Source-linked AI summary
Improving galaxy morphologies for SDSS with Deep Learning
H. Domínguez Sánchez, M. Huertas-Company, M. Bernardi, D. Tuccillo, J. L. Fischer
TL;DR
Large galaxy surveys require scalable morphological classification because visual labels are time-consuming and can be uncertain, especially for features such as bars and mergers. The paper trains CNNs on existing visual catalogues to produce GZ2-type probabilities and T-Types for approximately 670,000 SDSS galaxies. The resulting catalogue achieves high same-distribution GZ2 performance and provides a substantially larger T-Type sample with low offset and scatter.
Problem
Visual galaxy classification is difficult to scale and can contain uncertainty or bias, motivating accurate automated classifications for large samples.
Method
CNNs are trained on GZ2 and Nair & Abraham visual catalogues using SDSS colour images to predict GZ2 questions, T-Types, and supplementary classifications.
Results
More than 97% accuracy is achieved for GZ2 questions on a test sample matching the robust training classifications, while T-Types show b=0.03 and σ=1.1 against Nair & Abraham classifications.
Takeaways & Limitations
The publicly released catalogue provides homogeneous GZ2-type and T-Type classifications for 670,722 galaxies, roughly three times the GZ2 statistics and about 50 times the Nair & Abraham T-Type sample.
Takeaways & Limitations
GZ2 bar and merger labels retain identification limitations, and the supplied PS0 probability should only be used for galaxies with T-Type ≤0.
Abstract
from arXiv · showhide
We present a morphological catalogue for $\sim$ 670,000 galaxies in the Sloan Digital Sky Survey in two flavours: T-Type, related to the Hubble sequence, and Galaxy Zoo 2 (GZ2 hereafter) classification scheme. By combining accurate existing visual classification catalogues with machine learning, we provide the largest and most accurate morphological catalogue up to date. The classifications are obtained with Deep Learning algorithms using Convolutional Neural Networks (CNNs). We use two visual classification catalogues, GZ2 and Nair & Abraham (2010), for training CNNs with colour images in order to obtain T-Types and a series of GZ2 type questions (disk/features, edge-on galaxies, bar signature, bulge prominence, roundness and mergers). We also provide an additional probability enabling a separation between pure elliptical (E) from S0, where the T-Type model is not so efficient. For the T-Type, our results show smaller offset and scatter than previous models trained with support vector machines. For the GZ2 type questions, our models have large accuracy (> 97\%), precision and recall values (> 90\%) when applied to a test sample with the same characteristics as the one used for training. The catalogue is publicly released with the paper.
1 INTRODUCTION
Galaxy morphology spans a continuous range of bulge-, disk-, spiral-, and irregular-dominated forms, yet accurate classification is difficult to scale visually. This motivates automated methods that can handle large surveys while preserving useful morphological information.
- Morphological classification: Hubble’s scheme separates galaxies into early types dominated by bulges and late types with significant disks, including barred and unbarred spirals, S0s, and irregulars.A numerical T-Type can be assigned to each morphological type.
- Scientific motivation: Morphology correlates closely with stellar properties, including colour, mass, velocity dispersion, stellar population age, gas content, and rotation.These structural and intrinsic properties also evolve across cosmic time.
- Challenges: Visual classification is difficult because galaxy types transition smoothly and high-redshift images have poorer quality alongside structural evolution.These factors make subclass boundaries non-obvious.
- Challenges: Visual classification is prohibitively time-consuming for surveys containing millions of galaxy images.The problem affects large surveys such as SDSS, the Dark Energy Survey, and EUCLID.
- Existing approaches: Galaxy Zoo expanded citizen classification from three broad categories to the more complex GZ2 decision system, but volunteer classifications have caveats such as missed weak bars and uncertain intermediate choices.A separate prior approach used morphology-correlated parameters and nonlinear boundaries in parameter space.
- Deep learning motivation: CNNs offer a natural automated approach because they learn high-level image features directly from pixels instead of requiring a pre-selected parameter set.Earlier CNN work reproduced GZ2 classifications but could inherit biases from uncertain visual labels.
- This work: The study combines GZ2 and Nair & Abraham visual catalogues with CNNs to produce GZ2-type classifications and T-Types for a substantially larger SDSS sample.The catalogue also includes an E-versus-S0 probability and an alternative bar classification.
2 DATA SETS
The study trains and evaluates models using complementary visual catalogues and applies them to a large SDSS parent sample. The released sample benefits from detailed classifications and improved photometric measurements.
