Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network

Fu, He; Guo, Jianing; Deng, Chenguang; Liu, Heng; Wu, Jie; Shi, Zhengguo; Wang, Cailing; Xie, Xiaoning

doi:https://doi.org/10.5194/gmd-2023-158

Preprints

https://doi.org/10.5194/gmd-2023-158

Preprints

Submitted as: development and technical paper

15 Aug 2023

Submitted as: development and technical paper |

| 15 Aug 2023

Status: this preprint has been withdrawn by the authors.

Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network

He Fu, Jianing Guo, Chenguang Deng, Heng Liu, Jie Wu, Zhengguo Shi, Cailing Wang, and Xiaoning Xie

Abstract. The middle reaches of the Yellow River (MRYR), located in northern China, are the transition zone between semi-arid and semi-humid climates. As one of the climate-sensitive regions in China, MRYR has fragile ecological environment and serious soil loss, which leads to geological disasters such as landslides, collapses and mudslides caused by extreme precipitation to occur. However, scarceness of the high-resolution precipitation data over MRYR limits the assessment of the environmental impacts caused by climate change, especially for extreme precipitation. In this paper, we design a Residual in Residual Dense Block based network (RRDBNet) model for the statistical downscaling of precipitation in MRYR, and compare the proposed RRDBNet with the generalized linear regression model and two popular deep learning-based models. The results show that the proposed RRDBNet model has a good performance on precipitation simulations, which can well reproduce the spatial-temporal characteristics of high-resolution precipitation. Especially, RRDBNet has substantial improvements in extreme precipitation compared with other models. On the probability density function of daily precipitation, it is further demonstrated that RRDBNet performs better on extreme precipitation frequency. Our results suggest that the statistical downscaling based on RRDBNet may be an effective tool for historical and future climate simulations from global climate models.

This preprint has been withdrawn.

Received: 19 Jul 2023 – Discussion started: 15 Aug 2023

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this preprint. The responsibility to include appropriate place names lies with the authors.

Download & links

Preprint (PDF, 15368 KB)

Withdrawal notice
This preprint has been withdrawn.
Preprint (15368 KB)

Download & links

This preprint has been withdrawn.

He Fu, Jianing Guo, Chenguang Deng, Heng Liu, Jie Wu, Zhengguo Shi, Cailing Wang, and Xiaoning Xie

Interactive discussion

Status: closed

RC1:
'Comment on gmd-2023-158', Anonymous Referee #1, 24 Aug 2023
Review for “Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network” By Fu et al.

In this study, Fu et al. proposed a new deep learning approach, namely, Residual in Residual Dense Block based network (RRDBNet), to downscale ERA5 data at the middle reaches of the Yellow River (MRYR) in China. The proposed method was compared with three other methods (GLM, CNN and RDBNet). The authors concluded that RRDBNet performed better than other methods in terms of difference, RMSE and correlartion coefficient (CC). However, the study lacks novelty and improvements are very trivial. Thus, I suggest rejecting for publishing it at GMD. My major concerns are as follows:

The models were trained with daily data, but all the evaluations were performed at aggregated time scale (annual, monthly and seasons). How does the model perform in daily time scale (both daily statistics and extremes)? The aggregated time scales may not be critical and may hide important information on model evaluation.

The rationale of the proposed method is not clear. Compared to other deep learning methods, what are its advantages and why did the author propose this method? Without deep understanding the model itself, the manuscript gives readers limited insights.

The authors claimed the proposed method is much better than the other three methods, which may not be true. Given the stochastic nature of deep learning models and the slightly better statistics, the superiority may be purely due to stochasticity itself. Training the model multiple times may help discriminate the superiority and stochasticity. Furthermore, it is not fair to compare different deep learning models without giving model complexity (e.g., the number of trainable parameters).

Minor comments

In the introduction section, the authors described many GCM downscaling. However, this study is not about GCM downscaling but reanalysis data, which is very different story and may mislead readers. Thus, the introduction needs to be rewritten.

Line 69: the reanalysis data ERA5 has spatial resolution of 0.25x0.25 degree not 2x2 degree. Furthermore, the authors selected 5 predictors without giving any reasons.

Line 89: the three parameters came out first time without any explanations. Y ∈ Rt°x100°x159 came out without further information.

Lines 114 to 119: Does the statement about batch normalization come from model testing? If that is true, this statement needs to be included in the results section. If it is not true, where this come from?

Line 121: how did you get B=0.2?

Line 139: The statement “The final precipitation is obtained by multiplying p with the random values of the distribution with shape α scale β.” Why?

