Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. Compared to the PredRNN-v2, the improvement of SAM-Net in predicting typhoon cloud images is too small for 4 time steps，thus it not worth being highlighted in the abstract. Instead, there is a significant improvement of SAM-Net in predicting typhoon cloud images for 10 time steps, which can be mentioned in the abstract.

2. The self-attention memory module should be described in detail in the section 3.2. For example, in formula 7, what is the Km? The calculation formulas should be given, too.

3. There are 1200 series data of cloud images used in this study. Are they all typhoon cases? The introduction of data is unclear and the description of typhoon cloud images such as the source, spatial resolution and temporal scope should be supplemented in the article. In addition, the preprocessing of cloud images also need to be described in detail.

4. Which typhoon case is shown in Figures 7 and 8? This study lacks the introduction and discussion of typhoon cases.

5. The model in this paper is an improved model based on PredRNN-v2, but the structure of PredRNN was proposed three years ago. In recent years, the innovational image sequence prediction models such as SimVP and SCSTque have been developed. It would be better to compare the performance of SAM-Net with those of these advanced prediction models.

Comments on the Quality of English Language

The English writing of this article is generally easy to understand, but there are still some sentences that need improvement. For example, in line 393-394, 'According to our experimental data, PredRNN-v2’s performance in typhoon cloud image prediction is suboptimal. Thus, we improved upon the predRNN-v2 model and evaluated it …' It can be rewritten as 'Our experiment results indicate that the performance of PredRNN-v2 is not satisfied in predicting typhoon cloud images. Thus, we proposed an improved prediction model based on the predRNN-v2 and evaluated it …'.

Author Response

Response to Reviewers Comments

We would like to thank all Reviewers and the Editors for the valuable and constructive comments on our manuscript. We especially appreciate your conformation of the reviews about our contribution. We have revised the manuscript accordingly. We have submitted both the revised manuscript and the “Summary of changes” for your reference. If any further information is required, please feel free to contact us.

Summary of Changes

According to the reviewers’ comments, we have made five major improvements in the revised version:

[Major Change 1]

As suggested by reviewer # 1, we have improved the wording of the abstract and emphasized the significant improvement of SAM-Net in predicting typhoon cloud images in 10 time steps.

[Major Change 2]

As suggested by Reviewer # 1, we have provided a detailed introduction to the preprocessing of typhoon cloud imagery data and supplemented the sources, spatial resolution, and time range of the cloud imagery data used. Please refer to Section 4.1 (Typhoon cloud images) for more information.

[Major Change 3]

As suggested by Reviewer # 1 and Reviewer # 2, we have corrected the formula labeling errors that appeared in the paper and provided the missing calculation formulas.

[Major Change 4]

As suggested by reviewer # 3, we have added a set of generalization experiments for ordinary cloud images and provided a detailed description of the dataset used.

[Major Change 5]

As suggested by Reviewer # 1, Reviewer # 2, and Reviewer # 3, we have added a Discussion section to discuss the typhoon case studied in this paper, as well as the applicability, limitations, and future research directions of the SAM-Net model.

 

Responses to Reviewer #1

[Comment 1]

Compared to the PredRNN-v2, the improvement of SAM-Net in predicting typhoon cloud images is too small for 4 time steps thus it not worth being highlighted in the abstract. Instead, there is a significant improvement of SAM-Net in predicting typhoon cloud images for 10 time steps, which can be mentioned in the abstract.

Response:

Thanks for your careful review, We have revised the abstract to emphasize the significant improvement effect at a 10 time steps (cf. [Major change 1]).

[Comment 2]

The self-attention memory module should be described in detail in the section 3.2. For example, in formula 7, what is the ? The calculation formulas should be given, too.

Response:

Thanks for your careful review,  is similar to  mentioned in Formula 2, which is a key in the self attention memory mechanism. We have added the calculation formula for  in the description below Formula (7) (Page 6) (cf. [Major change 3]).

[Comment 3]

There are 1200 series data of cloud images used in this study. Are they all typhoon cases? The introduction of data is unclear and the description of typhoon cloud images such as the source, spatial resolution and temporal scope should be supplemented in the article. In addition, the preprocessing of cloud images also need to be described in detail.

Response:

Thanks for your valuable comments. The 1200 series of typhoon cloud images used in this study are all actual typhoon cases, including Typhoon Nida in 2016, Typhoon Hato in 2017, Typhoon Mangkhut in 2018, and Typhoon Higos in 2020. In Section 4.1 (Page 9), we provide a detailed description of the preprocessing(Section 4.2, Page 10) of typhoon cloud images, as well as the source, spatial resolution, and time range of typhoon cloud images(cf. [Major change 2]).

