Deep Learning Investigation of Mercury’s Explosive Volcanism

Round 1

Reviewer 1 Report (Previous Reviewer 3)

 The main comments are as follows.

The second paragraph in Introduction should be shortened. In third paragraph, the authors should identify why to study explosive volcanism and how to study it previously? What is the geological relationship between the explosive volcanism and the vents and pyroclastic deposits produced? Moreover, “…defining the shape and surface area of 55 deposits covering 110 vents”, what is the scientific background of this expression?

Lines 28-29: “However, the 28 transfer of knowledge is not immediate, and the gap broadens when addressing planetary remote 29 sensing data”, what does the expression mean? As far as I know, the deep leaning method has been widely used in the information extraction using the planetary remote sensing data.

The contents in lines 79 to 133 are lengthy, which must be shorten.

The contents in Pages 2 to 5 (before section 2.3) should be shorten, and the interpolation is not the main work of this paper. Moreover, in section 2.3, the content about the construction of the architecture is poor.

Line 201: what is explosive volcanic vents? Are there any other vents on the mercury? 

Section 3: The content about the Training the Deep Neural Network is poor. Here, what is the reason to analyze the Effect of patch size and the Effect of the latent space dimension and number of clusters in this section? Besides, what is the actual input and output of the network? How to train and use it?

Sections 4 and 5: I do not understand why the authors only paid attention to the extraction of the pyroclastic deposits and its area. Moreover, what is the geologic significance of the deposit areas?

Section 6: The content should be shorten into the concise expressions. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

Review: Deep Learning Investigation of Mercury’s Explosive Volcanism

 

The manuscript describes how the authors make use of a deep learning algorithm to study the extent of pyroclastic deposits.

The manuscript gives a very detailed description of the used methods (around half of the entire manuscript). From this, the first half of the manuscript reads somewhat lengthy. However, I think such deep learning approaches are not yet widespread in the scientific community (I have also no deeper expertise in this field). Therefore, I think, the manuscript can be used as kind of a “manual” of how to make use of such an algorithm in remote sensing. This is important because it makes such techniques “available” for the scientific community.

The second part of the manuscript describes the results gathered with the new method. This part shows that the machine-based method to determine the extend of the pyroclastic deposits gave other results than classical methods. A flaw for me is, that the differences of the results are not (cannot be?) explained because the exact function of the deep learning algorithm is not known. For me, it would be interesting to know which exact spectral or special feature(s) in the raw data are responsible for the discrimination of the deposits and the background rocks. However, the results gathered with the new method seem to be more correct or consistent than the IAU-values.

The manuscript is well-written.

 

Maybe, the authors can think of putting the whole first part in the supplement and only describe the results in the main text. That would be helpful for the readers that are only interested in your results and those who want to learn how deep learning can be implemented can read the supplement.

 

All over, I think, the manuscript is worth for publication of minor changes.

 

 

Detailed comments:

Line 409 “wince” must be since

Figure 11. I cannot see a red circle

Line 571 “tot” must be “to”

Line 656 you may add a reference for the BepiColombo mission

 

 

 

With kind regards

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report (Previous Reviewer 3)

The authors gave  good responses to my comments. The revision of the manuscript is proper. I have no new comments.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Fig. 1 please identify the couple four images with the relative number in the figure too.

Fig. 5 - please if you are able to erase the cute MISTRAL word it would be great.

 

Reviewer 2 Report

Dear Authors

I appreciate the authors' efforts in applying ML techniques on the surface of Mercury. It clearly shows the authors are sincerely engaged in the topic.

I would like to point out those for the improvement of the draft.

1) First of all, the draft is very verbose. I can't see any reason to represent the procedure and result in an excess of 21 pages.  

- Introduction should be far shortened event including parts of the results (some part of the result section should be included here with a more concise form).

- Section 2 . Introduce a flowchart to summarize the method and processing. The reader does not need a textbook-style statement of the backgrounds.

- I really DO NOT believe your draft has the deep contributions to be stated for more than 10 pages in sections 2 and 3.

