A Generative Adversarial Network for Pixel-Scale Lunar DEM Generation from High-Resolution Monocular Imagery and Low-Resolution DEM
Round 1
Reviewer 1 Report
This paper proposes a relevant solution to avoid the cost of high resolution DEM production, by taking advantage of available data. The presentation of the method is complete and the experiments are convincing. However, there are some aspects that need minor adjustments or clarifications:
Line 63-64 – this sentence is essential to understand the rest of the article. The idea should be stated more clearly.
Line 80 – "fair depth" : be more explicit
Line 264 – I understand that x and y represent two planimetric axes so "horizontal and vertical" is not appropriate and should be replaced by "directional" for example. Clarify "gradient differences" (in terms of slopes)
Line 368-369 – These metrics are very poor. Probably they do not allow to detect anisotropies, or too noisy or too smooth DEM. This limitation should at least be mentioned in the discussion.
Line 465 – DEM "comparable" with an image : clarify
The concept of resolution is addressed in a misleading way. There are almost 100 occurrences of the term "resolution", most often in the expressions high- or low-resolution, i.e. a qualitative concept which is well understood, but in other cases the term "resolution" is used with the meaning of "pixel size" or equivalent, which is a bit confusing. Indeed, pixel size is somewhat arbitrary, while the resolution is the real performance of an oversampling method to recover small terrain details. This abuse of language has no impact in some works, but in this one it should be avoided because the objective of the method is an improvement of the resolution.
A few more details :
Concerning numerical writing: insert a space between the number and the unit when it has been omitted
Reference 6 is not complete
In figure 5 and 8 : replace "difference" by "absolute difference" (since it seems to be the case)
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 2 Report
The proposed generative-adversarial-network(GAN) based method seems to well support generating high-resolution pixel-scale DEM from a single image aided by a low-resolution DEM as revealed from the experiments and the analyses. With that, the following questions are related to some technical issues which need to be further verified or explained before the merit of this study can be truly valued:
1.It is not clear about how the DEMs and images are registered? And is there any indicator for the registration quality? If there still exist significant misregistration between images and DEMs, how would it affect the reconstruction? Or how the proposed approach would tackle such problem?
2.Similarly, what is the actual mechanism to not let the reconstruction result inherit the severe systematic bias of DEM, like the sub-area IV?
3. Following the above context, the scale recovery by using Eq.(13) would be negatively affected by data with huge errors. Is there any treatment considered for error detection and correction?
4. How to determine the α, β, γ of Eq. (11)?
Last but not least:
English check: line 251: is consisted of; line 443: As can see.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 3 Report
The authors of the manuscript propose a methodology for reconstruction of DEM based on a single high resolution image using GAN. It is an interesting topic as also the authors have pointed out that there is a high number of high resolution images of the lunar surface, while the DEM products with relatively low, mainly due to the lack of stereoscopic camera on board of the satellites in lunar orbit.
The work presented here seems interesting, however, in my opinion it should pass a significant revision to make the manuscript more easy to be comprehended by the reader. For example, one my main suggestion is that the manuscript to be edited by a native speaker. Who will be able to edit it in such way that the excessive narrative parts are removed and most of the paragraphs are straight to the point. Some examples but not only (l.290-293;311-319;522-525).
Please keep consistent while referring to a figure (Figure and Fig.) is seen at many places (check template).
Figure 1 and sect 3.1 - not really, should be improved to be presented in much clearer manner.
Figures 3, 5, 6, 7, 8 - Some of them are with low quality, please fix. Areas I-IV are barely visible in fig.5, please fix. All of them should have subfigure naming, for example Figure 5a, Figure 7h,etc. And they should be properly included into your manuscript, otherwise it is really hard to follow your discussion and in the meantime search for the figure you've mentioned!
Check for minor spelling errors- example l.489 LROC.
Where is your discussion? - you should clear discuss your results, discussion similar works; advantages and what can be improved, etc.
regards.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors have well addressed the issues I mentioned and responded to each point with elaboration on solutions and future work. I have no further comments and suggestions.
Reviewer 3 Report
In my opinion, the manuscript has been improved.
Regards.