Line 142: “In the training phase” was repeated.
Citation: https://doi.org/10.5194/gmd-2023-158-RC1
- AC2: 'Reply on RC1', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Referee #1,
  We sincerely thank the Referee #1 for all your valuable comments and insightful suggestions to our manuscript to improve the manuscript. We have addressed all the specific comments in the revised manuscript. Since the responses are long and contain some figures and tables, we put the point-by-point responses in "Response_Referee#1.pdf". Please see "Response_Referee#1.pdf".
  
  Best regards,
  Xiaoning Xie et al.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC2
CEC1:
'Comment on gmd-2023-158', Juan Antonio Añel, 05 Sep 2023

Dear authors,
After checking your manuscript, a couple of issues regarding the code availability need to be fixed. They are related to the GitHub repositories that you mention in your statement. GitHub is not a suitable repository for scientific publication archival. Fortunately, in the case of your manuscript, you can solve this quickly. First, one of the GitHub repositories you mentioned already has a copy in Zenodo, which is linked in GitHub; it is https://zenodo.org/record/5007540. The other one corresponds to downscaleR.keras, and this code is under the GPLv3 license, so you can take it and create a new Zenodo repository containing it. This is what you should do.

Therefore, please modify the Code availability statement in your manuscript in any future reviewed version, changing the GitHub links for the Zenodo repositories and their DOIs. Also, please reply to this comment with the details for the Zenodo repository for downscaleR.keras, as this should have been addressed already at the submission of your manuscript.
Juan A. Añel
Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/gmd-2023-158-CEC1
- AC1: 'Reply on CEC1', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Executive Editor,
  We sincerely thank you for all your valuable comments, insightful suggestions, and thoughtful corrections to our manuscript. We have addressed all the specific comments in the revised manuscript, with the point-by-point responses detailed below.
  
  Best regards,
  Xiaoning Xie et al.
  
  “ After checking your manuscript, a couple of issues regarding the code availability need to be fixed. They are related to the GitHub repositories that you mention in your statement. GitHub is not a suitable repository for scientific publication archival. Fortunately, in the case of your manuscript, you can solve this quickly. First, one of the GitHub repositories you mentioned already has a copy in Zenodo, which is linked in GitHub; it is https://zenodo.org/record/5007540. The other one corresponds to downscaleR.keras, and this code is under the GPLv3 license, so you can take it and create a new Zenodo repository containing it. This is what you should do.
  Therefore, please modify the Code availability statement in your manuscript in any future reviewed version, changing the GitHub links for the Zenodo repositories and their DOIs. Also, please reply to this comment with the details for the Zenodo repository for downscaleR.keras, as this should have been addressed already at the submission of your manuscript.”
  Response: Thanks for your comments and suggestions. Based on your suggestions, we have made modifications to the code and data availability section. The modifications are as follows: All models are based on R software and build on the Climate4R framework (Iturbide et al., 2019). The code for the GLM and CNN models is obtained from https://zenodo.org/record/5007540 (Baño-Medina et al., 2020). To design the RDBNet and RRDBNet used, we rely on downscaleR.keras (https://doi.org/10.5281/zenodo.10077173). The code for the RDBNet and RRDBNet models is available at https://doi.org/10.5281/zenodo.8234006. The input data from ERA5 (https://doi.org/10.24381/cds.bd0915c6; Hersbach et al., 2023) underlying this study are publicly available. The observational data from GPM (http://gpm.nasa.gov/data-access/downloads/gpm; Huffman et al., 2015) underlying this study are publicly available. The output data can be found at https://doi.org/10.5281/zenodo.8234006.
  
  References
  Baño-Medina, J., Manzanas, R., and Gutiérrez, J. M.: Configuration and intercomparison of deep learning neural models for statistical downscaling, Geosci. Model Dev., 13, 2109–2124, 2020.
  Hersbach, H., Bell, B., Berrisford, P., Biavati, G., Horányi, A., Sabater, J. M., Nicolas, J., Peubey, C., Radu, R., Rozum, I., et al.: ERA5 hourly data on pressure levels from 1940 to present, 2023.
  Huffman, G. J., Bolvin, D. T., Braithwaite, D., Hsu, K., Joyce, R., Xie, P., and Yoo, S.-H.: NASA Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG), Algorithm theoretical basis document (ATBD) version, 4, 2015.
  Iturbide, M., Bedia, J., Herrera, S., Baño-Medina, J., Fernández, J., Frías, M. D., Manzanas, R., San-Martín, D., Cimadevilla, E., Cofiño, A. S., et al.: The R-based climate4R open framework for reproducible climate data access and post-processing, Environ. Model. Softw., 111, 42–54, 2019.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC1
RC2:
'Comment on gmd-2023-158', Anonymous Referee #2, 26 Sep 2023

