Spatial Downscaling of ERA5 Reanalysis Air Temperature Data Based on Stacking Ensemble Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

First of all, the title of this paper does not match with the system. Please check and revise the correct title.

 

Figure 1 shows the map of the study area and meteorological stations distribution which is very important.

 

The source of data used in this study is clearly stated in section 2.

 

The evaluation metric of this study is also clearly explained and suitable for this type of study.

 

The hyperparameter of the machine learning models are not mentioned in the text. Can the author explain why these parameters is not included?

 

It is very important to include them in order to improve the quality of the paper.

 

Figure 6 should use be written in English to align with the rest of the text.

 

Can the authors explain why is there an overlap between RMSE and Relevance in Figure 7? It does not seem to be clearly explained.

 

Can the author explain if any form of feature selection was used in this study?

 

The authors did mention the limitation of this study in section 4 but did not further explained how to improve this study in the conclusion.  

Comments on the Quality of English Language

The paper is being written quite loosely. Please consider to write in a more formal manner and also check the spelling and grammar throughout the text.

Author Response

1. Summary

Thank you very much for taking the time to review the manuscript. We have made revisions to the manuscript based on your suggestions, and the modified sections have been highlighted in red font in the revised manuscript. Detailed responses to your comments have been provided in Section 2. Once again, we appreciate your valuable feedback on the manuscript.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: First of all, the title of this paper does not match with the system. Please check and revise the correct title.

Response 1: Thank you very much for your suggestion. We agree with it and have already modified the title. The modified title is consistent with the one submitted by the system.

Comments 2: The hyperparameter of the machine learning models are not mentioned in the text. Can the author explain why these parameters is not included? It is very important to include them in order to improve the quality of the paper.

Response 2: Thank you very much for your suggestion. We agree with it. Stacking ensemble learning is used to integrate three machine learning methods in the article. Bayesian optimization is employed to optimize the model's hyperparameters in this study. Due to the simplification of sections, it was not described in the Methods section. In the revised manuscript, we describe the hyperparameter optimization method in Section 2.3.2 on page 6, line 235.

Comments 3: Figure 6 should use be written in English to align with the rest of the text.

Response 3: Thank you very much for this suggestion. We agree with it, and we have already made modifications to Figure 6.

Comments 4: Can the authors explain why is there an overlap between RMSE and Relevance in Figure 7? It does not seem to be clearly explained.

Response 4: Thank you for raising this question, and we will provide an explanation. Figure 7 is intended to illustrate the correlation between the RMSE of downscaled results and the quality of the reanalysis data with observed values. The higher the correlation, the smaller the RMSE. The correlation values range from 0 to 1, while the RMSE values range from 0 to 8. Therefore, as the values vary, intersections may occur in certain areas.

Comments 5: Can the author explain if any form of feature selection was used in this study?

Response 5: Thank you very much for raising this question. This study utilized selected independent variables and conducted correlation analysis with 0.1° ERA5-Land reanalysis temperature data to complete the selection of input variables for the model. The specific content has been added in Section 3.2, Line 299 of page nine.

Comments 6: The authors did mention the limitation of this study in section 4 but did not further explained how to improve this study in the conclusion.

Response 6: Thank you very much for raising this question, and we would like to provide an explanation. The limitation of this study lies in the validation of the downscaled results. Currently, the temperature measurements we can obtain are from meteorological stations, and the distribution of these stations is objective. Unfortunately, we do not have remote sensing data to validate this aspect, which is a common limitation. Therefore, there is a limitation in our ability to validate the downscaled results. However, utilizing measured temperature values from meteorological stations can largely demonstrate the feasibility of this research.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Fig. 1: There is another map inside the map representing the china border. It seems to be redundant. If so, you can remove it.

L 153: It seems you mean ERA5-Land. If so, please explicitly mention that. This data is more known as a 9 km resolution data, instead of 0.1 degree.

L 154: This address (apps.ecmwf.int/...) is no longer valid. The new data archive is at the Copernicus website.

L 167: Both are acquired at 10:30? Or perhaps the nighttime is 22:30?

L 172: Please give a reference for this software tool (MRT).

DEM data, ...: Please provide the information about the availability of the any input data and tools used in the study. If they are open source and available to public or not?

Fig. 2: Make another Figure/diagram, explicitly showing the features you used in the learning process. So the main features were the MODIS LST and DEM? If so, please clearly show them in a separate diagram or by modifying Fig. 2.
If MODIS data was a critical feature, it should be discussed in more details, with additional images. For example, the MODIS swath passing the domain of study. Is the MODIS footprint provides a uniform data over the region, or are there sporadic missing values due to clouds and other factors? How these missing values can be handled. Or is it okay to ignore them, ...?

L 252: one sentence (N represents ...) has been repeated.

L 260: Was it possible to apply the training process on hourly ERA5-Land data? Or could you test the algorithm on some hourly data and validate the results for some certain days, instead of months? I think if a daily validation shows any improvement, that can be considered as highly significant, in comparison to monthly validation. So the algorithm can be safely employed in (let's say) atmospheric operational systems and workflows.

