Radar-Based Precipitation Nowcasting Based on Improved U-Net Model
Round 1
Reviewer 1 Report (New Reviewer)
Comments and Suggestions for AuthorsThe manuscript proposes a Double Recurrent Residual Attention Gates U-Net (DR2A-UNet) deep learning model to carry out radar echo extrapolation, with the quantitative rainfall estimation obtained. The model is compared with the U-Net and U-Net++ models as benchmarks. It covers a currently common interest but is far from being accepted in the current form. Specific issues need to be addressed carefully.
1. Firstly, many tracked revisions are still shown in the submitted manuscript, affecting its readability.
2. Although the U-Net model has been known well, the description is far from enough in Section 3.1.1. For instance, what is the meaning of marked numbers in Fig. 3(a), and the arrows, boxes … In particular, the U-Net++, which has not been so popular, needs much more contents. The matrix X, the superscripts, and many other elements in the structure. Similar cases for the other models … Additionally, are the parameter settings (e.g., learning rate, batch size etc.) all the same among these models?
3. Overall, the paper uses the multiple DL models, i.e., U-Net, U-Net++, DR2A-UNet and ConvLSTM. May the authors write in detail the respective purposes of these models in the current study.
4. As the authors introduced, the U-Net-like models are characterized by capabilities of ‘improve the image feature extraction ability and feature reduction ability’. In such context, how can they be used for radar echo extrapolation? I have not seen any temporal information included.
5. The dataset description is not detailed enough. Any preprocessing? Besides, the division for model training and examination are not clear… are they independent here? Such information must be supplemented carefully.
6. The manuscript organization is a bit confusing. In prediction evaluations, why only one lead time is selected for Sections 4.1 (1 hour) and 4.2 (2 hours)? Can the authors provide more on the model performance with different lead times to see how the models behave with the increasing lead times? Besides, in Section 4.4, in addition to the presented metrics such as correlation coefficient and bias, the POD, FAR, CSI and F1 results are even more necessary for the nowcasting details.
7. Due to that the experiment settings and comparisons are not clear enough, I could not judge more on the currently analyzed results… The revised manuscript would be expected for further review.
8. English of the manuscript must be improved.
Comments on the Quality of English LanguageEnglish of the manuscript must be improved.
Author Response
On behalf of all the contributing authors, I would like to express our sincere appreciations of your letter and reviewers' constructive comments concerning our article entitled "Radar-based precipitation nowcasting based on improved U-Net model" (Manuscript Number: remotesensing-2922891) for resubmission. We have studied and discussed all your comments point-by-point carefully with our co-authors, and accordingly made substantial revisions to our paper. All the changes we have made were in the red-colored text in the revised manuscript. Our point-by-point responses to your comments are provided below in the blue-colored texts. Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report (New Reviewer)
Comments and Suggestions for AuthorsThank you very much for the great work presented here.
I few items caught my attention:
· Page 2, Line 51, should be “… which makes”, not “make”.
· Page 2, Lines 77-82, consider splitting into at least two sentences (e.g. “… ground truth echoes. The dynamic Z-R relation …”.
· Page 3, Lines 96-100, similarly as above, for example:
o Sentence 1 ending with “… Hebei Province China.”
o Sentence 2 “Data consist of hourly rainfalls … and Liulin.”
o Sentence 3 “The typical rainfall process … the accuracy of nowcasting.”.
· Page 5, Lines 145-148, similarly as above, first sentence shuld end before “and”, “… stacked and combined with each other.”. Second sentence should start with ”The network has two encoder-decoder …”
· Page 5, Line 150, end the sentence here “… the structure is shown in Fig. 4b.”
· Page 5, Line 151, typo, it should be “.” (dot), not comma in front of “The recurrent residual …”
· Page 5, Line 155, break sentence here “and gradient disappearance. The recurrent convolution …”
· Page 5, Line 162, break the sentence here “… decoding structure. Point-by-point addition…”
· Page 5, Line 170, dot at the end “… memory model. In this study, …”
· Page 6, Lines 186-187, maybe: ”Each loss function is shown in the following equations, (1) and (2). The weight changes with the radar reflectivity value is shown in equation (3).”
