CliGAN: A Structurally Sensitive Convolutional Neural Network Model for Statistical Downscaling of Precipitation from Multi-Model Ensembles
Round 1
Reviewer 1 Report
In this paper, the authors conduct a statistical downscaling methodology, called CliGAN, to transform large-scale annual maximum precipitation data generated by multiple AOGCMs to regional-level gridded data. They conduct an adversarial training approach in combination with a 3-loss component total loss downscaling method, to verify the effectiveness of the proposed downscaling structure. Although it is difficult to identify how the authors progressed beyond the state of the art and the text is generally mediocre-written; however, I find their effort interesting towards downscaling precipitation data with a robust, good-fitting and current methodology that exhibits good performance. The presentation of the results is satisfactory but there are several fragments in the text, or sentences that really do not bring forward a clear message or are confusing to the reader. Moreover, the authors in many cases don’t follow the MDPI template and many changes are needed to address this issue (e.g. in-text references numbering, spacing issues, e.t.c.). So, I propose some specific and general comments to the authors in order to improve their manuscript.
General comments
Firstly, you generally avoid using commas which many times makes it difficult for the reader to understand the text. So, I propose to read carefully the text and add commas where needed.
In introduction you give many details on specific aspects such as: i) why precipitation patterns are more difficult to be effectively simulated other weather variables, ii) why widely used loss functions fail to capture the special structure of precipitation, iii) why statistical downscaling outperforms the dynamic approach, iv) why transfer functions are gaining ground against other approached, v) why NNs and specifically CNNs are effective methodologies BUT too little is mentioned about the state of the art considering precipitation downscaling (thoroughly described examples) and how your approach outperforms other similar scientific efforts- meaning to identify the novelty of your study.
In Methods section you missed the subsection 2.2.2. while the titles of 2.1, 2.2, 2.2.1, 2.2.3 and 2.2.4 don’t need a colon at the end.
In Results and Discussion section you could use convergence diagnostic tools which can officially indicate the convergence period (how many iterations are needed for convergence- confirm the 10,000 iterations) and the burn-in period (how many initial iterations must be rejected from the sequence-confirm your findings).
In Results and Discussion, i) in lines 315-320 I don’t find the repetition of Table 2 results necessary- instead you could only refer to the best loss combination, ii) you could move the Figures S1-S4 from the end of this section, to the exact point in the text where they are discussed.
Line-specific comments
Line 28: Instead of “methodology” use “methodologies”
Line 101: add “that” after it should be noted…
Line 113: instead of “are” use “is”
Line 129: instead of “there are” use “there is”
Line 162: Dot at the end of the sentence is missing
Lines 176-177: Please rewrite the sentence to make clear to the reader what you want to say
Line 196: You could mention the interpolation method
Line 309: Instead of “function” use “functions”
Line 323-325: Please rewrite the sentence to make clear your statement
Line 328: Use a semicolon after the word “iterations”, otherwise rewrite the sentence
I would suggest to carefully read and correct many sentences in the text to succeed a smooth reading experience for the reader. (there are several sentences that need to be carefully corrected)
To this end, I recommend the manuscript for publication with major revisions as proposed above.
Author Response
Reviewer 1:
In this paper, the authors conduct a statistical downscaling methodology, called CliGAN, to transform large-scale annual maximum precipitation data generated by multiple AOGCMs to regional-level gridded data. They conduct an adversarial training approach in combination with a 3-loss component total loss downscaling method, to verify the effectiveness of the proposed downscaling structure. Although it is difficult to identify how the authors progressed beyond the state of the art and the text is generally mediocre-written; however, I find their effort interesting towards downscaling precipitation data with a robust, good-fitting and current methodology that exhibits good performance. The presentation of the results is satisfactory but there are several fragments in the text, or sentences that really do not bring forward a clear message or are confusing to the reader. Moreover, the authors in many cases don’t follow the MDPI template and many changes are needed to address this issue (e.g. in-text references numbering, spacing issues, e.t.c.). So, I propose some specific and general comments to the authors in order to improve their manuscript.
General comments
Firstly, you generally avoid using commas which many times makes it difficult for the reader to understand the text. So, I propose to read carefully the text and add commas where needed.
We corrected the syntaxes throughout the text and did a full review for grammar and flow.
In introduction you give many details on specific aspects such as: i) why precipitation patterns are more difficult to be effectively simulated other weather variables, ii) why widely used loss functions fail to capture the special structure of precipitation, iii) why statistical downscaling outperforms the dynamic approach, iv) why transfer functions are gaining ground against other approached, v) why NNs and specifically CNNs are effective methodologies BUT too little is mentioned about the state of the art considering precipitation downscaling (thoroughly described examples) and how your approach outperforms other similar scientific efforts- meaning to identify the novelty of your study.