Fig. 3: Captions of Figures should be 100% clear and complete. Correlation between air temperature of which data? The final product of the learning process? Different time scales should me mentioned.

Fig. 7: What is "Relevance" displayed by a blue line? There is no clear explanation about the meaning of "Relevance" in the manuscript. Or maybe you discussed about it without mentioning its name? Legends and captions must be clear.

4.1 Variable selection: I think this sub-section must be moved to methodology, before sub-section 3.3 which discusses about the results.

L 394-396: What do these terms (YNLST, ONLST, YLST, and NLST) stand for?
You used these abbreviations in Figs 8 and 9, but without any discussion about their nomenclature in the manuscript.

Figs 8 and 9: What was your criteria for the feature selection? What is the meaning of variable importance in Figs 7 and 8? Are they some kind of correlation? Please clarify.
How did you calculate the scores for each variable in Figs 7 and 8?

Author Response

1. Summary

Thank you very much for taking the time to review the manuscript. We have made revisions to the manuscript based on your suggestions, and the modified sections have been highlighted in red font in the revised manuscript. Detailed responses to your comments have been provided in Section 2. Once again, we appreciate your valuable feedback on the manuscript.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: Fig. 1: There is another map inside the map representing the china border. It seems to be redundant. If so, you can remove it.

Response 1: Thank you for pointing out this issue. Regarding this matter, I would like to provide an explanation. The inclusion of the Chinese border on the map was to illustrate the specific location of my research area within China. If you feel it is unnecessary, I will remove it in subsequent versions.

Comments 2: L 153: It seems you mean ERA5-Land. If so, please explicitly mention that. This data is more known as a 9 km resolution data, instead of 0.1 degree.

Response 2: Thank you very much for raising this question, we agree with it. ERA5-Land data is used in this paper, and we have mentioned it in line 150. Additionally, we found that the original spatial resolution of the ERA5-Land reanalysis dataset is 9 kilometers on a simplified Gaussian grid, and the data in CSD has been reset to a regular latitude-longitude grid of 0.1 x 0.1 degrees. The data downloaded for this paper is at a resolution of 0.1 degrees (approximately 10 kilometers).

Comments 3: L 154: This address (apps.ecmwf.int/...) is no longer valid. The new data archive is at the Copernicus website.

Response 3: Thank you very much for raising this question, we agree with it. Our data was downloaded from the European Centre for Medium-Range Weather Forecasts, and the download link has been updated in page 4, line 152.

Comments 4: Both are acquired at 10:30? Or perhaps the nighttime is 22:30?

Response 4: Thank you very much for raising this issue. We agree with the suggestion, and we have made the correction on page 5, line 167

Comments 5: L 172: Please give a reference for this software tool (MRT).

Response 5: Thank you very much for your input, and we agree with the suggestion. MRT is a tool provided by NASA for processing MODIS data, allowing for tasks such as mosaicking, reprojection, and format conversion. The download link for this tool is provided on page 5, line 72 of the document.

Comments 6: DEM data, ...: Please provide the information about the availability of the any input data and tools used in the study. If they are open source and available to public or not?

Response 6: Thank you very much for bringing up this issue. We agree that all the data used in the article are properly sourced, as mentioned in the text. Additionally, the main tools utilized in this research are Python, ArcGIS, and MRT. Python and ArcGIS are commonly used software tools, We did not provide download links in the article, the download address for MRT has been included in the revised manuscript.

Comments 7: Fig. 2: Make another Figure/diagram, explicitly showing the features you used in the learning process. So the main features were the MODIS LST and DEM? If so, please clearly show them in a separate diagram or by modifying Fig. 2.

Response 7: Thank you very much for your suggestions. We would like to provide an explanation. Figure 2 depicts the flowchart of Stacking ensemble learning. The first part comprises input data, which includes the training set and validation set consisting of all independent variables and dependent variables; hence, it is not possible to display LST and DEM separately. In addition to the base learners and meta-learner, the functionalities utilized include Bayesian optimization and 10-fold cross-validation, which are now shown in the new Figure 3 on page 6.

Comments 8: If MODIS data was a critical feature, it should be discussed in more details, with additional images. For example, the MODIS swath passing the domain of study. Is the MODIS footprint provides a uniform data over the region, or are there sporadic missing values due to clouds and other factors? How these missing values can be handled. Or is it okay to ignore them, ...?

Response 8: Thank you very much for bringing up this issue. Regarding the MODIS data, the data used in this article are all at the monthly scale. The 8-day LST (Land Surface Temperature) data mentioned in the article were downloaded, and we employed an averaging method to synthesize the eight-day data into monthly data. Upon inspecting the synthesized data, we found sporadic missing values in some months. To address this, we used ArcGIS to fill in the missing values. Additionally, MODIS NDVI (Normalized Difference Vegetation Index) data itself is monthly data, and only a small amount of data had sporadic missing values. We also used ArcGIS to fill in these missing values. Taking January 2003 LST data as an example, in Figure 2 on page five, we present the distribution of Aqua satellite surface land temperature data before and after filling in the gaps.

