Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 94 of the 94 references it cites were found in a registry or an index; 38 of the 48 sources its claims cite were checked for support.

Detecting galaxy tidal features using self-supervised representation learning

Open access
Desmons, Alice, Brough, Sarah, Lanusse, Francois · 2024 · Monthly Notices of the Royal Astronomical Society

Abstract

Low surface brightness substructures around galaxies, known as tidal features, are a valuable tool in the detection of past or ongoing galaxy mergers, and their properties can answer questions about the progenitor galaxies involved in the interactions. The assembly of current tidal feature samples is primarily achieved using visual classification, making it difficult to construct large samples and draw accurate and statistically robust conclusions about the galaxy evolution process. With upcoming large optical imaging surveys such as the Vera C. Rubin Observatory’s Legacy Survey of Space and Time, predicted to observe billions of galaxies, it is imperative that we refine our methods of detecting and classifying samples of merging galaxies. This paper presents promising results from a self-supervised machine learning model, trained on data from the Ultradeep layer of the Hyper Suprime-Cam Subaru Strategic Program optical imaging survey, designed to automate the detection of tidal features. We find that self-supervised models are capable of detecting tidal features, and that our model outperforms previous automated tidal feature detection methods, including a fully supervised model. An earlier method applied to real galaxy images achieved 76 per cent completeness for 22 per cent contamination, while our model achieves considerably higher (96 per cent) completeness for the same level of contamination. We emphasize a number of advantages of self-supervised models over fully supervised models including maintaining excellent performance when using only 50 labelled examples for training, and the ability to perform similarity searches using a single example of a galaxy with tidal features.

Overview

CoChat Summary

TL;DR

The paper presents a self-supervised NNCLR model with a linear classifier to detect galaxy tidal features in HSC-SSP Ultradeep imaging, and compares it to supervised models on SDSS and HSC data. It reports high completeness and ROC AUC, with strong performance even when only 50 labelled examples are used.

Key findings

  • On HSC-SSP, the SSL model reaches ROC AUC = 0.911 ± 0.002 and TPR = 0.94 ± 0.1 at FPR = 0.20; with only 50 labels, it maintains ROC AUC = 0.89 ± 0.01 and TPR = 0.90 ± 0.1 at FPR = 0.2.
  • The model reaches 96% completeness for 22% contamination, compared with 76% completeness for 22% contamination from an earlier automated method.
  • On SDSS merger classification, with 120 unique labels the SSL model reaches AUC = 0.85 while the supervised model drops to 0.77; with larger label sets both reach similar ROC AUC ≈ 0.96. On HSC, the SSL model consistently outperforms the supervised model (maximum ROC AUC 0.911 ± 0.002 vs 0.867 ± 0.004; at 50 labels 0.89 ± 0.01 vs 0.83 ± 0.01).
  • Training is inexpensive: the encoder required ∼30 minutes for 25 epochs on a single GPU, and the linear classifier ∼1 minute for 50 epochs; a similarity search using one labelled tidal-feature example retrieves other candidates from ∼44,000 unlabelled galaxies.

Methods

  • Two datasets: HSC-SSP PDR2 Ultradeep, with ∼44,000 unlabelled galaxies after filtering and a labelled sample of 760 galaxies (380 tidal, 380 non-tidal) split 600/60/100 for training/validation/testing; and SDSS DR7/Pearson et al. (2019), with 6000 galaxies (3000 merging, 3000 non-merging) split 4800/600/600, plus an unlabelled SDSS set expanded to 50,000 via rotations.
  • Model: NNCLR self-supervised encoder using ResNet-20, 128-dimensional representations, queue size 10,000, temperature 0.1, trained for 25 epochs on unlabelled data; a linear classifier with sigmoid activation, Adam optimiser, and binary cross entropy trained for 50 epochs on labelled data. Augmentations include random flips, Gaussian noise, and jitter/crop.
  • Evaluation: ROC AUC and TPR at FPR = 0.2, averaged over ten runs with outliers beyond 3σ removed; UMAP visualisation of encoded representations, similarity searches, and performance across five magnitude-radius regions with 200 galaxies each.