[Comment 4]

Which typhoon case is shown in Figures 7 and 8? This study lacks the introduction and discussion of typhoon cases.

Response:

Thanks for your valuable comments. The typhoons shown in Figures 7 and 8 are both Haigos typhoons, which we have added in the caption of the paper. In addition, we have included a Discussion section to introduce and discuss typhoon cases (Section 6, Page 17) (cf. [Major change 5]).

[Comment 5]

The model in this paper is an improved model based on PredRNN-v2, but the structure of PredRNN was proposed three years ago. In recent years, the innovational image sequence prediction models such as SimVP and SCSTque have been developed. It would be better to compare the performance of SAM-Net with those of these advanced prediction models.

Response:

Thanks for your valuable comments We tried to increase the experimental comparison between SimVP and SCSTque as soon as possible, but SCSTque has not yet released its own source code, so we cannot compare the SCSTque model. Some other prediction models, such as SimVP, concatenate two images together as input to the model, while our typhoon cloud image is a complete whole, so we did not compare with it. However, we will add them in the related work.

[Comment 6]

The English writing of this article is generally easy to understand, but there are still some sentences that need improvement ….

Response:

Thank you for your affirmation of our work. We have rewritten our sentences according to your suggestion (Section 7, Page 18).

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This study proposes the SAM-Net model, applied to the spatio-temporal sequence prediction of typhoon cloud maps, which combines the self-attention mechanism and the memory module to enhance the ability of capturing global spatial features with long-time dependencies. The paper is clear and the results show that the model outperforms the PredRNN-v2 model on several benchmark datasets. The performance of the SAM-Net model proposed in the paper in typhoon cloud prediction is impressive. However, the model needs to be improved considerably in terms of dataset size, length of prediction time, and model adaptability. In addition, the experimental results need further improvement, and the discussion about the limitations of the model is insufficient. It is recommended that the authors revise and resubmit the paper.

 

major:

(1) The article spends a lot of space in the results and analyses describing the performance of SAM-Net on the MovingMNIST and KTH datasets, demonstrating the spatio-temporal sequence prediction capability of SAM-Net, but the direct relevance of these datasets to the theme of the paper "typhoon cloud images prediction",  their correlation is weak. This makes it difficult for the reader to be fully convinced of the effectiveness of the model in meteorological applications.

(2) For typhoon cloud prediction, the paper only provides simple input and output frame prediction results, lacking in-depth analyses of specific cases. It is suggested that the model's prediction results should be analyzed and discussed in detail with real typhoon cases. The persuasive power of the article can be significantly enhanced by introducing case studies of real typhoons. For example, the performance of the model in typhoon generation, path prediction or intensity change is analyzed to demonstrate its advantages in handling complex typhoon cloud map data.

(3) The amount of data for typhoon cloud maps seems to be too small compared to the MovingMNIST and KTH datasets.

(4) In meteorological image prediction, besides the image quality assessment metrics MSE, SSIM, and PSNR, meteorological prediction-specific metrics such as CSI and FAR may be more applicable to the prediction of typhoon cloud maps. It is suggested that the authors supplement these indicators, which can more comprehensively assess the effectiveness of the model in real weather prediction.

(5) It is suggested to further discuss the adaptability and limitations of the model, especially the problems that may be encountered when dealing with larger scale or irregular data, and give corresponding solutions or future research directions.

 

minor:

(1) In the introduction of the 4.1 dataset, it is mentioned that for the typhoon cloud map dataset the first four frames are used to predict the last four frames, i.e., four hours prediction, which is different from the training of the other two datasets. Is it possible to do longer time prediction? However, the experimental results in 5.3.2 are predictions for the next ten frames, it confuses. Please explain how to get a prediction for ten frames and give a model error analysis similar to Figure 9?

(2) Too much formulas are used in the article, are they derived by authors? If not, please list the references. If it is some classical formulas, there is no need to list them all.

(3) Some symbols in the formulas and diagrams are wrongly labelled, e.g., ‘Q_h,K_h,V_h is the weight set of 1 x 1 convolution’ in line 194? What and where is the C? In the Figure 2 subfigure, what's the Qc?

(4) In the explanation of Eq. 28 for calculating the structural similarity SSIM, it is recommended to specify the value of pixel L in this study. For uint8 data, the maximum pixel value is 255, and for floating-point data, the maximum pixel value is 1 . In general, k1 = 0.01 and k2 = 0.03.

(5) Line 400, ‘In this study, we used 128 cloud images for prediction’, I think the authors may want to express that the image resolution is 128x128, not that 128 cloud images were used.

Comments for author File: Comments.pdf

Author Response

Response to Reviewers Comments

We would like to thank all Reviewers and the Editors for the valuable and constructive comments on our manuscript. We especially appreciate your conformation of the reviews about our contribution. We have revised the manuscript accordingly. We have submitted both the revised manuscript and the “Summary of changes” for your reference. If any further information is required, please feel free to contact us.

