A ResNet-Based Super-Resolution Approach for Constructing a High-Resolution Temperature Dataset from ERA5 Reanalysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors are requested to provide a set of contributions in the provided article in the introduction section such that readers understand the significance of the article.

It is unclear what the contributions of this paper are. It seems like the super resolution is applied to temperature data from ERA5 dataset. The authors are advised to showcase the uniqueness of the model. If the model is newly developed various ablation studies are required to show the effectiveness of the proposed model. In addition to this, residual networks are extensively used for super resolution purposes so the authors are advised to point out the new developments.

 

The authors are requested to provide details about the figure in the article. For example, figure 12 and figure 13 are not mentioned in the article. In addition to this, figure 12 and the rest of them need clarification. How are low resolution compared to true data or SR data? Are they not of different resolutions? Furthermore the figures provided in the article are of low resolution, the authors are advised to zoom in the section for better visibility.

 

The authors mention that a few studies were conducted where the authors had applied SR to temperature data. However, these comparisons are not provided. The authors are requested to provide comparisons quantitatively in separate tables and qualitatively by zooming in areas where there are differences with state of the art methods. Currently there is no comparison with the existing methods.

 

The authors are requested to provide experimental details i.e. how the models were trained, ablation studies regarding how different layers affect the SR process. Why only 8 residual blocks were used? How would increasing or decreasing the residual block affect the network and so on.

Comments on the Quality of English Language

There are a few formatting and grammar erros which need to be corrected.

Author Response

Reviewer2

Comments 1:

The authors are requested to provide a set of contributions in the provided article in the introduction section such that readers understand the significance of the article

Response 1:

Thank you for your valuable comments! We have added a paragraph in the Introduction to describe the work and contributions of the paper, to help readers understand the significance of our paper. (Line 103-111)

 

Comments 2:

It is unclear what the contributions of this paper are. It seems like the super resolution is applied to temperature data from ERA5 dataset. The authors are advised to showcase the uniqueness of the model. If the model is newly developed various ablation studies are required to show the effectiveness of the proposed model. In addition to this, residual networks are extensively used for super resolution purposes so the authors are advised to point out the new developments.

The authors are requested to provide experimental details i.e. how the models were trained, ablation studies regarding how different layers affect the SR process. Why only 8 residual blocks were used? How would increasing or decreasing the residual block affect the network and so on.

Response 2:

Thank you for your suggestions! In Section 4, we have introduced Section 4.3 Discussion, in which we present some of the ablation study results listed in Table 2. We have also refined the experimental data by comparing a CNN model without residual blocks with the ResNet model. The significance of incorporating residual blocks is also discussed.

 

Comments 3:

The authors are requested to provide details about the figure in the article. For example, figure 12 and figure 13 are not mentioned in the article. In addition to this, figure 12 and the rest of them need clarification. How are low resolution compared to true data or SR data? Are they not of different resolutions? Furthermore the figures provided in the article are of low resolution, the authors are advised to zoom in the section for better visibility.

Response 3:

Thanks! Figure 12 and figure 13 are mentioned in line 335 and line 353, respectively. For the figures in the paper, we have replaced them with clearer versions. However, due to the inherently low resolution of the data, some figures may still appear slightly blurry.

 

Comments 4:

The authors mention that a few studies were conducted where the authors had applied SR to temperature data. However, these comparisons are not provided. The authors are requested to provide comparisons quantitatively in separate tables and qualitatively by zooming in areas where there are differences with state of the art methods. Currently there is no comparison with the existing methods.

Response 4:

Thank you for your suggestions! In Section 4, we have added 4.3 Discussion, where we compare the results listed in Table 2 with some existing methods. Compared with existing studies, our method has achieved a certain improvement in performance.

 

Comments 5:

There are a few formatting and grammar erros which need to be corrected.

Response 5:

Thank you for your suggestions! We have revised some formatting issues and grammatical errors.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Overall evaluation:

(1) Main contribution: The authors perform SR reconstruction on temperature data using convolutional neural networks to enhance effective resolution and data information (ERA5 data: 1 year of data, spatial resolution of 0.25° × 0.25°, 2-meter temperature parameter). They used a model that combines sub-pixel convolution with ResNet architecture to perform super-resolution reconstruction of temperature data in the SC region.

(2) Defects of the paper:

(a) The model is based only on temperature data. One parameter is not enough to conclude anything. Your model must incorporate data about relative humidity, pressure, wind direction, and speed.

(b) This study's limitations are not well emphasised. The same idea applies to the novelty elements.

(c) A short description of the studied area and the climate from there is missing.

(d) Conclusions are not clear enough at this moment. I suggest using numbers for objectivity.

(3) Minor mistakes:

Eq. (4) and (5) are a bit outside the body text.

(4) Question:

How did you divide the dataset into training and validating datasets?

(5) Innovation elements:

This study shows that using CNN for the super-resolution reconstruction of temperature data is feasible. Statistical outcomes derived from the evaluation coefficients and the analysis of the comparison charts post-super-resolution underline the model's capabilities.