We have included a paragraph to explain the state of the art of precipitation downscaling and relevance of study around that – as follows (Line no. 133-149):
“Among the climatological variables which are in general downscaled in practice, precipitation, perhaps, is most susceptible to various uncertainties (Hashmi et. al. 2009). The non-Gaussian nature of extreme precipitation owing to its localized nature creates problems for classical statistical estimators (Benestad 2010, Wilks 1995). Some recent studies have utilized more advanced statistical techniques and Bayesian methods in particular to downscale extreme precipitation over data sparse regions (Ro’fah Nur 2019, Enrico 2020). An example of direct application of machine learning techniques in statistical downscaling of precipitation is Vandal et. al. 2017 who used a generalized stacked super-resolution CNN to downscale daily precipitation over the US. Despite several limitations in their experiments, the result shows the efficiency and robustness of the approach over other methods in predicting extremes. Cheng et. al. 2020 has also recently introduced a novel residual dense block (RDB) into the Laplacian pyramid super-resolution network (LapSRN) to generate high-resolution precipitation forecast. Onishi et. al. 2019 used super-resolution techniques to simulate high-resolution urban micrometeorology, while Ji et al. 2020 proposed several CNN-based architectures to forecast high-resolution precipitation. Underlying all of these models is the treatment of two-dimensional fields such as climate model outputs and gridded observations as analogous to non-geographic images which makes CNNs an ideal candidate as transfer functions in statistical downscaling.”
[paste new paragraph]
In Methods section you missed the subsection 2.2.2. while the titles of 2.1, 2.2, 2.2.1, 2.2.3 and 2.2.4 don’t need a colon at the end.
We revised the subsection headings accordingly in the revised manuscript.
In Results and Discussion section you could use convergence diagnostic tools which can officially indicate the convergence period (how many iterations are needed for convergence- confirm the 10,000 iterations) and the burn-in period (how many initial iterations must be rejected from the sequence-confirm your findings).
We used initial 500 iterations as burn-in and 10,000 iterations as convergence iterations. These details are now clearly reported in the text as follows (Line no. 300-303):
“The initial 500 iterations were discarded as burn-in and subsequently the model was trained for 10,000 iterations until convergence. These settings were found to be more than adequate for simulating realistic high-resolution precipitation patterns.”
In Results and Discussion, i) in lines 315-320 I don’t find the repetition of Table 2 results necessary- instead you could only refer to the best loss combination, ii) you could move the Figures S1-S4 from the end of this section, to the exact point in the text where they are discussed.
We have removed the summary of entire Table 2 and only reported the best combination of losses in the revised manuscript (Line no. 331-346).
We have moved the Figure S1 as Figure 6 in the revised manuscript. However, we still retained all the other auxiliary plots (Figure S1-S3) which are used to compare the results but not discussed in detail to the supplementary section of the paper.
Line-specific comments
Line 28: Instead of “methodology” use “methodologies”
Corrected in our revised manuscript (Line no. 27).
Line 101: add “that” after it should be noted…
Corrected in our revised manuscript (Line no. 102).
Line 113: instead of “are” use “is”
Corrected in our revised manuscript (Line no. 112).
Line 129: instead of “there are” use “there is”
Corrected in our revised manuscript (Line no. 128).
Line 162: Dot at the end of the sentence is missing
Corrected in our revised manuscript (Line no. 168).
Lines 176-177: Please rewrite the sentence to make clear to the reader what you want to say
Corrected in our revised manuscript. The revised line is as follows (Line no. 189-190):
“We developed this downscaling methodology using annual maximum daily precipitation as our target variable.”
Line 196: You could mention the interpolation method
The “bi-linear” interpolation method is now mentioned in the manuscript. The revised sentence reads as follows (Line no. 204-206):
“To overcome this difficulty, we interpolated precipitation from different AOGCMs onto 10km resolution grids using bi-linear interpolation method.”
Line 309: Instead of “function” use “functions”
Corrected in our revised manuscript (Line no. 329).
Line 323-325: Please rewrite the sentence to make clear your statement
This sentence has been removed from the revised manuscript.
Line 328: Use a semicolon after the word “iterations”, otherwise rewrite the sentence
Corrected in our revised manuscript (Line no. 334).
I would suggest to carefully read and correct many sentences in the text to succeed a smooth reading experience for the reader. (there are several sentences that need to be carefully corrected)
We have revised the paper for flow and grammar to increase readability.
