Deep Learning Segmentation of Spiral Arms and Bars
Mike Walmsley∗
Dunlap Institute for Astronomy and Astrophysics
University of Toronto
Toronto, ON M5S 3H4, Canada
m.walmsley@utoronto.ca
Ashley Spindler
Centre for Astrophysics Research, Department of Physics, Astronomy and Mathematics
University of Hertfordshire
Hatfield, Hertfordshire, AL10 9AB, UK
a.spindler@herts.ac.uk
Abstract
We present the first deep learning model for segmenting galactic spiral arms and
bars. In a blinded assessment by expert astronomers, our predicted spiral arm
masks are preferred over both current automated methods (99% of evaluations)
and our original volunteer labels (79% of evaluations). Experts rated our spiral
arm masks as ‘mostly good’ to ‘perfect’ in 89% of evaluations. Bar lengths
trivially derived from our predicted bar masks are in excellent agreement with a
dedicated crowdsourcing project. The pixelwise precision of our masks, previously
impossible at scale, will underpin new research into how spiral arms and bars
evolve.
1 Introduction
Astronomers do not know how spiral arms are made. There are two long-standing theories: density
waves [1] and swing-amplification [2]. Despite decades of work, it is unclear which theory dominates
[3, 4]. Spiral arms also show different characteristics in different galaxies; they vary in arrangement
(e.g. grand design vs. flocculent), arm count, and arm tightness (pitch angle). Precise explanations
for these characteristics remain elusive.
A closely-related galaxy structure is galactic bars; linear features in the center of galaxies, often
connecting two spiral arms [5]. Barred spirals make up about 30% of disk-type galaxies in the nearby
universe [6]. Bars alter the structure of a host galaxy by funnelling gas and dust from the disk into
the galaxy bulge [7, 8, 9]. Bars grow over time; whether long and short bars are distinct subclasses or
part of a continuum is unclear [10].
Progress in understanding both bars and spiral arms has been limited by the difficulty of measuring
their shapes [4, 11, 12] Shape measurements to-date have been made by either expert inspection [13],
rule-based image processing [14, 1], or crowdsourcing [15, 12].
We develop a deep learning model to jointly identify both spiral arms and bars.
Deep learning for galaxy morphology has focused almost exclusively on classifying entire images
([16, 17, 18, 19, 20, 21, 22, 23, 24, 25], etc.) Works on segmentation are rare. Hausen and Robertson
∗Second affiliation: Jodrell Bank Centre for Astrophysics, Department of Physics & Astronomy, University
of Manchester, Manchester M13 9PL, UK
Machine Learning and the Physical Sciences Workshop, NeurIPS 2023.
ar
X
iv
:2
31
2.
02
90
8v
1 
 [a
str
o-
ph
.G
A]
  5
 D
ec
 20
23
Input Volunteers
Masters+'21
sparcfire
Davis+'14
This Work
Figure 1: From left: input galaxy image, segmentation masks by volunteers (GZ3D, [11]), masks
by sparcfire [14](spiral only), and masks by our model. Rows show randomly-selected test set
galaxies, filtered to have GZ DESI [25] vote fractions above 0.5.
[26] trained a U-net-like architecture to predict ‘pixel level morphological classifications’ i.e. to
identify the pixels belonging to a single galaxy and then ascribe all those pixels to a single galaxy
class (e.g. spheroid, disk, etc.) Ostdiek et al. [27] used a U-Net-like architecture to classify pixels as
belonging to either lensing galaxies or lensed subhalos. Richards et al. [28] heavily extended Mask
R-CNN to classify both tidal structures around galaxies and background contamination. Bhambra
et al. [29] (hereafter PB+2022) trained a standard CNN (rather than U-Net or Mask R-CNN) to
classify galaxy images as barred or not barred, and then used saliency mapping to identify the pixels
most relevant to that classification. All of these works adopt a pixelwise classification framework.
2 Methods
2.1 Data and Loss Function Design
Our segmentation labels are sourced from Galaxy Zoo: 3D, published in Masters et al. [11]. GZ3D
collected crowdsourced pixel masks of the spiral arms and bars of the 29,831 galaxies listed as
potential targets for the MaNGA survey [30]. While the project was successful and released a public
catalogue of pixel masks, Masters et al. [11] highlighted the time-consuming and labour-intensive
nature of the task — particularly in regards to labelling the spiral arms.
2
Spiral
0%
20%
40%
60%
80%
100%
Ex
pe
rt-
pr
ef
er
re
d 
al
go
rit
hm
This work
GZ3D
Sparcfire
Tie
Bar
(a) Fraction of astronomer evaluations selecting each ‘algorithm’
as best-performing (N=367, 133). Astronomers strongly prefer
our model, even over the original labels we use for training.
0 5 10 15 20
Predicted bar length [h 1 kpc]
0
5
10
15
20
Ho
yl
e 
(2
01
1)
 b
ar
 le
ng
th
 [h
1  k
pc
]
r = 0.91
L1=1.2 kpc
(b) Bar length measurements from our model
vs. human measurements (Hoyle et al. [15]).
