A CatBoost-Based Model for the Intensity Detection of Tropical Cyclones over the Western North Pacific Based on Satellite Cloud Images

Round 1

Reviewer 1 Report

The paper is devoted to development and testing of deep learning algorithm to detect tropical cyclone (TC) intensity. The Catboost-based model uses infra-red images from geostationary satellites to determine parameters of cloud top brightness temperature field, particularly, its gradients related with TC maximum wind speed, to further train the neural network.

 

I do not feel qualified to judge the construction of the deep learning model, but the result seems promising. The idea of temporal interpolation greatly improves the estimation accuracy. Two case studies show that the model is able to reproduce the evolution of TC intensity quite satisfactory, though the agreement with the testing data is still not perfect, at least for the stage of rapid intensity growing, and the model still needs further improvement, probably in future studies.

 

My general recommendation is major revision, considering more accurate presentation of method and results. Style and language improvement are also needed, the present version seems quite crude with a lot of typos, incorrect formulations and even slip-ups, like on lines 105-111.

 

Specific comments are listed below.

 

1. As far as can be understood, this model is an improved version of Authors’ previous developments (Zhong et al. 2020), but it is not clear from the text what is the difference, starting from Introduction (lines 89-100). As mentioned in the text, there were already six physical factors used in the previous study. What factors are new here? Please specify in Introduction what is the difference and novelty of the method presented here.

1a. There are no figures or some quantitative comparison of the result of the previous and the present models. It is only pointed, lines 489-491, that “compared with the pure CNN models used in past studies, the Catboost-based model based on prior physical factors has smaller errors in detecting TC intensity and has a better intensity prediction performance, especially in the early stage of TC intensity.” Please add more information/tables/figures to prove this statement.

 

2. More information and references are needed to cover the previous developments on TC detection and their parameter description using satellite methods (not only machine-learning). Studies on Dvorak technique improvements (lines 57-58) should also be referred.

Also, there is no mention of the Catboost model of the boosting integrated learning algorithm in Introduction. Please add more details with references.

 

3. The Authors reduce the resolution of satellite images from 5 km to 10 km, but at the same time they argue, e.g. lines 433-434, that “potential temperature gradient information is masked due to the limited resolution of the satellite, which results in underestimation at high intensities” and that “satellite technology advances are expected to further improve the model effect” (lines 463-464). It is confusing, as original images already have better resolution. Though I understand that resolution reducing saves computational time, it is interesting to see the result of simulations using original resolution.

4. As seems from Sec 3.1-3.2, there was no need to introduce both RMSE and MAE. They both exhibit similar behavior, and single RMSE could be enough to present the method errors. The statement about difference between these parameters, lines 301-303 (“RMSE is very sensitive to the large or small errors in dataset”, “MAE can better reflect the actual situation of the error of the predicted value”) is disputable if not incorrect, and this difference is not discussed further.

 

5. More specified classification for TD, TS, TY, … would be helpful (e.g. wind speed range for each).

 

6. Lines 220-225 and Table 1. Lat_TC and Lon_TC are probably not the physical factors, but just additional information? The authors state that “these factors have a strong correlation with the TC intensity”. How the position of TC center can correlate with its intensity?

 

7. 400-461 “As analyzed above, when the intensity is too high, the CTBT gradient information will be masked, and the discrimination ability of characteristic factors will be reduced.” – Nevertheless, the maximum intensity, and the intensity at the TC decay phase are detected very well in this case. Thus, the high intensity (actually, 30-40 m/s is not too high for a TC) is not a proper explanation. Can any another reasoning be suggested to explain such strong overestimation during the stage of TC rapid intensification?

 

8. Abstract, lines 24-26 - Not clear (smaller-larger, stronger-weaker, etc), please rewrite, maybe in two sentences

 

9. Please provide the data source where the satellite data were taken from.

As mentioned above, editing of English language and style are required.

Here are some (but not all) misprints, but more thorough revision including sentence construction all over the text is needed.

 

L. 21 expand -> expands

L. 62 This -> this

L. 175 “the the”

L. 254 is continuously -> are continuously

L. 280 to analysis -> to analyze

L. 454 a extratropical -> an extratropical

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

General comments:

The authors proposed Catboost-based TC intensity detection model using various morphological factors, which could be obtained from the geostationary satellite. Since there is significant relationship between structural factors and TC intensity, this research could contribute to quantify the characteristics of TCs. However, there are several logical issues for the modeling. After revising them, we could consider the publication of this paper.

 

Major comment:

1.     Is there any reason for using Catboost? As you mentioned in the description, the catboost model is suitable to the categorization problem using categorical variables. However, the TC intensity detection is one of the regression problems and all variables you used in this study are continuous factors. There are several simple machine learning algorithms for regression modeling, therefore, you should compare with other regression models. Otherwise, you have to clearly describe the reason why you use the model.

2.     Line 272: The random distribution of all datasets for training and test does not make sense. Since temporally closed TCs have similar spatial characteristics and intensity, they should be used for training “or” evaluation. If you distributed the datasets year-wisely, it should be described in the description.

3.     Chapter 2.2: The variable MMV is the minimum DAV value for your modeling, but the average of low DAVs (ex. less than 1% of DAV values) would be more suitable for your model. The 1 pixel (~10 km) is too local to represent the symmetricity of the cyclonic system.

4.     Chapter 2.2: It is hard to understand the meaning of each input variable. I suggest providing some visual representations of each variable, especially DAV-based variables.

Minor comment:

1.     Chapter 2.1: You have to mention the difference of FY-2G and FY-2F. How they fill the temporal gap of each other? Are they have exactly same specification?

2.     Contents (i.e., figures and tables) should be mentioned prior to be shown.
FYI, Figure 1,4 and Table 2 were shown prior to the mention in the text.

3.     Line 194: What means 2400 deg? The exact meaning of each variable should be described in the draft.

4.     Figure 4: Add scheme information in the figure.

5.     Figure 4: In the case of depth 8 in figure 4 (c), there is a significant increase in error according to estimator increase from 3000 to 4000, whereas this increase is not significant when interpolation datasets are utilized. Is there any reason of that?

Moderate editing of English language required.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I’m quite satisfied with the answers and text improvements the Authors provided. The paper looks more accurate and clearer now. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

I appreciate your consideration. Most comments were well reflected, and readability is enhanced. However, there are some rooms for improving your paper.

Regarding on the response of Major#2, I partially agree with your opinion, but it is obvious that temporally closed samples, used for training, could be a hint for validation. I checked the papers mentioned in your response. They additionally tested on the independent duration, and demonstrated the robustness of the model. I suggest conducting additional evaluation on the samples in 2019 or , which were not used in your modeling, then, the robustness of your model could be proved.

As minor comments, there are numerous typos and incorrect numbers in the contents. It should be carefully checked and modified.

1.     Figure 6 and figure 9 are still cut off.

2.     Line 26: TCsthan à TCs than

3.     Line 37, 49, 51, 420, 422 … : Most numbering in the contents and reference are incorrect. They should be carefully modified. 

4.     Figure 2.3: Please increase the font size in the figure.

…

Should be modified

Author Response

Please see the attachment.

Author Response File: Author Response.docx