MSG-GAN-SD: A Multi-Scale Gradients GAN for Statistical Downscaling of 2-Meter Temperature over the EURO-CORDEX Domain

Round 1

Reviewer 1 Report

This manuscript proposes a downscaling framework using multi-scale gradients GAN to downscale 2-m temperature data from ~ 83 km to 13 km spatial resolution over the European domain. The methodology proposed in this paper is novel, especially for Climate data, and will interest readers of AI. This work also can be applied to various data but not just restricted to climate variable data. However, I feel the paper's organization is hard to understand. Here are my detailed comments:

Writing:

1. The introduction lacks major motivation. Especially, the last part of the introduction reads more like a summary of the methodology (Lines 83-90). Please consider rewriting that part.
2. Section 2.1 Data sounds more like data preprocessing. This part assumes that the readers know about the model architecture (See lines 181-192). It would be helpful if these descriptions were available after the model architecture section. Maybe creating a separate sub-sections "Data preprocessing" after Section 2.2 would enhance readability.
3. Section 2.2 also requires a more sequential explanation. According to lines 199-201, the 2x4 images generated from the LR image are given as the input. But this is not explicitly stated in the paper. Please include the details/steps sequentially. Maybe a flowchart will be helpful.

Appendix B:

1. I could not understand the terms \lambda{tr} and \lambda{tr-cv}. How are these values chosen? What are the ranges of these values?
2. Also, in Lines 577-579, it is stated that the lambda values in Eqn (2) are set to 1. Please add this detail in the main text as well. If the lambda values are not equal to one, what are the criteria to choose these terms?

Minor Comments:

1. Line 16-17: Please rephrase this sentence
2. Line 38: Downscaling is not a "spatio-temporal process." Please rephrase
3. Line 171 – how are the low-resolution images undersampled? Do the authors use padding here as well?
4. Line 184: "Instead": I am unsure if this is a correct transitioning word in the context. Please consider rephrasing.
5. Line 188: There is no Section V. Please check
6. Line 240: I think "Regarding" instead of "As regards"
7. Too many section 2.3.1 for "Training Set Arrangements", "Training Configurations", and "the Validation Framework." Change section number.
8. Line 250 talks about season-based analysis. But this is not explicitly described before. But, the seasons are listed in Line 282. This should be done prior.
9. Line 297: "orthogonal couples of factors"?? I think it should couple.
10. It would be helpful if the authors could also include the formula of Pearson and Spearman correlation as well.
11. Line 412: I think "daily" should be "day"
12. Equation numbers (1) and (2) in appendix-B are the same as Equation (1) and (2) in the main text. This can be organized better.
13. Table A3 – Please separate the Monthly and Seasonal columns into two columns, namely "number of discriminator updates" and "Epoch." It is hard to read now.   

 

Author Response

The response  to Reviewer #1  is reported in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Thank you for inviting me to review the manuscript “MSG-GAN-SD: A Multi-Scale Gradients GAN for statistical downscaling of 2-meter temperature over the EURO-CORDEX domain” which intend to develop a new architecture tailored for statistical downscaling, called MSG-GAN-SD, allowing interpretability and good stability during training, thanks to multi-scale gradient information. Then, this architecture was applied to downscale ERA-Interim 2-m temperature fields, from 83.25 km to 13.87 km resolution, covering the EURO-CORDEX domain within the 1979–2018 period.

The work is well designed. The authors provided a good literature background and a good amount of detail for the implementation of their work. The manuscript is well written in its current form.

I just have a few minor recommendations to improve the manuscript as follows:

• Please introduce the abbreviation of Generator (G) and the Discriminator (D) in the text before using them (e.g. line 181, 185);
• Line 188: check the cross-reference “Section V”;
• Lines 287-289: please clarify the meaning of the 3rd phase “final test procedure” in the Validation framework. In the current form, it is not clear the meaning of this phase. Moreover, if you want to keep this phase, please add the subsection “3. Final test” as you did for “1. Best model selection” and “2. Evaluation procedure”. In the current form, there is an unbalance among the 3 phases. The same recommendation for section 3. Results and discussion since you report the results of the first 2 phases but not the third one.
• Line 324. 325: I suggest reporting the formula from these two publications instead of putting the reference so that readers can understand them directly from the manuscript.
• Line 568: where is the usefulness of Appendix C? I do not see any cross-reference in the main manuscript to this Appendix. Please insert the cross-reference where is it is applicable.

 

Author Response

The response to Reviewer #2 is reported in the attached file.

Author Response File: Author Response.docx