Length predictions are derived by trivially
processing our bar masks.
GZ3D volunteers used a polygon drawing tool to enclose areas corresponding to bars or spiral
arms. Each galaxy was marked by 15 volunteers. We use the individual volunteer polygon vertices
published by Masters et al. [11] to calculate how many volunteers enclosed each pixel.2
Unlike previous astronomy works, we frame our segmentation task as pixelwise regression. The
fraction of volunteers enclosing each pixel reflects the confidence of the crowd. This confidence is a
critical training signal when attempting to learn the often-ambiguous task of identifying spiral arms.
We choose the loss function:
L(X,V,w) ∝
C∑
k
|
X∑
i,j
V∑
l
δ(Xij ∈ Vl,k)− ϕw(X)ijk | (1)
where, for spatial pixel index (i, j) in image X , we check if each pixel Xij is enclosed by the vertices
marked by volunteer Vl for class k (spiral or bar) and compare that to the model output for the same
pixel in channel k, ϕw(X)ijk. In short, we calculate the mean absolute error when predicting, for
each pixel, the fraction of volunteers who included that pixel in their mask, and sum over each output
class/channel.
We implement this loss by encoding the pixelwise volunteer vote fractions as greyscale JPEG images
(where e.g. a value of 255 corresponds to all 15 volunteers marking a pixel) and then loading and
augmenting these masks as if they were conventional images.3 Our final convolutional layer uses
ReLU activation [31] to ensure non-negative predictions. The same network jointly learns to predict
masks for both spiral arms and bars using a two-channel output.
Our images are sourced from the DESI Legacy Surveys [32]. DESI-LS images are comparable to the
original GZ3D images for identifying spiral arms and bars in nearby galaxies, but using them for
training allows for future work to make direct measurements on the full DESI-LS footprint. We use
fixed train/validation/test sets with 70/10/20% division. We only include galaxies with GZ3D masks
for spiral arms. We do not penalise the network for predicting bars (or, in principle, spiral arms) on
galaxies that volunteers chose not to mark as such, as we judge the lack of marks to be an unreliable
indicator for the lack of a feature.
2Unlike Masters et al. [11], we do not discount self-intersecting areas. Many volunteer marks are only self-
intersecting where closing line segments slightly intersect, and so discarding these areas removes a significant
amount of otherwise-useful annotations.
3We use only flip and interpolated rotation augmentations, thus preserving the pixel (vote fraction) values.
3
2.2 Model Architecture
Our model uses a U-Net architecture which consists of a downsampling encoder and an upsampling
decoder. For both submodels, a single step is comprised of two residual blocks, followed by a
strided convolution for the downsampling or a transpose convolution for the upsampling. Dropout
is used between the residual blocks and convolutions in the encoder to control overfitting. Skip
connections between each level of the encoder pass the outputs of the downsampling step to the
inputs of the corresponding upsampling step. Following Smith et al. [33] we use Mish activation
functions [34] in the residual blocks. We perform a hyperparameter sweep (on the validation set) to
select the batch size, dropout rates, and down/upsampling dimensions. Our code is publicly available
at https://github.com/mwalmsley/zoobot-3d.
2.3 Results
Figure 1 presents a random selection of predicted spiral and bar masks. Our predicted masks typically
capture spiral and bar features well. Further, they are typically smoother and more precise than the
original volunteer labels. The polygon drawing tool used by GZ3D volunteers creates a ‘blocky’
mask which fails to capture the soft edges and ambiguous locations of spiral arms. Small number
statistics (each galaxy is marked by 1-15 volunteers) then lead to unphysical step functions where
masks from individual volunteers overlap. Our model learns from these observed masks to predict
the expected mask at each pixel. Unlike the labels, the expected mask is smoothly varying.
Comparing to previous work is difficult because, to our knowledge, there are no existing auto-
mated methods for segmenting spiral arms. The closest comparable code is sparcfire [14], which
segments as an intermediate step towards mathematically describing the arm segments. We run
sparcfire with the recommended default parameters, except for disabling deprojection; deprojec-
tion transforms the resulting segmentation mask, making direct comparison impossible. Our model is
notably more accurate than sparcfire on the random examples shown. We also compared with the
recent general purpose ‘foundation’ model SegmentAnything [35]; we find that SegmentAnything
typically identifies galaxy vs. background pixels but cannot segment structures within galaxies.
To quantitatively assess performance, we recruited 20 astronomers for a blinded comparison between
methods. Participants were shown a grid of images (similar to 1) and asked to rate their preferred
and second-preferred mask. For each preferred algorithm, they then rated each mask by quality on a
four-point scale: ‘perfect’, ‘mostly perfect (some minor errors)’, ‘mostly poor (some major errors)’,
and ’totally failed’. Participants made a total of 500 evaluations on 100 galaxies. Participants were
not involved in this research and did not know which algorithm generated each mask.
