Assessing Galaxy Rotation Kinematics: Insights from Convolutional Neural Networks on Velocity Variations
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis is an interesting and useful paper on classifying galaxies on fast and slow rotators. The use of convolutional neural networks is well-motivated, and it proves useful for researchers working on this subject. I think the paper can be accepted as a contribution to the section ``Galaxies and Clusters''. Anyway, some suggestions could improve the paper:
Could the authors publish the code in Keras for the CNN they are using?
Either as an appendix or as extra material.
Have the authors compared with machine learning methods (random forests, for example)?
Sometimes, there are simpler machine learning methods that should be compared in speed and efficiency.
Author Response
Comment 1:
This is an interesting and useful paper on classifying galaxies on fast and slow rotators. The use of convolutional neural networks is well-motivated, and it proves useful for researchers working on this subject. I think the paper can be accepted as a contribution to the section ``Galaxies and Clusters''. Anyway, some suggestions could improve the paper:
Could the authors publish the code in Keras for the CNN they are using?
Response1:
>>>We appreciate the reviewer's suggestion. In response, we have made the code available in a public GitHub repository. The link to the repository has been included in the Data Availability Statement of the manuscript. In the GitHub repository we also put all python files and useful notebooks to facilitate reproducibility and further exploration of our implementation.
Comment 2:
Have the authors compared with machine learning methods (random forests, for example)?
Sometimes, there are simpler machine learning methods that should be compared in speed and efficiency.
Either as an appendix or as extra material.
Have the authors compared with machine learning methods (random forests, for example)?
Sometimes, there are simpler machine learning methods that should be compared in speed and efficiency.
Response 2:
>>>We appreciate the reviewer's valuable suggestion. In this study, our primary focus was on CNN-based classification; however, we fully acknowledge the importance of comparing different machine learning approaches. In our next paper, we plan to extend our analysis by comparing CNN with traditional machine learning classifiers such as random forests, support vector machines, and Naive Bayes in terms of both accuracy and computational efficiency. Additionally, we aim to explore model interpretability across these methods. For instance, in this study, we employed Integrated Gradients to interpret the CNN results. In our future work, we plan to apply techniques like LIME to machine learning classifiers such as random forests and compare their interpretability with CNN-based explanations. In the last paragraph of Sec. 4, we provide further explanation.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors use Convolutional Neural Network (CNN) to classify galaxies according to their stellar kinematic maps from the SAMI Galaxy Survey into slow and fast rotators. Such galaxy classification is crucial to our understanding of galaxy formation and evolution.
Achieving an accuracy and precision of approximately 91\% and 95\%, they apply their trained model to classify previously unknown rotator galaxies. Overall, the manuscript is well written, concise but presenting several details in their method and analysis. This study is an important precursors for the next-generation integral Field Spectrograph (IFS) surveys, highlighting the power of CNNs to improve our comprehension of galaxy dynamics. The publication in Universe - "Galaxies and Clusters" is justified, since the paper would be of great
interest for the large community investigating galaxy evolution. I have three major comments which I think will strengthen and complete the manuscript, with other few minor comments on the draft.
Major comments:
1. The authors note that they achieved the same conclusions using the MaNGA dataset intestead of the SAMI one. It would be useful to add some key plots also using only MaNGA and using MaNGA+SAMI. These would be important to address how singnal-to-noise impacts accuracy and precision of CNNs. How do Fig 1 and Fig 2 looks for MaNGA and MaNGA+SAMI?
2. It would be interesting to show a quantitative comparison with the current classification as a plot. How many unknown slow and fast rotators are recovered from the CNN that the classical method is not able to classify?
3. As a next step, it would be useful also to see the comparison with Jesse van de Sande et al. 2021 classification (https://ui.adsabs.harvard.edu/abs/2021MNRAS.505.3078V/abstract) - the classical method but also a comparison with the classification that uses higher-order kinematics.
Minor Comments:
-Lines 50-51 - please change references to the next-generation IFS surveys. For example, Hector (Bryant+2020,2024) instead of MaNGA which is now in the past.
-Line 70 - add citation for SAMI catalog and specify which data release has been used.
-Line 90 - SAMI was installed on the AAT.
-Line 104 - there are duplicates in the SAMI catalogue due to multiple observations. Which criteria have been applied to select the "best" datacubes/stellar kinematic maps and discard the other datacubes?
-Line 109 - explain the kinematic classification to which the flags are referring to. Add the proper citation. Do the results change if the classification from vandeSande et al. 2021 is used instead?
-Lines 157-158 - please add also the numbers of galaxies and not only the percentage.
-Line 206 - please specify numbers of galaxies in the training data, number of known slow and fast rotators.
-Line 241 - list numbers of unknown slow and fast rotators.
-Line 353 - MaNGA is a past survey and not a next-generation one, please provide a proper reference, such as Hector which will be characterised by 15,000 galaxies.