Summary of Changes

According to the reviewers’ comments, we have made five major improvements in the revised version:

[Major Change 1]

As suggested by reviewer # 1, we have improved the wording of the abstract and emphasized the significant improvement of SAM-Net in predicting typhoon cloud images in 10 time steps.

[Major Change 2]

As suggested by Reviewer # 1, we have provided a detailed introduction to the preprocessing of typhoon cloud imagery data and supplemented the sources, spatial resolution, and time range of the cloud imagery data used. Please refer to Section 4.1 (Typhoon cloud images) for more information.

[Major Change 3]

As suggested by Reviewer # 1 and Reviewer # 2, we have corrected the formula labeling errors that appeared in the paper and provided the missing calculation formulas.

[Major Change 4]

As suggested by reviewer # 3, we have added a set of generalization experiments for ordinary cloud images and provided a detailed description of the dataset used.

[Major Change 5]

As suggested by Reviewer # 1, Reviewer # 2, and Reviewer # 3, we have added a Discussion section to discuss the typhoon case studied in this paper, as well as the applicability, limitations, and future research directions of the SAM-Net model.

Responses to Reviewer #2

[Comment 1]

The article spends a lot of space in the results and analyses describing the performance of SAM-Net on the MovingMNIST and KTH datasets, demonstrating the spatio-temporal sequence prediction capability of SAM-Net, but the direct relevance of these datasets to the theme of the paper "typhoon cloud images prediction", their correlation is weak. This makes it difficult for the reader to be fully convinced of the effectiveness of the model in meteorological applications.

Response:

Thanks for your valuable comments. We have read some other papers on spatiotemporal sequence prediction of cloud images, which have compared the MovingMNIST and KTH datasets. Therefore, we also compared these two datasets. In addition, we conducted a set of experiments on ordinary cloud images and analyzed the results of typhoon cloud images and ordinary cloud images.

[Comment 2]

For typhoon cloud prediction, the paper only provides simple input and output frame prediction results, lacking in-depth analyses of specific cases. It is suggested that the model's prediction results should be analyzed and discussed in detail with real typhoon cases. ….

Response:

Thanks for your valuable comments. We have included a Discussion section to introduce and discuss typhoon cases (Section 6.Page 17). (cf. [Major change 5]).

[Comment 3]

The amount of data for typhoon cloud maps seems to be too small compared to the MovingMNIST and KTH datasets.

Response:

Thanks for your valuable comments. Due to the difficulty in obtaining typhoon datasets and the large amount of incomplete data in the obtained typhoon datasets, our dataset is relatively small compared to these two datasets. However, we are currently applying for more typhoon cloud map datasets. If we obtain the typhoon cloud map dataset, we will expand our dataset and reflect it in our future research work.

[Comment 4]

In meteorological image prediction, besides the image quality assessment metrics MSE, SSIM, and PSNR, meteorological prediction-specific metrics such as CSI and FAR may be more applicable to the prediction of typhoon cloud maps. ….

Response:

Thanks for your valuable comments. CSI and FAR indicators are more suitable for the accuracy of meteorological event forecasting, while the accuracy of cloud prediction in images may not be suitable. In addition, we have read some other papers on cloud image prediction, mostly using MSE, SSIM, PSNR and other indicators, and rarely using CSI and FAR to evaluate the accuracy of cloud image prediction.

[Comment 5]

It is suggested to further discuss the adaptability and limitations of the model, especially the problems that may be encountered when dealing with larger scale or irregular data, and give corresponding solutions or future research directions.

Response:

Thanks for your valuable comments. Following your suggestion, we have added a Discussion section to discuss the applicability and limitations of the model, as well as future research directions. (Section 6, Page 17) (cf. [Major change 5]).

 

[Minor Comment 1]

In the introduction of the 4.1 dataset, it is mentioned that for the typhoon cloud map dataset the first four frames are used to predict the last four frames, i.e., four hours prediction, which is different from the training of the other two datasets. Is it possible to do longer time prediction? However, the experimental results in 5.3.2 are predictions for the next ten frames, it confuses. Please explain how to get a prediction for ten frames and give a model error analysis similar to Figure 9?

Response:

Thanks for your valuable comments. In this study, we conducted experiments with both 4 and 10 time steps, and analyzed the results of the two different time steps in sections 5.3.1 and 5.3.2, respectively. In addition, as per your request, we have provided an error comparison chart for 10 time steps similar to Figure 9.

[Minor Comment 2]

Too much formulas are used in the article, are they derived by authors? If not, please list the references. If it is some classical formulas, there is no need to list them all.