We find that our model is strongly preferred by astronomers over both sparcfire and our original
volunteer labels. Fig. 2a shows the fraction of evaluations where astronomers rated each ‘algorithm’
as the best performing; for spiral arms, our model was preferred by an individual expert in 78%
of spiral evaluations (vs. 20% for the multi-volunteer GZ3D labels and 1% for sparcfire) and
68% of bar evaluations (vs. 11% for the GZ3D labels, with the remainder tied). Our model ranked
higher than sparcfire in 99% of spiral evaluations and higher than GZ3D in 79% of evaluations 4.
Astronomers rated our spiral arm masks as ‘mostly good’ to ‘perfect’ in 89% of evaluations, vs. 77%
for the GZ3D labels.
To quantitatively demonstrate that our masks are useful for astronomy, we process our bar masks to
calculate bar lengths. We calculate the length of the confidently-barred5 pixel mask and convert this
length to physical distances. We find (Fig. 2b excellent agreement with the measurements gathered
by Hoyle et al. [15] in a crowdsourcing project dedicated to finding bars.
4Formally, the one-tailed Binomial odds of an ensemble of experts preferring our model in k out of N
galaxies (excluding ties) under the null hypothesis of no true preference between our model and the original
labels is p << 0.01 for spirals (N = 91, k = 74) and p = 0.04 for bars (N = 26, k = 17)
5Defined as having predicted bar vote fractions greater than the median-confidence pixel
4
3 Conclusion and Discussion
We have presented the first deep learning model for segmenting spiral arms and bars. Our model
outperforms both the closest existing automated approach and the original crowdsourced labels on
which it was trained, as judged by independent astronomers in a blinded evaluation.
Accurate segmentation of spiral arms and bars is especially relevant for Integral Field Spectroscopy
(IFS) surveys, which measure spectra at every pixel. This was the original motivation for GZ3D.
There are now multiple large IFS surveys (e.g. MaNGA, SAMI) and IFS is available on observational
platforms including NIRSpec and MIRI on JWST and ESO-MUSE on the VLT. Accurate automated
segmentation masks could enhance research with any of these platforms.
Accurate segmentation is also relevant for the very largest surveys. Our model is trained on DESI-LS
images, enabling spiral and bar shape measurements for galaxies with upcoming DESI spectra.
Euclid [36] will shortly begin resolving the detailed morphology of higher-redshift (z ≈ 1) galaxies,
allowing us to measure bars and spirals as they form and grow to their present-day appearance.
We chose to use a U-Net architecture because it is appropriate for smaller labelled datasets (here,
5054 training galaxies) and to focus this work on introducing a first solution to this new task. More
recent architectures may improve performance.
In closing, we would like to note a negative result. We found that, by adding a Dirichlet regression
head [37] to the latent space of our U-Net, we could also predict Galaxy Zoo volunteer vote fractions
competitively with a dedicated classification model. However, adding this joint prediction task did
not improve the quality of our segmentation masks. This complicates the ‘astronomy foundation
model’ view presented by Walmsley et al. [38] (see also [39, 40, 41]). General purpose models may
not always outperform dedicated ones.
Acknowledgments and Disclosure of Funding
We are grateful to the Zooniverse volunteers and GZ:3D authors for creating the labels we rely on
here.
We would like to thank the astronomers who contributed blinded assessments, including: Dominic
Adams, Elisabeth M Baeten, Hugh Dickinson, James Tropp Garland, Antonio Herrera-Martin, Bill
Keel, Karen L Masters, James Pearson, Jonathon C S Pierce, Shravya Shenoy, Brooke Simmons,
Josh Speagle, Grant Stevens, Laura Trouille, and Klaas Wiersema.
This work was supported by a grant from Meta. MW is a Dunlap Fellow. The Dunlap Institute is
funded through an endowment established by the David Dunlap family and the University of Toronto.
This work has made use of the University of Hertfordshire high-performance computing facility
(https://uhhpc.herts.ac.uk/.
References
[1] C. C. LIN and FRANK H. SHU. ON THE SPIRAL STRUCTURE OF DISK GALAX-
IES. Selected Papers of C C Lin with Commentary, pages 561–570, 8 1987. doi:
10.1142/9789814415651{\_}0033.
[2] J. A. Sellwood, L. S. Sparke, J. A. Sellwood, and L. S. Sparke. Pattern speeds in barred spiral
galaxies. MNRAS, 231(1):25P–31, 3 1988. ISSN 0035-8711. doi: 10.1093/MNRAS/231.1.25P.
URL https://ui.adsabs.harvard.edu/abs/1988MNRAS.231P..25S/abstract.
[3] Ross E. Hart, Steven P. Bamford, William C. Keel, Sandor J. Kruk, Karen L. Masters, Brooke D.
Simmons, and Rebecca J. Smethurst. Galaxy Zoo: constraining the origin of spiral arms.