To this end, I recommend the manuscript for publication with major revisions as proposed above.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper proposed a new method to downscale GCM climate simulation using GAN with a generator network and a discriminator network. The total loss function is the sum of the adversarial loss, structural loss, and content loss. The results of downscaling 9 AOGCMs show better performance than using a subset of the loss functions or Mean Absolute Error, PCA. Overall, this is a good paper. I have only some comments on the presentation of the paper.
1.
Because the authors used the precipitation output from AOGCMs rather than other circulation predictors from GCMs, the "downscaling" method in the paper seems to conduct downscaling and bias-correction at the same time, rather than generating a fine resolution data of original simulated precipitation as downscaling implied.
2. The authors promote the methodology as CliGAN. Is the software published/shared in the community? Please clarify this information.
Other minor comments:
Lines 48-52: Though many physical mechanisms are incorporated in GCMs, many components/processes in the climate system are still represented by parameterization. These parameterizations differ from model to model. That is part of the reason for the wide ranges of simulation and projection of climate by different GCMs (another being the nonlinearity of the climate system). The degradation not only occurs on the higher-spatial and temporal-scale but also affects the large-scale simulation, as sh ? own by Fig. 3.
Line 176: "which is used a" -> "which is used as a"
Line 216: What is E stands for in Eq. 2?
Line 226: Eq(3), so the approximate earth moving distance is just D_thetaD(G_thedaG(P_GCM)) ?
Figure 5(c): Can the authors provide an explanation or hint why the structural loss increases quasi-periodically with epochs
Author Response
Reviewer 2:
The paper proposed a new method to downscale GCM climate simulation using GAN with a generator network and a discriminator network. The total loss function is the sum of the adversarial loss, structural loss, and content loss. The results of downscaling 9 AOGCMs show better performance than using a subset of the loss functions or Mean Absolute Error, PCA. Overall, this is a good paper. I have only some comments on the presentation of the paper.
1.Because the authors used the precipitation output from AOGCMs rather than other circulation predictors from GCMs, the "downscaling" method in the paper seems to conduct downscaling and bias-correction at the same time, rather than generating a fine resolution data of original simulated precipitation as downscaling implied.
We agree with the reviewer on this. We have made a note of this fact in our discussion of the input data (Line no. 200-201):
“Our statistical framework will not only create the fine resolution precipitation but will correct the biases of the AOGCM simulated precipitations as well.”
The authors promote the methodology as CliGAN. Is the software published/shared in the community? Please clarify this information.
The custom code used in this paper is still not released to the public and is available from authors upon request. We have made a note in this in the methods as follows (Line no. 281-282):
“The CliGAN model was implemented in python and is available via request from the authors.”
Other minor comments: Lines 48-52: Though many physical mechanisms are incorporated in GCMs, many components/processes in the climate system are still represented by parameterization. These parameterizations differ from model to model. That is part of the reason for the wide ranges of simulation and projection of climate by different GCMs (another being the nonlinearity of the climate system). The degradation not only occurs on the higher-spatial and temporal-scale but also affects the large-scale simulation, as shown by Fig. 3.
We agree the uncertainties in the AOGCM simulation is multi-faceted. So ideal downscaling method should not only make the spatial-temporal higher scale but also correct other inter-model and intra-model biases and differences. We have made a note of the sensitivity of the input data to model parameterization in the discussion (Line no. 64-66):
“As such ensemble methods attempt to capture a suite of AOGCMs by collating outputs from multiple model which in turn aims to reduce sensitivity to individual model biases through a quantitative framework (Knutti et al. 2010)”
Line 176: "which is used a" -> "which is used as a"
Corrected in the revised manuscript (Line no. 182).
Line 216: What is E stands for in Eq. 2?
E is the error of the discriminator. In our case, it is the earth moving distance between the input and the target distribution.
Line 226: Eq(3), so the approximate earth moving distance is just D_thetaD(G_thedaG(P_GCM))?
The response of the discriminator is the approximate earth moving distance. The discriminator is trained on the earth moving distance to calculate it for the simulated samples.
Figure 5(c): Can the authors provide an explanation or hint why the structural loss increases quasi-periodically with epochs
The generator learns the manifold of the data in stages due to the forcing of the discriminator response. We think that when it tries to learn a new manifold it moves away from its existing configuration resulting in increase of error and once it learns the manifold and stabilizes the error also reduces. Notice the decreasing amplitude of the peak of the error oscillation indicating learning of the new manifold of the data.
Round 2
Reviewer 1 Report
No comments