Limitations

  • The paper states that the labelled tidal-feature sample is brighter, lower-redshift, and slightly larger in apparent radius than the unlabelled and non-tidal samples, so the model is biased toward assigning high classifier scores to brighter galaxies; contamination is expected mainly from bright, larger, low-redshift non-tidal galaxies.
  • The paper states that validation and testing sets are very small (60 and 100 galaxies), making evaluation difficult; the SDSS dataset is selected for obvious interacting mergers rather than tidal features, and SDSS surface-brightness limits mean only the brightest tidal features would be visible.
  • The paper states that comparison with other works is limited by the scarcity of similar studies, and that the model is intended for use with visual classification rather than as a fully automated classifier; no optimal threshold is specified because the TPR/FPR trade-off is application-dependent.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/mnras/stae1402

Cites

94 of 94 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Example of galaxies with tidal features from HSC-SSP ( gri -band cutout images). Top row from left to right: shells, stream. Bottom row from left to right: asymmetric halo, and double nucleus.
Figure 1 Example of galaxies with tidal features from HSC-SSP ( gri -band cutout images). Top row from left to right: shells, stream. Bottom row from left to right: asymmetric halo, and double nucleus.From the PDF
Figure 2 Distribution of galaxy g -i colour (top), i -band magnitude (middle top), redshift (middle bottom), and radius (bottom) for the labelled datasets of galaxies with known tidal features (blue) and without tidal features (red) and the large unlabelled dataset (dashed green).
Figure 2 Distribution of galaxy g -i colour (top), i -band magnitude (middle top), redshift (middle bottom), and radius (bottom) for the labelled datasets of galaxies with known tidal features (blue) and without tidal features (red) and the large unlabelled dataset (dashed green).From the PDF
Figure 3 Illustration of the self-supervised model architecture. The model takes in a batch of images as input and creates two different views of the batch using random augmentations. Each view is encoded and the nearest neighbour of view 1 is located from the queue of example images. The loss between view 2 and the nearest neighbour of view 1 is calculated using a contrastive loss function which is minimised for similar pairs of images and maximised otherwise.
Figure 3 Illustration of the self-supervised model architecture. The model takes in a batch of images as input and creates two different views of the batch using random augmentations. Each view is encoded and the nearest neighbour of view 1 is located from the queue of example images. The loss between view 2 and the nearest neighbour of view 1 is calculated using a contrastive loss function which is minimised for similar pairs of images and maximised otherwise.From the PDF
Figure 4 Test set ROC AUC as a function of the number of SDSS labels used for training for a supervised (blue) and self-supervised (red) model. Both models show similar performance in the high number of labels regime. For fewer training labels, the self-supervised model outperforms the supervised model.
Figure 4 Test set ROC AUC as a function of the number of SDSS labels used for training for a supervised (blue) and self-supervised (red) model. Both models show similar performance in the high number of labels regime. For fewer training labels, the self-supervised model outperforms the supervised model.From the PDF
Figure 5 Test set ROC AUC as a function of the number of HSC-SSP labels used for training for a supervised (blue) and self-supervised (red) model. The points along the solid lines represent the average test set ROC AUC where each point is an average of ten runs. The points along the dashed lines show the ROC AUC for the model which reached the lowest final validation loss for the given number of training labels. When using the average ROC AUCthe self-supervised model performs consistently better, but both models remain consistent for all number of labels. When using the model with the lowest validation loss both models show similar performance in the high number of labels regime. For fewer training labels, the self-supervised model outperforms the supervised model.
Figure 5 Test set ROC AUC as a function of the number of HSC-SSP labels used for training for a supervised (blue) and self-supervised (red) model. The points along the solid lines represent the average test set ROC AUC where each point is an average of ten runs. The points along the dashed lines show the ROC AUC for the model which reached the lowest final validation loss for the given number of training labels. When using the average ROC AUCthe self-supervised model performs consistently better, but both models remain consistent for all number of labels. When using the model with the lowest validation loss both models show similar performance in the high number of labels regime. For fewer training labels, the self-supervised model outperforms the supervised model.From the PDF
Figure 6 Average test set TPR when FPR = 0.2 as a function of the number HSC-SSPlabelsusedfortraining for the supervised (blue) and self-supervised (red) model. Each point is an average of 10 runs. Both models show a slight decrease in TPR with decreasing training label number. The self-supervised model has a consistently higher TPR.
Figure 6 Average test set TPR when FPR = 0.2 as a function of the number HSC-SSPlabelsusedfortraining for the supervised (blue) and self-supervised (red) model. Each point is an average of 10 runs. Both models show a slight decrease in TPR with decreasing training label number. The self-supervised model has a consistently higher TPR.From the PDF
