Quantitative Precipitation Estimation Using Weather Radar Data and Machine Learning Algorithms for the Southern Region of Brazil

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The Quantitative Precipitation Estimation method is a foundational pillar in hydro-meteorological sciences, with far-reaching implications for energy generation agricultural planning, and environmental conservation. Utilizing tree-based machine learning algorithms, such as Random Forest and Gradient Boosting, this research analyzes polarimetric variables to capture intricate patterns within the Z-R relationship. This work is very important to assess local precipitation using Machine Learning. The context of this manuscript is approvable for aim of remote sensing. However, there are some points that should be improved before considering its publication in Remote Sensing.

Major comments:

1.     This work is very good. For the structure of this manuscript, the authors have integrated the results and discussion. Considering that the aim of this study is proposed a Machine Learning method, I think it would be better to deal with the results and discussion separately.

2.     I noted that the accuracy of the model test phase was better than that of the validation phase. How should one go about understanding the problem. In addition, the modeling results of RFRF and GBRF were the best from Table 7, while the GBGB was the best model to capture rainfall from the Fig.5 and Table 9. All models overestimate rainfall, so which one is best for us to use instead. Please further explain.

Specific comments:

1.      L22: Please explain the Z-R relationship in detail and its importance to capture a rainfall.

2.      L68-72: These sentence present an importance of this study, which should have been removed to “Introduction”

3.      L99: “These relationships are chosen for comparison, incorporating methodologies from [4], [15], and [14]”. No clear.

4.      L253 and L254: Equation 2.5? Change to equations (11) and (12), respectively.

5.      L403-408: Please give further evidence.

6.      L458-461: Please add a reference.

7.      L482-486: These sentences were short of this manuscript, not conclusion. Please remove to discussion. In addition, the conclusion is needed to regroup.

 

Figures and Tables

1.      Figure 1: Add the longitude and latitude.

2.      Table 1: spatial denotes spatial resolution? 1°Í 250 m_ 1 elevation ?

Author Response

We would like to express our gratitude to the reviewer for the constructive feedback and guidance. The insights presented have been invaluable in enhancing the quality and clarity of our manuscript.

This manuscript presents our efforts on improving quantitative precipitation estimation using dual polarization radars in Brazil. These results are incorporated in our operational routines and are very reliable in our daily weather monitoring.

Regarding the reviewer's comments, we would like to clarify that:

1 - We concur that, for certain readers, the separation of these sections would improve on the text. However, we chose to join the Results and Discussion sections to improve readability, given that these are the first results of our ongoing research. We believe that keeping these sections together can create a smoother flow for the reader. They can see the findings and their interpretation immediately, without flipping back and forth between sections.

2 - The difference in accuracy between the test and validation phases is an important observation. The higher accuracy in the test phase may indicate an overfitting of the model to specific features of the test set. 

The difference in accuracy between the test and validation phases is an important observation. The higher accuracy found for the test indicates the specific features of the test set. The training dataset was balanced, meaning that half of its events were of non precipitation cases and the other half was of precipitation events. This was done to ensure that the model had sufficient precipitation events to train with, since most of the time there is no rain in our region of interest. In contrast to that, the test dataset was not balanced, meaning that it contained more non precipitation events than precipitation events. This was done to represent the reality of rain occurrences in the region. Thus, the model displayed a higher accuracy, considering the greater number of non precipitation events. 

Regarding the choice of the model, the results vary depending on the metrics and the characteristics of the dataset considered. The GBGB model showed the best performance in capturing precipitation as per Figure 5 and Table 9, suggesting its superiority under specific conditions. However, all models tend to overestimate precipitation, which will be a focus for adjustment and improvement in future studies.

Reviewer 2 Report

Comments and Suggestions for Authors

Please use other tree-based methods to perform the analysis such as the ExtraTreesClassifier to strengthen your analysis.

Rest all looks fine

Q2. Why do you use KGE?

 

Q3. Are there any other parameters that can be considered before feature selection? If yes, can you include them and carry out an analysis of different feature selection techniques.