(6) Overall recommendation: With the identified issues addressed, the paper has the potential to be accepted after minor revisions. The authors are encouraged to consider the feedback provided and make the necessary adjustments to improve their paper’s clarity, significance, and objectivity.

Author Response

Comments 1:

The model is based only on temperature data. One parameter is not enough to conclude anything. Your model must incorporate data about relative humidity, pressure, wind direction, and speed.

Response 1:

Thank you for your suggestions! Current research has focused on temperature data, but future directions will consider incorporating additional parameters, such as wind direction.

 

Comments 2:

This study's limitations are not well emphasised. The same idea applies to the novelty elements.

Response 2:

Thanks! Super-resolution studies of meteorological data share similar limitations: concerns over the generalizability of methods. Specifically, whether a model trained on data from one region can be used in different region. In the future, we will also investigate the generalizability of the model. We will conduct research on the generalizability of models, exploring their performance on temperature data across different regions and investigating the factors that influence a model's transfer learning capabilities.

 

Comments 3:

A short description of the studied area and the climate from there is missing.

Response 3:

Thanks! We have added a brief description of the regional climate in Section 2. Data. The description covers the general characteristics of the regional climate, as well as the average temperature, and the highest and lowest temperature conditions.

 

Comments 4:

Conclusions are not clear enough at this moment. I suggest using numbers for objectivity.

Response 4:

Thanks! In Section 4, subsections 4.2 and 4.3 have been enhanced with additional data tables that compare the performance of our model with those from related studies.

 

Comments 5:

Eq. (4) and (5) are a bit outside the body text.

Response 5:

Thanks! Eq. (4) and (5) have been revised to ensure standardization.

 

Comments 6:

How did you divide the dataset into training and validating datasets?

Response 6:

Thank you for the question! The temperature data spanning from January 1, 2023, to December 31, 2023, were designated as the training set, while the dataset covering January 1, 2020, to December 31, 2020, was used as the validation set. (Line 132-134)

 

Comments 7:

This study shows that using CNN for the super-resolution reconstruction of temperature data is feasible. Statistical outcomes derived from the evaluation coefficients and the analysis of the comparison charts post-super-resolution underline the model's capabilities.

Overall recommendation: With the identified issues addressed, the paper has the potential to be accepted after minor revisions. The authors are encouraged to consider the feedback provided and make the necessary adjustments to improve their paper’s clarity, significance, and objectivity.

Response 7:

Thank you very much.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors mention that one of the contributions is "The network improves the effective resolution and detail of the data while requiring less time and computational resources compared to conventional methods. ", however, there is no comparison to show that the proposed method is faster than the rest of the methods. 

In addition to this, there is no comparison to show " The model performs better with different sizes of LR data, which enhances its generalizability." The authors need to provide tables that contain quantitative and qualitative results with different input sizes to showcase this.

The authors are requested to revise the ablation study. Generally, an ablation study analyzes the contribution of different components of a model by systematically removing or modifying them and evaluating the impact on performance. This helps in understanding the significance of each component and justifying its inclusion. In this case, the number of residual blocks can be of one of them. 

The authors are requested to add additional metrics such as PSNR, SSIM as most of the recent methods utilize this to showcase the performance of the provided techniques.

 

The authors are requested to provide additional images to showcase qualitative analysis.

 

The authors are requested to provide information about Table 1. Is this table an average across the entire test dataset or is it based on one image? If it is on one image, an average of all the images in a dataset is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have considered the recommendations, adding information where necessary and improving the quality of some of the figures for the reader's sake.

The recommendation is for publishing.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

The authors are requested to provide citations for statements like this -  "When the numerical model is used to obtain an output with twice the original spatial resolution, the computational cost will increase by a factor of 8 (as the resolution doubles in longitude, latitude, and time)."

The authors mention that the computational cost is reduced by 1/8. Where are these validated? There are no performance comparison charts provided to illustrate the same.

The authors are requested to clarify what average moments mean?

The authors are requested to cross verify the metrics in table 3. While the PSNR is inverse of RMSE, the proposed method which has RMSE 0.289 has PSNR of 58.8929, while dADR-SR method has RMSE 0.310 which is slightly higher than 0.289 but the PSNR is 37.9?  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 4

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed majority of the comments; however, a clarification is required.

Are all results provided in Table 3 for the same dataset?
While browsing through the papers it seemed like most of the results for the provided state-of-the-art were from different dataset. For example, Vandal [19] proposed DeepSD and used PRISM dataset for precipitation and GTOPO30 elevation (30 arcsec spatial resolution) dataset distributed by the Land Processes Distributed Active Archive Center (LP DAAC) for topography. However, the proposed paper uses a different dataset namely, the European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis. The authors are advised to provide comparison for the same dataset since it is unfair and invalid if the test dataset is completely different for different methods.


In addition to this RMSE - 2.949 can be found for SRCNN in the Vandal et al., paper while DeepSD produces - 2.529. The authors are requested to clarify and correct this if it is a mistake? Could the authors provide an explanation on how PSNR value was generated since Vandal et al., does not provide that metric?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 5

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed the comments promptly