- Datasets: The training and testing datasets comprise GZ2, Nair & Abraham, and additional catalogues not used for training.Tests combine evaluations on training catalogues with independent available catalogues.
- Training catalogues: GZ2 provides volunteer classifications for approximately 240,000 SDSS DR7 galaxies through an 11-question decision tree with 37 possible responses.The catalogue records weighted and debiased vote counts and answer fractions.
- Training catalogues: Nair & Abraham classifies 14,034 SDSS-DR4 galaxies using expert visual inspection of monochrome g-band images and includes T-Types plus detailed structural features.Its bar labels distinguish strong, intermediate, and weak bars.
- Testing catalogues: The Cheng et al. catalogue supplies an independent test set of 984 non-starforming SDSS galaxies focused on distinctions among E, S0, and Sa-like bulge classes.It includes visual and automated classifications.
- Released sample: The released parent sample follows Meert et al. and contains improved two-dimensional multi-band decompositions and added stellar masses.These measurements address known SDSS pipeline photometry underestimates for luminous galaxies.
3 DEEP LEARNING MORPHOLOGICAL CLASSIFICATION MODEL
The model uses CNN-based deep learning to classify SDSS colour images, with a common architecture across tasks and outputs tailored to binary GZ2 questions or continuous T-Types.
- Model approach: Deep learning automatically extracts relevant features from raw data through nonlinear transformations, avoiding pre-processing based on hand-selected parameters.The approach requires many already classified images for training.
- Architecture: The study uses a similar CNN configuration for every classification task rather than comparing different network architectures.The models are implemented with the KERAS library.
- Image preprocessing: The input consists of 424×424-pixel RGB SDSS cutouts down-sampled to 69×69×3 flux matrices.Down-sampling reduces computing time and helps avoid over-fitting.
- Outputs: Binary GZ2 models output probabilities from 0 to 1, whereas the regression model outputs T-Type values ranging from -3 to 10.The T-Type output is directly interpreted as the galaxy’s T-Type.
4 GALAXY ZOO 2 BASED MODELS
The GZ2 models independently classify selected morphology questions using robust volunteer classifications, while restricting the decision tree to statistically and resolution-supported tasks. The models achieve strong performance on robust test samples, with clearer separation for extreme classes but greater uncertainty for intermediate categories.
- Training methodology: The models use GZ2 classifications with low uncertainty to train independent binary questions rather than reproduce the full decision tree.The analysis excludes lower-level tasks with fewer robust examples and omits spiral-arm signature and bulge-shape questions because the binned input resolution is insufficient.
- Training methodology: Training requires P > 0.8 in one binary answer and at least five votes, relaxed to P > 0.7 when statistics are limited.This selection removes noisy human classifications and supports faster model convergence.
- Training methodology: The binary models use binary cross-entropy, the Adam optimizer, and sigmoid activation, with outputs interpreted as probabilities of positive morphology.Each output ranges from 0 to 1 and represents the degree of the corresponding feature, such as bulge importance or roundness.
- Training methodology: The addressed questions include disk/features, bulge prominence, and mergers, with training availability constrained by the GZ2 decision-tree paths and positive-example counts.For example, bulge prominence is asked only after disk/features and non-edge-on selections, while fewer than 7% of galaxies have more than five merger votes.
- Results: More than 96% accuracy is obtained across the binary questions, reaching 98% for Q1, while precision and recall both exceed 0.9 except for the bar task.The bar task reaches a maximum of approximately 0.8 for both precision and recall, with global accuracy of 96.6%.
- Results: Extreme morphology classes are clearly separated, whereas intermediate categories show broader probability distributions and larger uncertainty.For bulge prominence, bulge-dominated galaxies have 2% false negatives and fewer than 0.1% false positives; intermediate bulges span all probability values. For cigar-versus-round classification, false positives and false negatives are each below 0.1%, but 27% of intermediate cases have Pcigar < 0.1.
5 N10 BASED MODELS
The N10-based models extend the catalogue with continuous T-Types, an E–S0 probability, and an alternative bar classifier. They improve T-Type agreement over the M15 comparison and provide useful but bounded performance across these morphology tasks.
- T-Type model: The T-Type model uses N10 visual classifications to learn a continuous structural sequence spanning early types, spirals, and irregular galaxies.The training set contains 10,000 galaxies with certain classifications, and the model uses mean squared error for linear regression.
- T-Type model: 0.03 median offset is achieved against N10 for T-Type ≤6, while the model fails to converge at higher values where statistics are scarce.The test sample contains approximately 500 galaxies not used for training.