Monthly Notices of the Royal Astronomical Society, 478(1):932–949, 7 2018. ISSN 0035-
8711. doi: 10.1093/MNRAS/STY1201. URL https://researchportal.port.ac.uk/en/
publications/galaxy-zoo-constraining-the-origin-of-spiral-arms.
[4] Si-Yue Yu and Luis C. Ho. The Statistical Properties of Spiral Arms in Nearby Disk Galaxies.
The Astrophysical Journal, 900(2):150, 9 2020. ISSN 0004-637X. doi: 10.3847/1538-4357/
ABAC5B. URL https://iopscience.iop.org/article/10.3847/1538-4357/
abac5bhttps://iopscience.iop.org/article/10.3847/1538-4357/abac5b/meta.
5
[5] Ronald J. Buta. Galaxy Morphology. 2 2011. URL http://arxiv.org/abs/1102.0550.
[6] Karen L. Masters, Robert C. Nichol, Ben Hoyle, Chris Lintott, Steven P. Bamford, Edward M.
Edmondson, Lucy Fortson, William C. Keel, Kevin Schawinski, Arfon M. Smith, and Daniel
Thomas. Galaxy Zoo: Bars in disc galaxies. Monthly Notices of the Royal Astronomi-
cal Society, 411(3):2026–2034, 3 2011. ISSN 13652966. doi: 10.1111/j.1365-2966.2010.
17834.x. URL https://academic.oup.com/mnras/article-lookup/doi/10.1111/j.
1365-2966.2010.17834.x.
[7] F. Combes, F. Debbasch, D. Friedli, D. Pfenniger, F. Combes, F. Debbasch, D. Friedli, and
D. Pfenniger. Box and peanut shapes generated by stellar bars. A&A, 233:82, 1990. ISSN 0004-
6361. URL https://ui.adsabs.harvard.edu/abs/1990A&A...233...82C/abstract.
[8] E. Athanassoula and A. Misiriotis. Morphology, photometry and kinematics of N-body bars
- I. Three models with different halo central concentrations. Monthly Notices of the Royal
Astronomical Society, 330(1):35–52, 2 2002. ISSN 00358711. doi: 10.1046/j.1365-8711.2002.
05028.x. URL https://academic.oup.com/mnras/article-lookup/doi/10.1046/j.
1365-8711.2002.05028.x.
[9] S. Khoperskov, M. Haywood, P. Di Matteo, M. D. Lehnert, and F. Combes. Bar quenching
in gas-rich galaxies. Astronomy and Astrophysics, 609:A60, 1 2018. ISSN 14320746. doi:
10.1051/0004-6361/201731211. URL https://ui.adsabs.harvard.edu/abs/2018A&A.
..609A..60K/abstract.
[10] Tobias Geron, Rebecca J Smethurst, Chris Lintott, Sandor Kruk, Karen L Masters, Brooke
Simmons, Kameswara Bharadwaj Mantha, Mike Walmsley, L Garma-Oehmichen, Niv Drory,
and Richard R Lane. Galaxy Zoo: kinematics of strongly and weakly barred galaxies. Monthly
Notices of the Royal Astronomical Society, 521(2):1775–1793, 3 2023. ISSN 0035-8711. doi:
10.1093/MNRAS/STAD501. URL https://academic.oup.com/mnras/article/521/2/
1775/7043473.
[11] Karen L. Masters, Coleman Krawczyk, Shoaib Shamsi, Alexander Todd, Daniel Finnegan,
Matthew Bershady, Kevin Bundy, Brian Cherinka, Amelia Fraser-Mckelvie, Dhanesh Krish-
narao, Sandor Kruk, Richard R. Lane, David Law, Chris Lintott, Michael Merrifield, Brooke
Simmons, Anne Marie Weijmans, and Renbin Yan. Galaxy Zoo: 3D – crowdsourced bar, spiral,
and foreground star masks for MaNGA target galaxies. Monthly Notices of the Royal Astronom-
ical Society, 507(3):3923–3935, 9 2021. ISSN 0035-8711. doi: 10.1093/MNRAS/STAB2282.
URL https://dx.doi.org/10.1093/mnras/stab2282.
[12] Ross E. Hart, Steven P. Bamford, Kyle W. Willett, Karen L. Masters, Carolin Cardamone,
Chris J. Lintott, Robert J. Mackay, Robert C. Nichol, Christopher K. Rosslowe, Brooke D.
Simmons, and Rebecca J. Smethurst. Galaxy Zoo: Comparing the demographics of spiral
arm number and a new method for correcting redshift bias. Monthly Notices of the Royal
Astronomical Society, 461(4):3663–3682, 2016. ISSN 13652966. doi: 10.1093/mnras/stw1588.
[13] Robert Kennicutt and Paul Hodge. A COMPARISON OF MEASURED SPIRAL ARM PROP-
ERTIES WITH MODEL PREDICTIONS. The Astrophysical Journal, 101:101–107, 1982.