· Page 6, Line 193, typo, missing “-“ and “e”, “… in the n-th image …”
· Page 7, Line 199, break sentence: “and predicting the second 20 maps). Then, the hourly rainfall intensity …”
· Page 8, Line 242, “… in equations …”, plural.
· Page 8, Lines 261-262, “… used as thresholds for the proposed evaluation the method…”
· Page 8, Lines 262-263, avoid personal pronouns, maybe “The 0/1 matrix was calculated by analyzing whether the reflectivity level exceeds the defined threshold.”
· Page 8, Line 264, present tense “Table 2 shows …”
· Page 8, Lines 265-267, maybe “All deep learning models outperform the prediction of 20dBZ threshold echoes when compared to the prediction of 30dBZ threshold echoes, as it was shown in Han, et al. [19].”
· Page 8, Lines 270-271, maybe “Compared to the baseline U-Net model, the UNet++, DR2A-UNet demonstrates improved accuracy of echo extrapolation.”
· Page 9, Line 302, no dot nedded after “Figures”
· Page 9-10, Lines 302-308, consider breaking into multiple sentences, the meaning is unclear.
· Page 12, Line 349, present tense: “Figure 9 shows…”
· Page 13, Line 367, similar as above “… accuracy are shown in Table 5.”
· Page 13, Lines 378-380, I don’t understand the meaning of this sentence “In addition, systematic errors in the radar data, homogenization of echoes generated by deep learning algorithms [33], and the DR2A-UNet model significantly outperforms the other control models for heavier rainfall.” Portion in front of “and” is incomplete on its own. The second part doesn’t seem to connect to the firs part either.
· Page 13, Line 388, present tense, “Figure 10 shows …”
· Page 14, Lines 395-401, consider breaking into multiple sentences, it is very difficult to follow the intended meaning.
· Page 14, Lines 402-404. Please elaborate how is the algorithm time error of 1 hour for nowcasting acceptable. Doesn’t that defeat the whole purpose of nowcasting? Are you concerned only on predicting the rain rate, but it is irrelevant when it happened?
Comments on the Quality of English Language
There is a persistent usage of long and convoluted sentences in the paper. I would suggest revising all those cases before publishing. I have listed more prominent issues in the review part.
Author Response
On behalf of all the contributing authors, I would like to express our sincere appreciations of your letter and reviewers' constructive comments concerning our article entitled "Radar-based precipitation nowcasting based on improved U-Net model" (Manuscript Number: remotesensing-2922891) for resubmission. We have studied and discussed all your comments point-by-point carefully with our co-authors, and accordingly made substantial revisions to our paper. All the changes we have made were in the red-colored text in the revised manuscript. Our point-by-point responses to your comments are provided below in the blue-colored texts. Please see the attachment.
Author Response File: Author Response.docx
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsGeneral comments:
This manuscript studies the radar-based rainfall nowcasting. The backbone is a U-Net structure which is modified by recurrent residual and attention gates. Balanced loss functions and dynamic Z-R relationships are also used for improving nowcasting. However, the ideas are not innovative enough as there are already several related research in the field. The experimental design is also flawed, and the presentation of results does not clearly show the advantages. The overall writing is too poor and some spellings are also naive. In conclusion, I am not recommended to accept this manuscript for publication in its current form. Some specific comments are listed below:
Specific comments:
1. Consider the Title, the expression of the “radar rainfall nowcasting” and “U-Net optimization model” is inadequate, which can be revised to “radar-based precipitation nowcasting” and “improved U-Net model”.
2. In Section 1, the cited literature is not at the forefront, the introduction is only a simple list and lacks logic. The authors should emphasize their advantages and weaknesses and then lead to their innovation points.
3. In Section 2.1, does “the local area of short ephemeral precipitation occurs frequently” really caused by “the small size of the watershed area”? It’s hard to understand.
4. In Section 2.2, the events 20211005 and 20221001 are close in total rainfall and maximum rainfall intensity, how to distinguish they are different types of rainfall processes? And does it can use technical terms to refer to three events, not in digits?
5. In Section 2.3, it seems like the study watershed only takes up a small part of the radar scanning scope, which area is exactly used for training and testing? The time and count information of the train and test set should also be provided.