The authors have used a residual-based approach to downscale precipitation over the reaches of China. The residual-based approach significantly improves on traditional machine learning approaches when compared to observational data.
While many aspects of the research are useful in this manuscript, there is little novelty in this approach, and many aspects of the research have been widely investigated elsewhere. The authors have many redundant figures, and the manuscript could be condensed significantly. Additionally, you only use 4 years of data to evaluate extremes. I suggest doing another experiment with a larger test set.

Citation: https://doi.org/10.5194/gmd-2023-158-RC2
- AC3: 'Reply on RC2', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Referee #2,
  We sincerely thank you for all your valuable comments, insightful suggestions, and thoughtful corrections to our manuscript. These comments and suggestions will undoubtedly help us improve the quality of the manuscript. In the revised manuscript, all changes and additions are highlighted in yellow. Since the responses are long and contain some figures and tables, we put the point-by-point responses in "Response_Referee#2.pdf". Please see "Response_Referee#2.pdf".
  
  Best regards,
  Xiaoning Xie et al.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC3
EC1: 'Comment on gmd-2023-158', Di Tian, 07 Nov 2023

Dear Authors,
Thanks for submitting your point-to-point response response to the reviewers' comments. As I read the manuscript as well as the response to the reviewers' comments, I agree with both reviewers that the study lacks novelty and that many aspects of the research have been investigated in many recent studies which are mostly neglected in the manuscript. The response to the reviewers' comments does not provide sufficient evidence to justify its novelty. Given that this is a fundamental issue, it requires redesigning methods, experiments, or rewriting the manuscript, which is not likely to be addressed in a limited time. Therefore, my recommendation is to reject and discourage the submission of a revised manuscript.

Citation: https://doi.org/10.5194/gmd-2023-158-EC1

Interactive discussion

Status: closed

RC1:
'Comment on gmd-2023-158', Anonymous Referee #1, 24 Aug 2023
Review for “Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network” By Fu et al.

In this study, Fu et al. proposed a new deep learning approach, namely, Residual in Residual Dense Block based network (RRDBNet), to downscale ERA5 data at the middle reaches of the Yellow River (MRYR) in China. The proposed method was compared with three other methods (GLM, CNN and RDBNet). The authors concluded that RRDBNet performed better than other methods in terms of difference, RMSE and correlartion coefficient (CC). However, the study lacks novelty and improvements are very trivial. Thus, I suggest rejecting for publishing it at GMD. My major concerns are as follows:

The models were trained with daily data, but all the evaluations were performed at aggregated time scale (annual, monthly and seasons). How does the model perform in daily time scale (both daily statistics and extremes)? The aggregated time scales may not be critical and may hide important information on model evaluation.

The rationale of the proposed method is not clear. Compared to other deep learning methods, what are its advantages and why did the author propose this method? Without deep understanding the model itself, the manuscript gives readers limited insights.

The authors claimed the proposed method is much better than the other three methods, which may not be true. Given the stochastic nature of deep learning models and the slightly better statistics, the superiority may be purely due to stochasticity itself. Training the model multiple times may help discriminate the superiority and stochasticity. Furthermore, it is not fair to compare different deep learning models without giving model complexity (e.g., the number of trainable parameters).

Minor comments

In the introduction section, the authors described many GCM downscaling. However, this study is not about GCM downscaling but reanalysis data, which is very different story and may mislead readers. Thus, the introduction needs to be rewritten.

Line 69: the reanalysis data ERA5 has spatial resolution of 0.25x0.25 degree not 2x2 degree. Furthermore, the authors selected 5 predictors without giving any reasons.

Line 89: the three parameters came out first time without any explanations. Y ∈ Rt°x100°x159 came out without further information.

Lines 114 to 119: Does the statement about batch normalization come from model testing? If that is true, this statement needs to be included in the results section. If it is not true, where this come from?

Line 121: how did you get B=0.2?

Line 139: The statement “The final precipitation is obtained by multiplying p with the random values of the distribution with shape α scale β.” Why?