[14] Darren R. Davis and Wayne B. Hayes. SpArcFiRe: Scalable automated detection of spiral
galaxy arm segments. Astrophysical Journal, 790(2):87, 7 2014. ISSN 15384357. doi:
10.1088/0004-637X/790/2/87. URL http://stacks.iop.org/0004-637X/790/i=2/a=
87?key=crossref.73c5cbd1359d0e439cd559ab23a013d0.
[15] Ben Hoyle, Karen L. Masters, Robert C. Nichol, Edward M. Edmondson, Arfon M. Smith,
Chris Lintott, Ryan Scranton, Steven Bamford, Kevin Schawinski, and Daniel Thomas. Galaxy
Zoo: bar lengths in local disc galaxies. Monthly Notices of the Royal Astronomical Society, 415
(4):3627–3640, 8 2011. ISSN 1365-2966. doi: 10.1111/J.1365-2966.2011.18979.X. URL
https://onlinelibrary.wiley.com/doi/full/10.1111/j.1365-2966.2011.18979.
xhttps://onlinelibrary.wiley.com/doi/abs/10.1111/j.1365-2966.2011.18979.
xhttps://onlinelibrary.wiley.com/doi/10.1111/j.1365-2966.2011.18979.x.
6
[16] S. Dieleman, K. W. Willett, and J. Dambre. Rotation-invariant convolutional neural networks
for galaxy morphology prediction. Monthly Notices of the Royal Astronomical Society, 450(2):
1441–1459, 2015. ISSN 0035-8711. doi: 10.1093/mnras/stv632. URL http://arxiv.org/
abs/1503.07077.
[17] M. Huertas-Company, R. Gravet, G. Cabrera-Vives, P. G. Pérez-González, J. S. Kartaltepe,
G. Barro, M. Bernardi, S. Mei, F. Shankar, P. Dimauro, E. F. Bell, D. Kocevski, D. C. Koo, S. M.
Faber, and D. H. McIntosh. A catalog of visual-like morphologies in the 5 candels fields using
deep learning. Astrophysical Journal, Supplement Series, 221(1):8, 2015. ISSN 00670049. doi:
10.1088/0067-0049/221/1/8. URL http://dx.doi.org/10.1088/0067-0049/221/1/8.
[18] Asad Khan, E. A. Huerta, Sibo Wang, Robert Gruendl, Elise Jennings, and Huihuo Zheng. Deep
learning at scale for the construction of galaxy catalogs in the Dark Energy Survey. Physics
Letters, Section B: Nuclear, Elementary Particle and High-Energy Physics, 795:248–258, 2019.
ISSN 03702693. doi: 10.1016/j.physletb.2019.06.009. URL https://www.sciencedirect.
com/science/article/pii/S0370269319303879?via%3Dihub.
[19] M. Huertas-Company, J. R. Primack, A. Dekel, D. C. Koo, S. Lapiner, D. Ceverino, R. C.
Simons, G. F. Snyder, M. Bernardi, Z. Chen, H. Domínguez-Sánchez, Z. Chen, C. T. Lee,
B. Margalef-Bentabol, and D. Tuccillo. Deep Learning Identifies High-z Galaxies in a Central
Blue Nugget Phase in a Characteristic Mass Range. The Astrophysical Journal, 858(2):114, 5
2018. ISSN 1538-4357. doi: 10.3847/1538-4357/aabfed. URL http://stacks.iop.org/
0004-637X/858/i=2/a=114?key=crossref.4e09299072a147355484d972ba56e818.
[20] H. Dominguez Sanchez, M. Huertas-Company, M. Bernardi, S. Kaviraj, J. L. Fischer, T. M.C.
Abbott, F. B. Abdalla, J. Annis, S. Avila, D. Brooks, E. Buckley-Geer, A. Carnero Rosell,
M. Carrasco Kind, J. Carretero, C. E. Cunha, C. B. D’Andrea, L. N. Da Costa, C. Davis,
J. De Vicente, P. Doel, A. E. Evrard, P. Fosalba, J. Frieman, J. Garc´ia-Bellido, E. Gaztanaga,
D. W. Gerdes, D. Gruen, R. A. Gruendl, J. Gschwend, G. Gutierrez, W. G. Hartley, D. L.
Hollowood, K. Honscheid, B. Hoyle, D. J. James, K. Kuehn, N. Kuropatkin, O. Lahav, M. A.G.
Maia, M. March, P. Melchior, F. Menanteau, R. Miquel, B. Nord, A. A. Plazas, E. Sanchez,
V. Scarpine, R. Schindler, M. Schubnell, M. Smith, R. C. Smith, M. Soares-Santos, F. Sobreira,
E. Suchyta, M. E.C. Swanson, G. Tarle, D. Thomas, A. R. Walker, and J. Zuntz. Transfer
learning for galaxy morphology from one survey to another. Monthly Notices of the Royal
Astronomical Society, 484(1):93–100, 3 2019. ISSN 13652966. doi: 10.1093/mnras/sty3497.
URL https://academic.oup.com/mnras/article/484/1/93/5266389.