 

 

Comments on the Quality of English Language

The quality of English is good

Author Response

Thank you for your feedback and constructive suggestions for improving our article. We acknowledge your recommendations and updated the text accordingly.

1-  We appreciate the suggestion to include other tree-based methods, such as the ExtraTreesClassifier, in our analysis. We agree that using various methods can strengthen our analysis, and we look forward to exploring these techniques in future studies. This will allow us to compare and contrast different approaches and assess their effectiveness in predicting precipitation in the southern region of Brazil.

2 - We chose to use KGE as a validation metric due to its ability to provide a comprehensive and comparable assessment of the performance of precipitation prediction models. KGE is a universal metric that can be applied in any region of the planet, facilitating comparison between different studies and enabling consistent evaluation of model performance in various geographic and climatic contexts.

3 -We placed emphasis on maintaining a physically meaningful feature set, aligning with the inherent characteristics of meteorological data. However, we also acknowledged the potential benefits of incorporating alternative feature types to enhance prediction accuracy. 

Reviewer 3 Report

Comments and Suggestions for Authors

However, there are a number of studies related to this subject. The novelty of the current study should be sufficiently delivered with results and comparisons.

1. please test more ML methods available including SVM, other RF based methods and ann.

2. provided results are not much well organized and even not much concise.

- Improve the quality of Figure 1 better map can be found.

- Eleminate Figure2

- All figures should not contain the title.

- Captions of all figures must be improved.

- Figure 3 caption should be improved. It it not English. I don't see much value of this figure.

- Figure 4 should be improved to show the whole procedure excluding its diagram. Improve caption.

- Figure 5 not all lines were explained. What are the lines of red and black lines.

- Figure 6 use (a),(b),(c), and (d). What do you mean filters and what is the threshold to reject. The caption should be designed to be self-explanatory.

- Fig7-10 use symbols (a)-(f) what each expalins should be included in detail.

 

 

Comments on the Quality of English Language

Extensive english editing is required.

Author Response

We appreciate the valuable feedback provided by the reviewer, which has helped us identify areas for enhancement in our manuscript. We have incorporated your suggestions and provided some comments on your questions, which will clarify our reasoning.

1 - We appreciate the suggestion to test additional machine learning methods. In response, we have initially considered including a broader spectrum of algorithms. However, our proposal utilizes tree-based machine learning algorithms due to their computational efficiency and effectiveness in handling large datasets typical in meteorological applications. While Support Vector Machines (SVM) are powerful for classification tasks, their application in operational environments for precipitation prediction is limited due to significant computational demands. SVMs often exhibit slow performance with large datasets, a critical drawback in real-time forecasting scenarios. Moreover, although neural networks offer extensive capabilities, especially in leveraging large volumes of radar data for more accurate forecasts, their implementation is part of our future work to expand the methodology. For the current study, we prioritized computational efficiency and scalability, which are crucial for practical applicability in operational forecasting systems. Thus, our focus remains on enhancing and comparing tree-based models like Random Forest and Gradient Boosting, which are well-suited for this context.

2 - We have reorganized the results section to enhance clarity and conciseness. We revised the text to ensure that it is intelligible and that our interpretation of the results is explicit. We also restructured the presentation of our data to facilitate an easier comparison across the different machine learning models developed in our study.

Reviewer 4 Report

Comments and Suggestions for Authors

Please go over the attached comments.

Comments for author File: Comments.pdf

Author Response

We would like to express our gratitude for the thorough review and insightful comments, which have provided us with an invaluable perspective on our manuscript. We appreciate the work you put forward in the revision, and have incorporated your suggestions in the revised text. Hereafter in this letter we wish to provide some comments on your suggestions, to assure that we were able to clearly communicate our work.