Line 142: “In the training phase” was repeated.
Citation: https://doi.org/10.5194/gmd-2023-158-RC1
- AC2: 'Reply on RC1', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Referee #1,
  We sincerely thank the Referee #1 for all your valuable comments and insightful suggestions to our manuscript to improve the manuscript. We have addressed all the specific comments in the revised manuscript. Since the responses are long and contain some figures and tables, we put the point-by-point responses in "Response_Referee#1.pdf". Please see "Response_Referee#1.pdf".
  
  Best regards,
  Xiaoning Xie et al.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC2
CEC1:
'Comment on gmd-2023-158', Juan Antonio Añel, 05 Sep 2023

Dear authors,
After checking your manuscript, a couple of issues regarding the code availability need to be fixed. They are related to the GitHub repositories that you mention in your statement. GitHub is not a suitable repository for scientific publication archival. Fortunately, in the case of your manuscript, you can solve this quickly. First, one of the GitHub repositories you mentioned already has a copy in Zenodo, which is linked in GitHub; it is https://zenodo.org/record/5007540. The other one corresponds to downscaleR.keras, and this code is under the GPLv3 license, so you can take it and create a new Zenodo repository containing it. This is what you should do.

Therefore, please modify the Code availability statement in your manuscript in any future reviewed version, changing the GitHub links for the Zenodo repositories and their DOIs. Also, please reply to this comment with the details for the Zenodo repository for downscaleR.keras, as this should have been addressed already at the submission of your manuscript.
Juan A. Añel
Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/gmd-2023-158-CEC1
- AC1: 'Reply on CEC1', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Executive Editor,
  We sincerely thank you for all your valuable comments, insightful suggestions, and thoughtful corrections to our manuscript. We have addressed all the specific comments in the revised manuscript, with the point-by-point responses detailed below.
  
  Best regards,
  Xiaoning Xie et al.
  
  “ After checking your manuscript, a couple of issues regarding the code availability need to be fixed. They are related to the GitHub repositories that you mention in your statement. GitHub is not a suitable repository for scientific publication archival. Fortunately, in the case of your manuscript, you can solve this quickly. First, one of the GitHub repositories you mentioned already has a copy in Zenodo, which is linked in GitHub; it is https://zenodo.org/record/5007540. The other one corresponds to downscaleR.keras, and this code is under the GPLv3 license, so you can take it and create a new Zenodo repository containing it. This is what you should do.
  Therefore, please modify the Code availability statement in your manuscript in any future reviewed version, changing the GitHub links for the Zenodo repositories and their DOIs. Also, please reply to this comment with the details for the Zenodo repository for downscaleR.keras, as this should have been addressed already at the submission of your manuscript.”
  Response: Thanks for your comments and suggestions. Based on your suggestions, we have made modifications to the code and data availability section. The modifications are as follows: All models are based on R software and build on the Climate4R framework (Iturbide et al., 2019). The code for the GLM and CNN models is obtained from https://zenodo.org/record/5007540 (Baño-Medina et al., 2020). To design the RDBNet and RRDBNet used, we rely on downscaleR.keras (https://doi.org/10.5281/zenodo.10077173). The code for the RDBNet and RRDBNet models is available at https://doi.org/10.5281/zenodo.8234006. The input data from ERA5 (https://doi.org/10.24381/cds.bd0915c6; Hersbach et al., 2023) underlying this study are publicly available. The observational data from GPM (http://gpm.nasa.gov/data-access/downloads/gpm; Huffman et al., 2015) underlying this study are publicly available. The output data can be found at https://doi.org/10.5281/zenodo.8234006.
  
  References
  Baño-Medina, J., Manzanas, R., and Gutiérrez, J. M.: Configuration and intercomparison of deep learning neural models for statistical downscaling, Geosci. Model Dev., 13, 2109–2124, 2020.
  Hersbach, H., Bell, B., Berrisford, P., Biavati, G., Horányi, A., Sabater, J. M., Nicolas, J., Peubey, C., Radu, R., Rozum, I., et al.: ERA5 hourly data on pressure levels from 1940 to present, 2023.
  Huffman, G. J., Bolvin, D. T., Braithwaite, D., Hsu, K., Joyce, R., Xie, P., and Yoo, S.-H.: NASA Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG), Algorithm theoretical basis document (ATBD) version, 4, 2015.
  Iturbide, M., Bedia, J., Herrera, S., Baño-Medina, J., Fernández, J., Frías, M. D., Manzanas, R., San-Martín, D., Cimadevilla, E., Cofiño, A. S., et al.: The R-based climate4R open framework for reproducible climate data access and post-processing, Environ. Model. Softw., 111, 42–54, 2019.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC1
RC2:
'Comment on gmd-2023-158', Anonymous Referee #2, 26 Sep 2023

The authors have used a residual-based approach to downscale precipitation over the reaches of China. The residual-based approach significantly improves on traditional machine learning approaches when compared to observational data.
While many aspects of the research are useful in this manuscript, there is little novelty in this approach, and many aspects of the research have been widely investigated elsewhere. The authors have many redundant figures, and the manuscript could be condensed significantly. Additionally, you only use 4 years of data to evaluate extremes. I suggest doing another experiment with a larger test set.