[21] Mike Walmsley, Lewis Smith, Chris Lintott, Yarin Gal, Steven Bamford, Hugh Dickinson,
Lucy Fortson, Sandor Kruk, Karen Masters, Claudia Scarlata, Brooke Simmons, Rebecca
Smethurst, and Darryl Wright. Galaxy Zoo: probabilistic morphology through Bayesian CNNs
and active learning. Monthly Notices of the Royal Astronomical Society, 491(2):1554–1574, 1
2020. ISSN 0035-8711. doi: 10.1093/mnras/stz2816. URL https://academic.oup.com/
mnras/article/491/2/1554/5583078.
[22] Mike Walmsley, Chris Lintott, Tobias Géron, Sandor Kruk, Coleman Krawczyk, Kyle W. Willett,
Steven Bamford, Lee S. Kelvin, Lucy Fortson, Yarin Gal, William Keel, Karen L. Masters,
Vihang Mehta, Brooke D. Simmons, Rebecca Smethurst, Lewis Smith, Elisabeth M. Baeten, and
Christine MacMillan. Galaxy Zoo DECaLS: Detailed visual morphology measurements from
volunteers and deep learning for 314 000 galaxies. Monthly Notices of the Royal Astronomical
Society, 509(3):3966–3988, 12 2022. ISSN 13652966. doi: 10.1093/mnras/stab2093. URL
https://doi.org/10.1093/mnras/stab2093.
[23] Ting-Yun Cheng, H Domínguez Sánchez, J Vega-Ferrero, C J Conselice, M Siudek, A Aragón-
Salamanca, M Bernardi, R Cooke, L Ferreira, M Huertas-Company, J Krywult, A Palmese,
A Pieres, A A Plazas Malagón, A Carnero Rosell, D Gruen, D Thomas, D Bacon, D Brooks, D J
James, D L Hollowood, D Friedel, E Suchyta, E Sanchez, F Menanteau, F Paz-Chinchón,
G Gutierrez, G Tarle, I Sevilla-Noarbe, I Ferrero, J Annis, J Frieman, J García-Bellido,
J Mena-Fernández, K Honscheid, K Kuehn, L N da Costa, M Gatti, M Raveri, M E S Pereira,
M Rodriguez-Monroy, M Smith, M Carrasco Kind, M Aguena, M E C Swanson, N Weaverdyck,
P Doel, R Miquel, R L C Ogando, R A Gruendl, S Allam, S R Hinton, S Dodelson, S Bocquet,
S Desai, S Everett, and V Scarpine. Lessons learned from the two largest Galaxy morphological
7
classification catalogues built by convolutional neural networks. Monthly Notices of the Royal
Astronomical Society, 518(2):2794–2809, 11 2022. ISSN 0035-8711. doi: 10.1093/MNRAS/
STAC3228. URL https://academic.oup.com/mnras/article/518/2/2794/6823716.
[24] J. Vega-Ferrero, H. Domínguez Sánchez, M. Bernardi, M. Huertas-Company, R. Morgan,
B. Margalef, M. Aguena, S. Allam, J. Annis, S. Avila, D. Bacon, E. Bertin, D. Brooks,
A. Carnero Rosell, M. Carrasco Kind, J. Carretero, A. Choi, C. Conselice, M. Costanzi, L. N.
Da Costa, M. E.S. Pereira, J. De Vicente, S. Desai, I. Ferrero, P. Fosalba, J. Frieman, J. García-
Bellido, D. Gruen, R. A. Gruendl, J. Gschwend, G. Gutierrez, W. G. Hartley, S. R. Hinton,
D. L. Hollowood, K. Honscheid, B. Hoyle, M. Jarvis, A. G. Kim, K. Kuehn, N. Kuropatkin,
M. Lima, M. A.G. Maia, F. Menanteau, R. Miquel, R. L.C. Ogando, A. Palmese, F. Paz-
Chinchón, A. A. Plazas, A. K. Romer, E. Sanchez, V. Scarpine, M. Schubnell, S. Serrano,
I. Sevilla-Noarbe, M. Smith, E. Suchyta, M. E.C. Swanson, G. Tarle, F. Tarsitano, C. To, D. L.
Tucker, T. N. Varga, and R. D. Wilkinson. Pushing automated morphological classifications
to their limits with the Dark Energy Survey. Monthly Notices of the Royal Astronomical
Society, 506(2):1927–1943, 7 2021. ISSN 0035-8711. doi: 10.1093/MNRAS/STAB594. URL
https://academic.oup.com/mnras/article/506/2/1927/6156625.
[25] Mike Walmsley, Tobias Géron, Sandor Kruk, Anna M M Scaife, Chris Lintott, Karen L
Masters, James M Dawson, Hugh Dickinson, Lucy Fortson, Izzy L Garland, Kameswara
Mantha, David O’, Ryan 10, Jürgen Popp, Brooke Simmons, Elisabeth M Baeten, and Christine
Macmillan. Galaxy Zoo DESI: Detailed Morphology Measurements for 8.7M Galaxies in the
DESI Legacy Imaging Surveys. MNRAS, 000:1–18, 9 2023. URL https://arxiv.org/abs/
2309.11425v1.