Comments 9: L 252: one sentence (N represents ...) has been repeated.

Response9: Thank you for your feedback. We accept it and have removed the mentioned content in the new draft.

Comments 10: L 260: Was it possible to apply the training process on hourly ERA5-Land data? Or could you test the algorithm on some hourly data and validate the results for some certain days, instead of months? I think if a daily validation shows any improvement, that can be considered as highly significant, in comparison to monthly validation. So the algorithm can be safely employed in (let's say) atmospheric operational systems and workflows.

Response 10: Thank you very much for raising this issue. We would like to provide an explanation. Due to the lack of hourly time-series data for MODIS land surface temperature and NDVI, our training process could not be applied to downscale the ERA5-Land data at hourly resolution. Daily MODIS data, due to factors such as cloud cover, contains too many missing values, making it impractical for precise imputation. Consequently, we were unable to conduct daily ERA5-Land downscaling training and validation. If suitable daily or hourly land surface temperature data and NDVI data become available in the future, subsequent research could consider increasing the temporal resolution to daily or hourly scales.

Comments 11: Fig. 3: Captions of Figures should be 100% clear and complete. Correlation between air temperature of which data? The final product of the learning process? Different time scales should me mentioned.

Response 11: Thank you very much for bringing up this issue. We agree with the suggestion, and the modified title has been updated on page 8, line 297. The figure illustrates the credibility of monthly ERA5-Land data in the Yellow River basin. It depicts the correlation between monthly average ERA5-Land data and observed data from meteorological stations. The higher the correlation, the higher the credibility.

Comments 12: Fig. 7: What is "Relevance" displayed by a blue line? There is no clear explanation about the meaning of "Relevance" in the manuscript. Or maybe you discussed about it without mentioning its name? Legends and captions must be clear.

Response 12: Thank you very much for raising this issue. We agree with the suggestion. This relevance refers to the correlation coefficient between the measured temperature data at meteorological stations and the downscaled results at the corresponding locations of the stations. The formula for calculating the correlation coefficient has been introduced in Section 2.3.5 on page 7. Figure 7 in the original manuscript has been changed to Figure 9 in the revised manuscript, and modifications have been made to the axis titles and legend in the figure.

Comments 13: 4.1 Variable selection: I think this sub-section must be moved to methodology, before sub-section 3.3 which discusses about the results.

Response 13: Thank you very much for suggesting this change. We agree and have made the necessary modifications. Regarding section 4.1, there was a naming issue which we have rectified. This section pertains to the analysis of feature importance. We have made the necessary adjustments in the revised manuscript. As for the variable selection section, we have reintroduced it in the new revised manuscript, as detailed in Section 3.2 on page 9. In this section, we analyzed the correlation between various variables and ERA5-Land in the analysis of temperature data, thereby completing the variable selection process.

Comments 14: L 394-396: What do these terms (YNLST, ONLST, YLST, and NLST) stand for? You used these abbreviations in Figs 8 and 9, but without any discussion about their nomenclature in the manuscript.

Response 14: Thank you very much for your suggestion. We agree with it and have made the necessary modifications. (YNLST, ONLST, YLST, and NLST) are the names of the independent variables, and we have provided explanations for them in Table 2 on page nine.

Comments 15: Figs 8 and 9: What was your criteria for the feature selection? What is the meaning of variable importance in Figs 7 and 8? Are they some kind of correlation? Please clarify.

Response 15: Thank you very much for this suggestion. We will provide an explanation accordingly. Regarding the variable selection in our paper, it was not previously mentioned. We selected the features for model training based on their correlation with ERA5-Land in the analysis of air temperature data. We have now included this information in the article, specifically in Section 3.2 on page 9. Additionally, variable importance refers to a metric used in data analysis, modeling, or machine learning tasks to measure the impact of each variable on the model output. In machine learning, variable importance is typically employed to determine which variables are most crucial for explaining or predicting the target variable. In this context, we aim to use variable importance to assist in analyzing the suitability of the variables chosen in our study.

Comments 16: How did you calculate the scores for each variable in Figs 7 and 8?

Response 16: Thank you very much for bringing up this question. We will provide an explanation accordingly. The models chosen as base learners in the document are all tree-based models. Calculating the feature importance of tree models involves measuring the information gain of each feature during node splits across all decision trees. This process is implemented using Python code.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have made overall improvements to the manuscript taking into consideration of the comments of the reviewer. Thus, I have no further objection for this paper to be published. 

Comments on the Quality of English Language

Please double check the grammar and spelling throughout the text before final publication. 

Author Response

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The authors have made overall improvements to the manuscript taking into consideration of the comments of the reviewer. Thus, I have no further objection for this paper to be published.

Response 1: Thank you once again for taking the time to review the revised manuscript. We sincerely appreciate your acknowledgment of the changes made to this manuscript.

Comments 2: Please double check the grammar and spelling throughout the text before final publication.

Response 2: Thank you for your suggestions. We are grateful for them. We have already checked the grammar and spelling of the entire text and made corrections where necessary. If there are still issues later on, we will make further adjustments.

Author Response File: Author Response.pdf