Citation: https://doi.org/10.5194/gmd-2023-158-RC2
- AC3: 'Reply on RC2', Xiaoning Xie, 07 Nov 2023
  
  Original Manuscript ID: GMD-2023-158
  Original Article Title: Deep-learning statistical downscaling of precipitation in the middle reaches of the Yellow River: A Residual in Residual Dense Block based network
  
  Dear Referee #2,
  We sincerely thank you for all your valuable comments, insightful suggestions, and thoughtful corrections to our manuscript. These comments and suggestions will undoubtedly help us improve the quality of the manuscript. In the revised manuscript, all changes and additions are highlighted in yellow. Since the responses are long and contain some figures and tables, we put the point-by-point responses in "Response_Referee#2.pdf". Please see "Response_Referee#2.pdf".
  
  Best regards,
  Xiaoning Xie et al.
  
  Citation: https://doi.org/10.5194/gmd-2023-158-AC3
EC1: 'Comment on gmd-2023-158', Di Tian, 07 Nov 2023

Dear Authors,
Thanks for submitting your point-to-point response response to the reviewers' comments. As I read the manuscript as well as the response to the reviewers' comments, I agree with both reviewers that the study lacks novelty and that many aspects of the research have been investigated in many recent studies which are mostly neglected in the manuscript. The response to the reviewers' comments does not provide sufficient evidence to justify its novelty. Given that this is a fundamental issue, it requires redesigning methods, experiments, or rewriting the manuscript, which is not likely to be addressed in a limited time. Therefore, my recommendation is to reject and discourage the submission of a revised manuscript.

Citation: https://doi.org/10.5194/gmd-2023-158-EC1

He Fu, Jianing Guo, Chenguang Deng, Heng Liu, Jie Wu, Zhengguo Shi, Cailing Wang, and Xiaoning Xie

Viewed

Total article views: 806 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
581	183	42	806	29	34

HTML: 581
PDF: 183
XML: 42
Total: 806
BibTeX: 29
EndNote: 34

Views and downloads (calculated since 15 Aug 2023)

Month	HTML	PDF	XML	Total
Aug 2023	156	41	5	202
Sep 2023	84	21	5	110
Oct 2023	38	9	0	47
Nov 2023	69	18	10	97
Dec 2023	17	8	1	26
Jan 2024	12	8	1	21
Feb 2024	18	10	4	32
Mar 2024	22	22	5	49
Apr 2024	21	10	5	36
May 2024	17	7	2	26
Jun 2024	13	5	1	19
Jul 2024	19	7	1	27
Aug 2024	22	5	0	27
Sep 2024	13	4	0	17
Oct 2024	13	1	0	14
Nov 2024	37	6	2	45
Dec 2024	10	1	0	11

Cumulative views and downloads (calculated since 15 Aug 2023)

Month	HTML	PDF	XML	Total
Aug 2023	156	41	5	202
Sep 2023	84	21	5	110
Oct 2023	38	9	0	47
Nov 2023	69	18	10	97
Dec 2023	17	8	1	26
Jan 2024	12	8	1	21
Feb 2024	18	10	4	32
Mar 2024	22	22	5	49
Apr 2024	21	10	5	36
May 2024	17	7	2	26
Jun 2024	13	5	1	19
Jul 2024	19	7	1	27
Aug 2024	22	5	0	27
Sep 2024	13	4	0	17
Oct 2024	13	1	0	14
Nov 2024	37	6	2	45
Dec 2024	10	1	0	11

Viewed (geographical distribution)

Total article views: 779 (including HTML, PDF, and XML) Thereof 779 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 13 Dec 2024

Short summary

A Residual in Residual Dense Block based network model (RRDBNet) is designed for statistical downscaling of precipitation in the middle reaches of the Yellow River. RRDBNet has a good performance on precipitation simulations, well reproducing the spatial-temporal characteristics of high-resolution precipitation. RRDBNet has substantial improvements in extreme precipitation compared with generalized linear regression model and two deep learning-based models.


Total:	0
HTML:	0
PDF:	0
XML:	0