[26] Ryan Hausen and Brant E. Robertson. Morpheus : A Deep Learning Framework for the
Pixel-level Analysis of Astronomical Image Data . The Astrophysical Journal Supplement
Series, 248(1):20, 6 2020. ISSN 00670049. doi: 10.3847/1538-4365/ab8868. URL http:
//arxiv.org/abs/1906.11248.
[27] Bryan Ostdiek, Ana Diaz Rivero, and Cora Dvorkin. Extracting the Subhalo Mass
Function from Strong Lens Images with Image Segmentation. The Astrophysical
Journal, 927(1):83, 3 2022. ISSN 0004-637X. doi: 10.3847/1538-4357/AC2D8D.
URL https://iopscience.iop.org/article/10.3847/1538-4357/ac2d8dhttps://
iopscience.iop.org/article/10.3847/1538-4357/ac2d8d/meta.
[28] Felix Richards, Adeline Paiement, Xianghua Xie, Elisabeth Sola, and Pierre Alain Duc. Panoptic
Segmentation of Galactic Structures in LSB Images. Proceedings of MVA 2023 - 18th Interna-
tional Conference on Machine Vision and Applications, 2023. doi: 10.23919/MVA57639.2023.
10216057.
[29] Prabh Bhambra, Benjamin Joachimi, and Ofer Lahav. Explaining deep learning of galaxy
morphology with saliency mapping. Monthly Notices of the Royal Astronomical Society,
511(4):5032–5041, 10 2022. ISSN 13652966. doi: 10.1093/mnras/stac368. URL https:
//academic.oup.com/mnras/article/511/4/5032/6529251.
[30] Eder C.M. Gomes, Paulo R.P. Amora, Elvis M. Teixeira, Antonio G.S. Lima, Felipe T. Brito,
Juliano F.C. Ciocari, and Javam C. Machado. UTTOS: A tool for testing UEFI code in
OS environment. Lecture Notes in Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics), 9976 LNCS(1):218–224, 12 2016.
ISSN 16113349. doi: 10.1007/978-3-319-47443-4{\_}14. URL http://stacks.iop.org/
0004-637X/798/i=1/a=7?key=crossref.ce8f697114eeb3752ff6da50e3e66986.
[31] Abien Fred Agarap. Deep Learning using Rectified Linear Units (ReLU), 3 2018. URL
http://arxiv.org/abs/1803.08375.
[32] Arjun Dey, David J. Schlegel, Dustin Lang, Robert Blum, Kaylan Burleigh, Xiaohui Fan,
Joseph R. Findlay, Doug Finkbeiner, David Herrera, Stéphanie Juneau, Martin Landriau,
Michael Levi, Ian McGreer, Aaron Meisner, Adam D. Myers, John Moustakas, Peter Nugent,
Anna Patej, Edward F. Schlafly, Alistair R. Walker, Francisco Valdes, Benjamin A. Weaver,
Christophe Yèche, Hu Zou, Xu Zhou, Behzad Abareshi, T. M. C. Abbott, Bela Abolfathi,
C. Aguilera, Shadab Alam, Lori Allen, A. Alvarez, James Annis, Behzad Ansarinejad, Marie
8
Aubert, Jacqueline Beechert, Eric F. Bell, Segev Y. BenZvi, Florian Beutler, Richard M. Bielby,
Adam S. Bolton, César Briceño, Elizabeth J. Buckley-Geer, Karen Butler, Annalisa Calamida,
Raymond G. Carlberg, Paul Carter, Ricard Casas, Francisco J. Castander, Yumi Choi, Johan
Comparat, Elena Cukanovaite, Timothée Delubac, Kaitlin DeVries, Sharmila Dey, Govinda
Dhungana, Mark Dickinson, Zhejie Ding, John B. Donaldson, Yutong Duan, Christopher J.
Duckworth, Sarah Eftekharzadeh, Daniel J. Eisenstein, Thomas Etourneau, Parker A. Fagrelius,
Jay Farihi, Mike Fitzpatrick, Andreu Font-Ribera, Leah Fulmer, Boris T. Gänsicke, Enrique
Gaztanaga, Koshy George, David W. Gerdes, Satya Gontcho A Gontcho, Claudio Gorgoni,
Gregory Green, Julien Guy, Diane Harmer, M. Hernandez, Klaus Honscheid, Lijuan (Wendy)
Huang, David J. James, Buell T. Jannuzi, Linhua Jiang, Richard Joyce, Armin Karcher, Sonia
Karkar, Robert Kehoe, Jean-Paul Kneib, Andrea Kueter-Young, Ting-Wen Lan, Tod R. Lauer,
Laurent Le Guillou, Auguste Le Van Suu, Jae Hyeon Lee, Michael Lesser, Laurence Perreault
Levasseur, Ting S. Li, Justin L. Mann, Robert Marshall, C. E. Martínez-Vázquez, Paul Martini,
Hélion du Mas des Bourboux, Sean McManus, Tobias Gabriel Meier, Brice Ménard, Nigel
Metcalfe, Andrea Muñoz-Gutiérrez, Joan Najita, Kevin Napier, Gautham Narayan, Jeffrey A.