 

1 - The difference in accuracy between the test and validation phases is an important observation. The higher accuracy found for the test indicates the specific features of the test set. The training dataset was balanced, meaning that half of its events were of non precipitation cases and the other half was of precipitation events. This was done to ensure that the model had sufficient precipitation events to train with, since most of the time there is no rain in our region of interest. In contrast to that, the test dataset was not balanced, meaning that it contained more non precipitation events than precipitation events. This was done to represent the reality of rain occurrences in the region. Thus, the model displayed a higher accuracy, considering the greater number of non precipitation events.

 

2 - The slopes are defined according to the model’s attributes. Given that we are dealing with a quantitative estimative precipitation task, this implies a rapidly updated model. Thus the models become increasingly different as they learn from the dataset provided.

 

3 - We acknowledge the importance of keeping our results reproducible. We updated the text to include a more in-depth description of the model’s training method, including the fine-tuning employed and hyperparameters values selected. 

Reviewer 5 Report

Comments and Suggestions for Authors

Dear authors, my general comments and suggestions can be found below:

 

Abstract: Please clarify “to capture intricate patterns within the Z-R relationship.” Also, the authors should provide quantitative values of the better results found using dual-polarization weather radar data and ML algorithms in contrast with QPE from ZR relationships.

Line 27: The presented limitations of ZR relationships are not intrinsic to “Western Paraná” region. The authors must rewrite and provide more general information in this sentence.

Line 29: It is kind of strange to study the “limitations” of something. Perhaps applications and accuracy are more suitable.

Line 31: The authors mention three ML models but specify only two. Please check.

Line 35: What is a “hybrid approach”? Please specify.

Line 46: Now are two ML models. Not anymore, three.

Line 72: When did the radar start to operate?

Lines 77-85: The explanation of Figure 1 should placed before it. Similar to Table 1.

Lines 97-103: Why are the authors using non-calibrated ZR relationships? Why was the choice of the three ZR presented?

Line 108: Please explain the choice of the 0.2 mm/15 min threshold in detail.7

Line 113: Chance subjective term “majority” for a quantitative value.

Lines 125-126: What techniques were used to fill the data gaps?

Lines 142-143: What consistency methods were used?

Lines 154-157: The authors did not mention that they also used altitude and Distance as variables for ML models.

Lines 159-160: Please cite the references.

Line 162: Please explain in detail the conversion of dBZ to mm6m-3.

Table 3 is poorly explained. Please give details about the coefficients presented.

Lines 196-197: Is it essential to use classification ML models to distinguish between “rain” and “no rain” events? Why not start with the “rain” events (above 0.2 mm/15 min?)? The authors mentioned such rain thresholds but now used ML models to do such work. The methodology is not clear. Please check and rewrite.

Table 6: The authors must present the equations obtained from the regression results using the ML models.

Table 7: Please explain in detail. Were the ZR relationships also used in ML approaches?

Table 9: Please explain why moderate rain presented better results than light rain.

 

Section 3.2: Instead of presenting all the coverage radar, the authors could zoom in on the area with rain gauges and show a comparative analysis between the observed rainfall and the estimated precipitation.

 

 

Comments on the Quality of English Language

Dear Editor,

Please check my comments to the authors.

Best regards.

Author Response

We wish to express our gratitude for your insightful comments and suggestions, which have significantly contributed to enhancing the quality and clarity of our manuscript. We have carefully considered each of your points and have made the following amendments to the revised manuscript.

We updated the abstract to better present our objectives and goals, which included the analysis of the complex patterns present in the ZR relationships. Quantitative values were also included to better denote the performance of our developed model in contrast to other methods.

We appreciate your remark about the limitations of the ZR relationships. The text has been rewritten to provide a more broad sense of the use of the ZR relationships in our region of interest. Additionally, we also changed “limitations” to “applications”, which better express our proposed goals.

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors improved the quality of the manuscript following the suggested comments. It seems now it is desirable to publish the manuscript as is. 

Reviewer 4 Report

Comments and Suggestions for Authors

I would like to recommend the manuscript for publication.

Reviewer 5 Report

Comments and Suggestions for Authors

The authors made all the modifications suggested in my previous revision.

I consider this new version able to be published in the Remote Sensing Journal.

Best regards.