- T-Type model: σ=1.1 average scatter is lower than M15’s σ=1.4 for T-Type ≤6, and the resulting catalogue contains approximately 50 times more classified galaxies than N10.The model’s galaxies follow the expected smooth transition from elliptical to spiral morphologies.
- Ell versus S0 models: The additional PS0 model distinguishes pure ellipticals from S0-related systems, with only 6% of N10 T-Type=-5 galaxies assigned PS0 > 0.5.For intermediate S0 classes, PS0 spans nearly the full probability range, while most S0/a galaxies receive high PS0.
- Ell versus S0 models: 95% of Cheng BC=3 galaxies have PS0 > 0.5, compared with 11% for BC=1 and 62% for BC=2, consistent with Sa, E, and intermediate classes.The classification has higher purity and completeness than Cheng et al.’s visual classifications, and is more efficient than its automated method for distinguishing E from Sa.
- Barred galaxies: The alternative N10-based bar model correctly classifies 90% of unbarred galaxies and reaches 88% and 80% true-positive rates for intermediate and weak bars at Pthr > 0.4.Its Pbar distribution for weak bars is shifted lower than for stronger bars, indicating that Pbar may serve as a proxy for bar strength.
6 COMPARING THIS CATALOGUE WITH THE GALAXY ZOO 2
The catalogue produces probability-based GZ2-type classifications for 670,722 galaxies and generally agrees with GZ2 while providing more decisive outputs. Comparisons with morphological parameters and targeted examples support the physical relevance of the probabilities, but merger classifications and some hierarchical questions require caution.
- Catalogue content: 670,722 galaxies receive a probability value for each catalogue question, with user-selected Pthr values controlling positive-example selection.The catalogue increases the available statistics relative to GZ2 and reports precision and TPR values for threshold choices.
- Probability validation: Pbulge correlates more strongly with bulge-to-total ratio and Sérsic index in the new catalogue than in GZ2.Pcigar and Pedge also correlate with ellipticity for both catalogues.
- Uncertain GZ2 cases: The models produce high probabilities for some galaxies with uncertain GZ2 classifications, including disk/features, edge-on and bar-signature examples.The examples use model probabilities above 0.9 and GZ2 agreement values below 0.25.
- Agreement with GZ2: 2.5%, 1.7% and 1.9% are the catastrophic-error fractions for Q1, Q2 and Q3, respectively, remaining below 3% for each task.These fractions count cases where one classification probability exceeds 0.8 while the other is below 0.4.
- Agreement with GZ2: 7.2% is the discrepancy fraction for Q6 mergers, for which comparison with GZ2 is especially difficult.The paper explicitly highlights the difficulty of interpreting this comparison.
- Caveats: GZ2 vote-independent probabilities still reflect selection effects because upper-level questions were trained only on galaxies receiving at least five relevant votes.Probabilities for such questions should be trusted completely only for positive examples of the corresponding answer.
- Caveats: PS0 should be applied only to galaxies with T-Type ≤0, while Pmerger may better trace clustered or projected pairs than ongoing mergers.The paper attributes the merger limitation to the small number of merger examples.
7 CONCLUSIONS
The paper uses CNNs trained on robust visual catalogues to build GZ2-type and T-Type classifications for a large SDSS sample. The models achieve high performance on robust GZ2 cases, while the T-Type model shows low offset and scatter and the catalogue substantially expands existing coverage.
- Method: CNNs are trained on GZ2 and Nair & Abraham visual catalogues to classify approximately 670,000 SDSS galaxies.The GZ2 models cover multiple questions, while Nair & Abraham supplies the T-Type training labels.
- GZ2 models: Each GZ2 question is trained independently using only galaxies with a(p) ≥0.3 and binary yes/no outputs between 0 and 1.This training design targets galaxies with small GZ2 classification uncertainties.
- GZ2 results: >97% accuracy is obtained on test samples matching the robust training classifications, with precision and TPR >90% for each question except bar signatures.For bar signatures, both precision and TPR reach approximately 80%.
- GZ2 results: 56% to 86% is the increase in galaxies with a(p) >0.3 for disk/features classifications when comparing GZ2 with the new catalogue.The catalogue provides homogeneous classifications for 670,722 galaxies.
- T-Type results: b=0.03 and σ=1.1 describe the T-Type model's offset and scatter relative to Nair & Abraham classifications.The catalogue is approximately 50 times larger than the previous Nair & Abraham T-Type catalogue and includes an E/S0 separation model.
- Implications: Applying the trained models to unclassified SDSS images is straightforward and not time-consuming, enabling planned extension to other SDSS samples.The paper specifically mentions the MaNGA dataset as a forthcoming application.