Newman, Jundan Nie, Brian Nord, Dara J. Norman, Knut A. G. Olsen, Anthony Paat, Nathalie
Palanque-Delabrouille, Xiyan Peng, Claire L. Poppett, Megan R. Poremba, Abhishek Prakash,
David Rabinowitz, Anand Raichoor, Mehdi Rezaie, A. N. Robertson, Natalie A. Roe, Ashley J.
Ross, Nicholas P. Ross, Gregory Rudnick, Sasha Safonova, Abhijit Saha, F. Javier Sánchez,
Elodie Savary, Heidi Schweiker, Adam Scott, Hee-Jong Seo, Huanyuan Shan, David R. Silva,
Zachary Slepian, Christian Soto, David Sprayberry, Ryan Staten, Coley M. Stillman, Robert J.
Stupak, David L. Summers, Suk Sien Tie, H. Tirado, Mariana Vargas-Magaña, A. Katherina
Vivas, Risa H. Wechsler, Doug Williams, Jinyi Yang, Qian Yang, Tolga Yapici, Dennis Zaritsky,
A. Zenteno, Kai Zhang, Tianmeng Zhang, Rongpu Zhou, and Zhimin Zhou. Overview of the
DESI Legacy Imaging Surveys. The Astronomical Journal, 157(5):168, 4 2019. ISSN 00046256.
doi: 10.3847/1538-3881/ab089d. URL http://arxiv.org/abs/1804.08657.
[33] Michael J. Smith, James E. Geach, Ryan A. Jackson, Nikhil Arora, Connor Stone, and Stéphane
Courteau. Realistic galaxy image simulation via score-based generative models. Monthly
Notices of the Royal Astronomical Society, 511(2):1808–1818, 11 2022. ISSN 13652966. doi:
10.1093/mnras/stac130. URL https://arxiv.org/abs/2111.01713v1.
[34] Diganta Misra. Mish: A Self Regularized Non-Monotonic Activation Function. 8 2019. URL
https://arxiv.org/abs/1908.08681v3.
[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson,
Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick.
Segment Anything. 4 2023. URL https://arxiv.org/abs/2304.02643v1.
[36] Nelson Christensen, Renate Meyer, Lloyd Knox, and Ben Luey. Bayesian methods for cos-
mological parameter estimation from cosmic microwave background measurements, 10 2001.
ISSN 02649381. URL http://arxiv.org/abs/1110.3193.
[37] Mike Walmsley, Campbell Allen, Ben Aussel, Micah Bowles, Kasia Gregorowicz, Inigo Val
Slijepcevic, Chris J Lintott, Anna M m. Scaife, Maja Jabłon´ska, Kosio Karchev, Denise
Lanzieri, Devina Mohan, David O’Ryan, Bharath Saiguhan, Crisel Suárez, Nicolás Guerra-
Varas, and Renuka Velu. Zoobot: Adaptable Deep Learning Models for Galaxy Morphology.
Journal of Open Source Software, 8(85):5312, 2023. doi: 10.21105/joss.05312. URL https:
//doi.org/10.21105/joss.05312.
[38] Mike Walmsley, Inigo Val Slijepcevic, Micah Bowles, and Anna M. M. Scaife. Towards Galaxy
Foundation Models with Hybrid Contrastive Learning. Machine Learning for Astrophysics
Workshop at the Thirty Ninth International Conference on Machine Learning (ICML 2022), 6
2022. doi: 10.48550/arxiv.2206.11927.
[39] Tomasz Róz˙an´ski, Yuan-Sen Ting, and Maja Jabłon´ska. Toward a Spectral Foundation Model:
An Attention-Based Approach with Domain-Inspired Fine-Tuning and Wavelength Parameteri-
zation. 6 2023. URL https://arxiv.org/abs/2306.15703v1.
[40] Henry W. Leung, Jo Bovy, Henry W. Leung, and Jo Bovy. Towards an astronomical foun-
dation model for stars with a Transformer-based model. arXiv, 000:arXiv:2308.10944, 8
9
2023. doi: 10.48550/ARXIV.2308.10944. URL https://ui.adsabs.harvard.edu/abs/
2023arXiv230810944L/abstract.
[41] Inigo V. Slijepcevic, Anna M. M. Scaife, Mike Walmsley, Micah Bowles, O. Ivy Wong,
Stanislav S. Shabala, and Sarah V. White. Radio Galaxy Zoo: Building a multi-purpose
foundation model for radio astronomy with self-supervised learning. pages 1–13, 5 2023. URL
https://arxiv.org/abs/2305.16127v2.
10