6. In Section 3.1, the introduction of both the control model and DR2A-UNet is unreadable. The figures and corresponding text are hard to relate and understand. The structure parameters are not given. The authors should also compare their methods to more state-of-the-art and RNN-based or GAN-based models.
7. In Section 3.2, how does the author set the weight specifically?
8. In Section 3.3, the sub-title “Loss function” is wrong.
9. In Section 4, why did the authors choose 20 and 30 dBZ thresholds? Can they represent moderate or strong rainfall? And the quantitative evaluation should done on the whole test set to verify the superior of methods, only case studies are insufficient.
10. On Page 14, Lines 390 and 391, the use of “rainfall pictures” and “radar photos” are too unprofessional and inconsistent, the authors should check all mistakes of this type and make corrects.
11. The Figures of this paper seem a little blurry, the authors should replace them with more clearer versions.
Comments on the Quality of English LanguageThe overall writing is too poor and some spellings are also naive, some expressions are even hard to understand. The authors should try their best to polish the article.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper “Radar rainfall nowcasting based on U-Net optimization model” compares three nowcasting models based on deep learning. The accuracy of precipitation forecasts for 1h and 2h lead times are evaluated. In some parts the text is unclear and needs clarification. Since I consider the ambiguities essential for the evaluation of the methods used, I suggest a major revision.
I have the following main comments:
1. I did not find explicitly written in the text the horizontal resolution of the forecast and the form of the forecast (accumulated precipitation, current precipitation, etc.)
2. What data was used to train the model and test the model?
3. Three types of rain rate estimations based on radar data are mentioned but it is not clear which is used in the single verifications.
4. It is typical to calculate bias as a ratio of forecasted and measured values. The used additive bias is not suitable when we do not know total precipitation.
5. Sometimes a word “relative” or “slightly worse” is used in places where more specific wording should be used (e.g. L20, L290, L336, L420, L430(was low)).
Specific comments:
L170: I do not understand what do you mean by “frame”.
L182: Could you clarify the sentence expressing how coefficients are calculated? What data are used?
L206: What do you mean by mean bias and average root mean square error?
L228: Could you explain “indicating that … extrapolation”. What is better echo extrapolation?
L246: The sentence needs reformulation. I think that “strong” is not the right word.
L248: What do you mean by “in 30 min”?
L270: In my opinion “certain degree” is too polite. The model simply underestimates.
L272: I do not understand the sentence.
Table 4: What do you mean by /mm? Here you use 20211006 (and also in L324) but in another places I see 20211005. Large negative correlations need explanations.
L338: I am puzzled by the description of Figure 8. I think that the prediction in the figure is not good. How am I to understand that the hourly forecast is off by an hour? This is not very encouraging.
L353: The sentence is too optimistic. You did not show that the model nowcast actual rainfall process.
L355: I do not think that the model has small error.
L362: I do not understand the meaning of the sentence
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors endeavored to conduct echo extrapolation using the Double Recurrent Residual Attention Gates U-Net (DR2A-UNet) deep learning model and subsequently compared the rainfall estimates derived from the dynamic Z-R relationship with those from U-Net and U-Net++. This study shows both strengths and weaknesses across different aspects, depending on specific cases. Accordingly, the study doesn't seem to demonstrate significantly improved results when compared to alternative approaches. Nonetheless, the significance of this study lies in the expansion of methodology into the area of deep learning. In particular, if this study is further expanded, it is expected that the rainfall nowcasting can be improved to some extent. I recommend manuscript publication once the manuscript addresses and answers the following questions.
Major comments:
1. In the case of 20211005, characterized by scattered convective cells, the echo extrapolation for all approaches appears less effective. However, for organized systems like that of 20221001, the echo extrapolation shows improvement. This suggests a potential limitation in predicting echo evolution only based on radar data for disorganized cases, even when employing deep learning models. Hence, it becomes crucial to incorporate not only radar echoes but also environmental variables in future studies to furnish comprehensive storm evolution dynamics to deep learning models.
2. The number of data points used to compute the correlations presented in Table 4 and Figure 7 is insufficient. It is essential for the authors to augment the datasets to ensure the meaningfulness of the comparisons relying on correlation coefficients.
Minor comments:
1. line 25: There is a typo in the word ‘activite’.
2. line 80-84: The sentence appears to have grammatical issues.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