Figure 7 ROC curves for the ten self-supervised models trained on 600 labels. The model with the highest ROC AUC is shown in red, the other nine models are shown in grey. The model shown in red was an outlier and hence was not included in the average ROC AUC calculations in Figure 5.
Figure 7 ROC curves for the ten self-supervised models trained on 600 labels. The model with the highest ROC AUC is shown in red, the other nine models are shown in grey. The model shown in red was an outlier and hence was not included in the average ROC AUC calculations in Figure 5.From the PDF
Figure 8 Results from a similarity search using two random galaxies with tidal features as query images. The two query galaxies are displayed on the left, alongside the top 24 galaxies with the highest similarity scores for each similarity search. The similarity and classifier scores, denoted by S and C respectively, are displayed in the top left corner for each image. The red outlines indicate galaxies which would be visually classified as hosting tidal features, regardless of whether this galaxy is the central object in the image e.g. first row, second column.
Figure 8 Results from a similarity search using two random galaxies with tidal features as query images. The two query galaxies are displayed on the left, alongside the top 24 galaxies with the highest similarity scores for each similarity search. The similarity and classifier scores, denoted by S and C respectively, are displayed in the top left corner for each image. The red outlines indicate galaxies which would be visually classified as hosting tidal features, regardless of whether this galaxy is the central object in the image e.g. first row, second column.From the PDF
Figure 9 2D UMAP projection of the self-supervised representations. Made by binning the space into 100 × 100 cells and randomly selecting a sample from that cell to plot in the corresponding cell location.
Figure 9 2D UMAP projection of the self-supervised representations. Made by binning the space into 100 × 100 cells and randomly selecting a sample from that cell to plot in the corresponding cell location.From the PDF
Figure 10 Left: The same 2D UMAP projection as Figure 9, binning the space into 40 × 40 cells, coloured according to the median classifier score (assigned to each galaxy by the linear classifier) in each bin. Bins with less than 5 galaxies are coloured white. Right: The same 2D UMAP projection, this time coloured according to the median absolute deviation of the classifier scores in each bin.
Figure 10 Left: The same 2D UMAP projection as Figure 9, binning the space into 40 × 40 cells, coloured according to the median classifier score (assigned to each galaxy by the linear classifier) in each bin. Bins with less than 5 galaxies are coloured white. Right: The same 2D UMAP projection, this time coloured according to the median absolute deviation of the classifier scores in each bin.From the PDF
Figure 11 A Panels: Identical to Figure 2 - Distribution of galaxy properties for the labelled datasets of galaxies with known tidal features (blue) and without tidal features (red) and the large unlabelled dataset (dashed green). B Panels: Distribution of galaxy properties for galaxies in the unlabelled dataset assigned a classifier score ≥ 0.8 (dashed blue) and ≤ 0.2 (dashed red). From left to right: g -i colour, i -band cmodel fit magnitude, photometric redshift, and i -band MGE fit radius.
Figure 11 A Panels: Identical to Figure 2 - Distribution of galaxy properties for the labelled datasets of galaxies with known tidal features (blue) and without tidal features (red) and the large unlabelled dataset (dashed green). B Panels: Distribution of galaxy properties for galaxies in the unlabelled dataset assigned a classifier score ≥ 0.8 (dashed blue) and ≤ 0.2 (dashed red). From left to right: g -i colour, i -band cmodel fit magnitude, photometric redshift, and i -band MGE fit radius.From the PDF
Figure 12 Galaxy i -band MGE radius as a function of i -band cmodel magnitude for the ∼ 44,000 galaxies in the unlabelled dataset. The red numbered boxes show the five regions in which we evaluate the model performance. The blue dashed box indicates the region marked out by the 25 th -75 th percentiles of radius and magnitude for the galaxies in the labelled tidal feature dataset.
Figure 12 Galaxy i -band MGE radius as a function of i -band cmodel magnitude for the ∼ 44,000 galaxies in the unlabelled dataset. The red numbered boxes show the five regions in which we evaluate the model performance. The blue dashed box indicates the region marked out by the 25 th -75 th percentiles of radius and magnitude for the galaxies in the labelled tidal feature dataset.From the PDF
Figure 13 The upper panels show the distribution of classifier scores assigned by the model to the 200 galaxies in each box. Galaxies visually-classified as having tidal features are shown in purple, and galaxies without tidal features are shown in blue. Each subplot also shows the tidal feature fraction in the respective box, according to visual classification. The lower panels show the Tidal feature fraction in each classifier score bin. Uncertainties are given for 1 𝜎 and estimated as described in the text. The red dashed line in each subplot indicates the tidal feature fraction for the relevant box, according to visual classification.
Figure 13 The upper panels show the distribution of classifier scores assigned by the model to the 200 galaxies in each box. Galaxies visually-classified as having tidal features are shown in purple, and galaxies without tidal features are shown in blue. Each subplot also shows the tidal feature fraction in the respective box, according to visual classification. The lower panels show the Tidal feature fraction in each classifier score bin. Uncertainties are given for 1 𝜎 and estimated as described in the text. The red dashed line in each subplot indicates the tidal feature fraction for the relevant box, according to visual classification.From the PDF