- Compared to huge descriptions of technical background, I can't find an adequate reason why your scientific finding requires deep learning (in other words the merits of deep learning to have your results rather than conventional hyperspectral analyses or even visual interpretation).

- Does all clustering methods in section 3 have any reason to be combined in the concept of deep learning? From my point of view, those are just applications of clustering algorithms combined with simple gridding methods.

Otherwise, you need to state how your deep-leaning is superior than the conventional approach.

- Thus the scientific finding in Section 4  is based on the achievement of the deep-learning approach. From my point of view, it is not clear why the interpretation of this section is available on the deep-learning approach. I mean the lack of connection between methodology and outcomes. 

- If you want to draw a scientific conclusion in section 4, it will be better to interpret the finding together with topographic information such as LIDAR DEM or other resources. Otherwise summarize please make clear unique advantage of your approach together with scientific finding  in sections 4 and 5 deleting unnecessary verbose statements.  

Other points

- L51: ref [10] missing

- L89: As with some other of your writing,  tautological expressions, Please check the whole writing.

- L148 : returned?

- L151-155: Rather than those statements, a simple diagram of sensor/spectral response will be better. So many such cases

- after L187 : IDW is a simple grinding method. Do you need these long statements?

- after L364 : BriCH etc.. put acronyms with references. State simple those as part of your deep-learning scheme

- after L430: Many statements  need to go introduction and method

- after L490: I can't see the bases on which the statements were driven.

Best Regards

Please see  Comments and Suggestions for Authors. Writing is very verbose, 

Reviewer 3 Report

Understanding the origin of volatiles and identifying opaque phases on the surface of Mercury are important tasks to be completed to understand the planet's building blocks and early evolution. In this study, the authors investigated the application of unsupervised deep learning to explore the 6diversity and constrain the extent of the Hermean pyroclastic deposits. Finally, they found the 3DCAE an effective technique to analyse sparse 13observations in planetary sciences. This is a technique paper to explore the surface deposits over the planetary surface using the deep learning method.

However, several expressions in the paper are wrong. The main comments are as follows.

(1)The authors’ understanding about the pyroclastic deposit is problematic. Then, the authors did not provide the real extent of the deposit in the paper. The color of pyroclastic deposit should be dark, but it is not provided in the paper. I suggested to add the spectral analysis of the deposit. So that, as a method, this is a good attempt to explore the surface deposits; but this paper is failed to extract the pyroclastic deposit.

(2) “The working principle of 246an autoencoder is to learn the representative features from an input data set by first encoding the 247data into a lower dimension "latent" representation and then decoding this reduced representation 248aiming to reconstruct the input data.”  is this a supervised or unsupervised method? “….trained with 80% of the data and validated with 424the remaining 20%...”? where are the training data?

(3)Line 258: “= (184×1×1×48). The dimensionality is then reduced following a three dimensional pooling 258operation to size (10×1×1×48) and” What method are used to reduce the dimensionality of the original data. Did it use in the following processing? If not, please delete it.

(4) Line 268: ” In the later case, the spectral dimension is less complex (typically 268RGB images) and the spatial resolution is very high”, the figure used in the paper is not high.

(5)Figure 1 is problematic.

(6)“Figure 5 illustrates the effect of the patch size on the final cluster maps, through an example 306 for the pyroclastic deposit on the Mistral crater.”Where is the pyroclastic deposit in Figure 5?  (1)What is the spectral features of the pyroclastic deposit? (2) Where is the real location of the pyroclasltic deposit? The accurate location is significant to extract the deposit. The cluster maps show a clear different from the geologic map, maybe hinting the bad cluster results?

(7)It is hard to understand why the authors define the number of the clusters and the space dimension.

(8)“The reflectance spectra of Mercury 415present no absorption bands and are mainly characterised by the spectral slopes and the reflectance 416at certain wavelengths”. The expression is wrong. Correspondingly, the expression related to the DBSCAN technique is wrong.