Comparing ML Methods for Downscaling Near-Surface Air Temperature over the Eastern Mediterranean

Round 1

Reviewer 1 Report

Comments to the Author

This study focused on comparing XGBoost and Deep Learning methods for downscaling near-surface air temperature over the eastern mediterranean. While the structure of the paper is satisfactory, there are significant shortcomings, particularly regarding the introduction, methods and results section, that need to be addressed before considering it for publication.

Comments:

1.The introduction provides sufficient references and background, but they are very long. I would suggest some parts merging in the introduction to better express ideas more concisely and clearly.

2. Please provide a detailed explanation for Figure 1.

3. The text in most of the figures is barely readable. The resolution of the figures should be improved. 

4. Compared with different ML methods, how about these ML models' complexity, parameter number and inference time?

5. It is known that the performance of Deep Learning generally will be better than XGBoost, so I think there is no comparison between XGBoost and Deep Learning methods.

6. The result section's content is insufficient. It should be added or expanded.

Author Response

Thanks for your good comments.  

Please see reply in attached pdf. 

Author Response File: Author Response.pdf

Reviewer 2 Report

This article employed machine learning and deep learning algorithms to downscale two air temperature products to the resolution of 1 km. The content of this article is novelty for which the advanced downscaling techniques of air temperature. Nevertheless, I think this article is hard to read for the scholars, and many results I hope to read are not provided. I would like the author to meticulously revise the article before discussing whether or not to publish the article. Specially, there are some typically issues that need to be answered or corrected:

1. Although the data generated via the Global climate model often present a coarse resolution, the most vital issue for the model is lack for the influence on man-made scenarios, like cities. This article uses MODIS LST to downscale air temperature, which is theoretically able to consider this issue. So I suggest the author make the validation attend to regions related to human activities, which may be a highlight of this article.

2. The author should polish the contents. The language is not clear (or smooth) in many places.

3. I suggest the author rewrite the section "Introduction". Actually, I don't understand why this work is necessary.

4. Why does the author divide "historical data" and "future data"? What is the difference between them? Also, the goal of this article only downscaled two reanalysis products, what the mean of "future data"?

5. The author should correct the unprecise statements in this article, such as:

+ line 72, "classical machine learning technique", is this statement correct?

6. Why each figure is blurry? I didn't read anything from all of these figures! Additionally, several figures do not prove the fluctuation relationship between the prediction and ground truth.

7. How to match the time between reanalysis and MODIS satellite? The detail should be provided.

8. Why the author express "Our model"? To my knowledge, the work of this article only makes a comparison between the data product in accordance with existing machine learning algorithms, so is this article presented the improvement or advancement for the algorithm? If it is, please explicitly illustrate the process. Moreover, I think the author should draw the flowcharts of selected downscale algorithm, especially for deep learning model.

9. The result maps of air temperature should be provided, after which downscaled.

The author should polish the contents. The language is not clear (or smooth) in many places.

Author Response

Thanks for your good comments.  

Please see reply in attached pdf. 

Author Response File: Author Response.pdf

Reviewer 3 Report

Thank you for inviting me to review the entitled manuscript "Comparing ML Methods for Downscaling Near-Surface Air Temperature over the Eastern Mediterranean". To obtain fine-resolution near-surface air temperature (Ta), this study employ two different machine learning algorithms (XGBoost and Deep Learning) to downscale the resolution of ERA5 (9 km) and CMIP6 (27 km) Ta to 1 km, and to extend the application for future climate predictions, which is innovative to some extent but still has some issues:

1. All figures in this article are coarse- resolution, so it is difficult to see clearly what is in the figure, so that it seriously affects the reading, such as figure1, figure4, figure5, figure6, and figure7. Please carefully check that the resolution of these figures meets the requirements of the journal.

2. This paper mentioned “We added MODIS daily 1 km LST and 250 m NDVI to the hourly dataset after up-sampling to hourly resolution using mean interpolation”. How to obtain hourly LST and NDVI data by means of average upsampling?

3. The input of high-resolution auxiliary variables is crucial to the downscaling results. In this work, the author uses a variety of features to train different models to achieve downscaling. However, we find that except the resolution of DEM and MODIS data >=1km, other features come from ERA5 and CMIP6 that have coarse resolution, how to align these features with great differences in resolution scale？

4. This paper chosen RMSE to estimate the models’ performance, which is reasonable. However, RMSE is sensitive to outliers, the model evaluation is inaccurate. It is recommended that multiple evaluation indexes, such as MAE and R square, can be used together.

5. In this paper, ERA5 is used in training the models for downscaling historical data, and daily predictions models are made since CMIP6 does not have hourly resolution. As far as I know, while ERA5 has daily data, why not uniformly make daily downscaling forecasts to avoid inconsistency in time resolution?

6. The RMSE in Table 1 and Table 3 requires units, please check carefully and indicate the units.

7. The RMSE in Figure 4 and 5 requires units, please check carefully and indicate the units.

8. The key words of the article is up to ten, and it is suggested that it can be appropriately reduced.

9. It will be helpful to explain the meaning of different colour lines in Figure 1, please add some description.

10. The font of the picture in the paper seems to be inconsistent with the text font, please check and suggest that the font be consistent.

11. Ta stands for near-surface air temperature, and its unit should be ‘°C’，not ‘C’，please check it in Figure 1 and correct it.

12. In the third paragraph of discussion, there is an abbreviation called "RMASE", but it is not seen in the full paper. May I ask whether the RMASE is incorrectly written? Please check and correct it!

Minor editing of English language required.

Author Response

Thanks for your good comments.  

Please see reply in attached pdf. 

Author Response File: Author Response.pdf

Reviewer 4 Report

The study compared the performances of two machine learning methods for downscaling near surface air temperatures. It probed the potential of incorporating satellite data and topographic information into the algorithms and also evaluated the accuracy of downscaling the projections of GCM models for climate change scenarios in future. The results of the study are encouraging and worth of publication in Remote Sensing. As far as I’m concerned, some issues of the study need to be clarified before the acceptance of the ms.

1.       To check if the inclusion of satellite data and topographic information contributes to the improvement of the downscaling accuracy, you need to compare the results of with and without the satellite data and topographic information. In the present study, there is no comparison.

2.       Table 1 shows the RMSEs of other studies. In Table 3, the RMSEs were lower than those in Table 1 if considering only the no-LOGO. I wonder if all the results in Table 1 are of no-LOGO, and the authors need to clarify it.

3.       The Discussion is weak and I suggest the authors to make profound discussion about the results and the methodologies of the present study.

4.       Line 66 and Line 207: where the lapse rate of “9.88°C / 1000m” come from?

5.       Line 258: what’s the “mean interpolation”? please explain it.

6.       Line 256: please explain the “temporal data” at its first place.

7.       Line 317: 1.39 should be 1.29.

8.       Figure 6 and Figure 7 show the feature weights of the XGBoost method. For comparison, please show also that of the NN methods and make relevant description of the comparison.

9.       All the figures are blurred and hard to read, please make them more readable with high quality.

Author Response

Thanks for your good comments.  

Please see reply in attached pdf. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I would like to thank the authors for their careful revision as per the provided comments.I recommend acceptance. 

Reviewer 2 Report

I have no further suggestions.

Reviewer 4 Report

The revised ms responds to my concerns completely and I have no more critical comment to it.