Verification

Claims the paper makes

30 Valid · 9 Not valid · 3 Not checked · 76 Own results and interpretations

45 more claims · 8 with their cited sources checked

Create a free account to see all 118 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 56
  • Probably not a substantial claim. 27
  • Borderline: one of the tests for a substantial claim is unclear. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 93 references

Research works printed but not linked: 2

  1. Ćiprijanović, 2023, MLST, 4, 025013
  2. Huang, 2019, Unagi

Cited by

See all 15

Metadata

Authors
Desmons, A., Brough, S., & Lanusse, F.
Affiliations
  • UNSW Sydney (AU)
  • AIM, CEA, CNRS, Université Paris-Saclay, Université Paris Diderot , Sorbonne Paris Cité, F-91191 Gif-sur-Yvette , France
Corresponding author
Desmons, Alice
Venue
Monthly Notices of the Royal Astronomical Society
Published
2024-06-05
Volume
531
Issue
4
Pages
4070-4084
ISSN
  • 0035-8711 (ISSN-L)
  • 1365-2966
OpenAlex
W4399363409
Cited by
15
References
93
Citations by year
2026: 6 · 2025: 7 · 2024: 1
FWCI
2.63
Type
journal article
Language
en
Open access
Locations
  • Monthly Notices of the Royal Astronomical Society (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • HAL (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
Australian Research Council (DP190101943)
Topics
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Galaxies: Formation, Evolution, Phenomena
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Stellar, planetary, and galactic studies
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Astrophysical Phenomena and Observations
Keywords
tidal features, galaxy mergers, self-supervised representation learning, automated feature detection, galaxy interactions, Hyper Suprime-Cam Subaru Strategic Program, Vera C. Rubin Observatory Legacy Survey of Space and Time, similarity search, galaxy evolution

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Detecting galaxy tidal features using self-supervised representation learning. https://app.cochat.ai/papers/10.1093/mnras/stae1402
The paper
Desmons, A., Brough, S., & Lanusse, F. (2024). Detecting galaxy tidal features using self-supervised representation learning. Monthly Notices of the Royal Astronomical Society. https://doi.org/10.1093/mnras/stae1402

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.