$ 2.7	93.0 $\pm$ 1.8
low- $e^{-}$	84.8 $\pm$ 3.9	84.9 $\pm$ 4.5	84.2 $\pm$ 5.0
cloudy	89.1 $\pm$ 3.4	89.2 $\pm$ 2.2	90.5 $\pm$ 1.6
dawn/dusk	64.2 $\pm$ 25.4	68.4 $\pm$ 25.4	67.1 $\pm$ 23.5
[N I] 520	85.1 $\pm$ 3.2	85.6 $\pm$ 3.4	86.0 $\pm$ 2.6

Table 3: Per-class AP (%) and macro-AP (

\pm

standard deviation across folds) of the strategies of Table 2 (frozen encoder, head only, 100 epochs; full fine-tuning, 15 epochs) and of the expert features (probe of Table 1 ); 811 spectra, 7 classes,

5{\times}3

cross-validation. Best value per row in bold.

Explore similar work

SpecTM: Spectral Targeted Masking for Trustworthy Foundation Models
Mar 23, 2026Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin AlamdariMasked Image Modeling Geospatial Foundation Models
Department of Civil and Environmental Engineering Florida State University, Tallahassee, FL, USA
AstroSpecLM: A Spectrum-Language Model for Evidence-Grounded Astronomical Spectral Analysis
Sep 7, 2026Jinghang Shi, Yanxia Zhang, Ali Luo +2Language Modeling Scientific QA
Attributing Preprocessing Invariance in Spectral Foundation Models
Aug 14, 2026Dongjun Wei, Hongyi WuTransfer Learning Raman Spectroscopy
ESCP Business School, Paris · OpluxCare Tech Group Ltd. · The University of Hong Kong, Hong Kong

Mar 23, 2026J/K move · Enter open · S save

SpecTM: Spectral Targeted Masking for Trustworthy Foundation Models

Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin Alamdari

Department of Civil and Environmental Engineering Florida State University, Tallahassee, FL, USA

Masked Image Modeling Geospatial Foundation Models Self-Supervised Visual Representation Learning Physics-Informed ML Hyperspectral Imaging

Open paper AskPDF arXiv

Foundation models are now increasingly being developed for Earth observation (EO), yet they often rely on stochastic masking that do not explicitly enforce physics constraints; a critical trustworthiness limitation, in particular for predictive models that guide public health decisions. In this work, we propose SpecTM (Spectral Targeted Masking), a physics-informed masking design that encourages the reconstruction of targeted bands from cross-spectral context during pretraining. To achieve this, we developed an adaptable multi-task (band reconstruction, bio-optical index inference, and 8-day-ahead temporal prediction) self-supervised learning (SSL) framework that encodes spectrally intrinsic representations via joint optimization, and evaluated it on a downstream microcystin concentration regression model using NASA PACE hyperspectral imagery over Lake Erie. SpecTM achieves R^2 = 0.695 (current week) and R^2 = 0.620 (8-day-ahead) predictions surpassing all baseline models by (+34% (0.51 Ridge) and +99% (SVR 0.31)) respectively. Our ablation experiments show targeted masking improves predictions by +0.037 R^2 over random masking. Furthermore, it outperforms strong baselines with 2.2x superior label efficiency under extreme scarcity. SpecTM enables physics-informed representation learning across EO domains and improves the interpretability of foundation models.

Abstract

Explore similar work

SpecTM: Spectral Targeted Masking for Trustworthy Foundation Models

AstroSpecLM: A Spectrum-Language Model for Evidence-Grounded Astronomical Spectral Analysis

Attributing Preprocessing Invariance in Spectral Foundation Models