Author Response
• Comment1: The authors note that they achieved the same conclusions using the
MaNGA dataset intestead of the SAMI one. It would be useful to add some key plots also
using only MaNGA and using MaNGA+SAMI. These would be important to address how
singnal-to-noise impacts accuracy and precision of CNNs. How do Fig 1 and Fig 2 looks for
MaNGA and MaNGA+SAMI?
Answer1: >>>We appreciate the reviewer’s suggestion. We applied our CNN classifier to a
small sample of MaNGA data with known λ and ellipticity values. However, given the
limited sample size, we do not find it valuable to report these results at this stage. In the
next phase of our research, we plan to expand our analysis by incorporating additional
machine learning approaches, such as random forests, support vector machines, and Naive
Bayes, to evaluate their performance in terms of both accuracy and model interpretability.
Our goal is to integrate these methods alongside CNNs within a unified pipeline, ultimately
developing a library tool that can be applied to various datasets. Afterwards, we intend to
include MaNGA, as well as data from future surveys, to further investigate the impact of
signal-to-noise on model performance. In the second footnote of page 2, we provide further
explanation.
• Comment2: It would be interesting to show a quantitative comparison with the current
classification as a plot. How many unknown slow and fast rotators are recovered from the
CNN that the classical method is not able to classify.
Answer2: >>>In Figure 3 of the paper, we present a distribution plot comparing the CNN
model’s predictions against the real test data. The test data consists of labeled samples,
where fast and slow rotators are classified based on the classical method. Notably, out
of 1,112 samples in the SAMI catalogue, 854 were originally unclassified by the classical
approach. However, our CNN model successfully assigned classifications to these previously
unknown cases. To reflect this, we have added a relevant sentence at the beginning of
second paragraph of the Section 3.2.
• Comment3: As a next step, it would be useful also to see the comparison with Jesse van
de Sande et al. 2021 classification https://ui.adsabs.harvard.edu/abs/2021MNRAS.
505.3078V/abstract - the classical method but also a comparison with the classification
that uses higher-order kinematics.
Answer3: >>>We appreciate the reviewer’s valuable suggestion. The classification method
presented in Jesse van de Sande et al. (2021) provides an important non-ML approach that
we can consider for future work. In our next study, we plan to explore a broader comparison
by including classical machine learning methods such as Random Forest, Na¨Ä±ve Bayes,
and Support Vector Machines, alongside our CNN model. This would allow for a more
comprehensive evaluation of different classification techniques, including both traditional
statistical methods and deep learning approaches.
• Comment4: Lines 50-51 - please change references to the next-generation IFS sur-
veys. For example, Hector (Bryant+2020,2024) instead of MaNGA which is now in the past.
Answer4: >>>They are fixed now.
Comment5: Line 70 - add citation for SAMI catalog and specify which data release has
been used.
Answer5: >>>We have added a citation for the SAMI survey. We previously included a reference
to Section 2 to provide more details on the data source, preprocessing steps, and dataset
preparation.
• Comment6: Line 90 - SAMI was installed on the AAT.
Answer6:>>> It is fixed now.
• Comment7: Line 104 - there are duplicates in the SAMI catalogue due to multiple
observations. Which criteria have been applied to select the ”best” datacubes/stellar
kinematic maps and discard the other datacubes?
Answer7: >>>We have resolved the issue of duplicated labels in our dataset by eliminating any
redundant entries from our label array, ensuring that each galaxy in our sample is uniquely
represented. Additionally, we have included further explanations in the second paragraph
of Section 2 for clarity.
• Comment8: Line 109 - explain the kinematic classification to which the flags are referring
to. Add the proper citation. Do the results change if the classification from vandeSande et
al. 2021 is used instead?
Answer8: >>>Based on the spin-ellipticity relations established by Emsellem et al. (2011)
we classified each galaxy as either a fast rotator (labeled as 1) or a slow rotator (labeled
as 0). To provide further clarity on this classification process, we have added additional
explanations in the second paragraph of Section 2.
• Comment9: Lines 157-158 - please add also the numbers of galaxies and not only the
percentage.
Answer9: >>>We have added some sentences at the beginning of the second paragraph in
Section 2.2 to provide further clarification.
• Comment10: Line 206 - please specify numbers of galaxies in the training data, number
of known slow and fast rotators.
Answer10: >>>We have added more details about the training data in Section 3.1.
• Comment11: Line 241 - list numbers of unknown slow and fast rotators.
Answer11:>>> We appreciate the reviewer’s insightful comment. In response, we have added
a new figure to the paper (Figure 6), which presents a histogram of the CNN prediction
probabilities (P ) for the unknown rotators. This figure highlights the distribution of
predicted probabilities, with fast rotators (P ≥ 0.95) and slow rotators (P ≤ 0.95).
Additionally, we have included the exact numbers of unknown fast and slow rotators in
Section 3.2 to provide a clearer quantitative analysis.
• Comment12: Line 353 - MaNGA is a past survey and not a next-generation one, please
provide a proper reference, such as Hector which will be characterised by 15,000 galaxies.
Answer12: >>>It is fixed now.
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors changed the manuscripts according to the suggestions and I do not have any more comments.