Response:

Thanks for your valuable comments. We believe that readers who may come across this paper in the future may not be familiar with some classic models, so we have listed all the formulas one by one.

[Minor Comment 3]

Some symbols in the formulas and diagrams are wrongly labelled, e.g., ‘, ,  is the weight set of 1 x 1 convolution’ in line 194? What and where is the C? In the Figure 2 subfigure, what's the ?

Response:

Thanks for your careful review. Thank you very much for pointing out the mistake. We have corrected the correct weight set. In the formula, C represents the number of channels. In the subgraph of Figure 2, we mistakenly wrote  as . We have now corrected it.

[Minor Comment 4]

In the explanation of Eq. 28 for calculating the structural similarity SSIM, it is recommended to specify the value of pixel L in this study. For uint8 data, the maximum pixel value is 255, and for floating-point data, the maximum pixel value is 1 . In general,  = 0.01 and  = 0.03.

Response:

Thanks for your valuable comments. Based on your suggestion, we have explained in the article the range of pixel values and the values of  and .(Section 4.3, Page 10)

[Minor Comment 5]

Line 400, ‘In this study, we used 128 cloud images for prediction’, I think the authors may want to express that the image resolution is 128x128, not that 128 cloud images were used.

Response:

Thanks for your careful review. We have revised our way of expression.

 

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

I would like to thank you for submitting your article entitled "SAM-Net: Spatio-Temporal Sequence Typhoon Cloud Images Prediction Net with Self-Attention Memory." After carefully reading your paper, I wish to congratulate you on the innovation in your work. You propose a novel model that integrates a self-attention memory module (SAM-Net) to solve the complex problem of predicting cloud images in spatio-temporal sequences, particularly for typhoon forecasting. This approach is highly relevant, as allows for better capture of long-term spatial and temporal characteristics in meteorological images, representing a significant improvement over traditional models such as PREDRNN-v2. Your results also show substantial improvements across several datasets, including MovingMNIST, KTH, and satellite images of typhoon clouds.

However, a few aspects of your work require adjustments to strengthen the impact of your contribution, or better justification on why they are not addressed in this paper. First, it would be beneficial to explore optimization methods, such as model compression, to reduce resource consumption while maintaining an acceptable level of performance. This optimization would make your model more accessible and usable in practical contexts.

Additionally, although you have evaluated SAM-Net on diverse datasets, it would be relevant to extend testing to other meteorological phenomena, such as precipitation or strong winds. This would help assess the robustness and generalization of your model in various conditions beyond typhoons. SAM-Net's ability to adapt to other types of climate data would further enhance its utility to the scientific community.

Moreover, a more detailed analysis of prediction errors in complex cases would be beneficial. A deeper exploration of situations where the model fails to accurately predict certain irregular cloud movements could provide insights into potential biases and offer directions for further performance improvements. This could include a qualitative study of errors in specific cases, and you might also consider enhancing your discussion on this topic.

Finally, while SAM-Net appears to be effective on smaller images, you mention a degradation in results on larger-scale images. You should address this scenario and propose image processing techniques, such as as super-resolution or dimensionality reduction methods, to improve the clarity of predictions on large-scale images while minimizing information loss.

 

Overall, your work provides a valuable contribution to meteorological forecasting based on spatio-temporal sequences, and I encourage you to consider these suggestions to enhance the reach and accuracy of your model.

 

Author Response

Response to Reviewers Comments

We would like to thank all Reviewers and the Editors for the valuable and constructive comments on our manuscript. We especially appreciate your conformation of the reviews about our contribution. We have revised the manuscript accordingly. We have submitted both the revised manuscript and the “Summary of changes” for your reference. If any further information is required, please feel free to contact us.

Summary of Changes

According to the reviewers’ comments, we have made five major improvements in the revised version:

[Major Change 1]

As suggested by reviewer # 1, we have improved the wording of the abstract and emphasized the significant improvement of SAM-Net in predicting typhoon cloud images in 10 time steps.

[Major Change 2]

As suggested by Reviewer # 1, we have provided a detailed introduction to the preprocessing of typhoon cloud imagery data and supplemented the sources, spatial resolution, and time range of the cloud imagery data used. Please refer to Section 4.1 (Typhoon cloud images) for more information.

[Major Change 3]

As suggested by Reviewer # 1 and Reviewer # 2, we have corrected the formula labeling errors that appeared in the paper and provided the missing calculation formulas.

[Major Change 4]

As suggested by reviewer # 3, we have added a set of generalization experiments for ordinary cloud images and provided a detailed description of the dataset used.

[Major Change 5]

As suggested by Reviewer # 1, Reviewer # 2, and Reviewer # 3, we have added a Discussion section to discuss the typhoon case studied in this paper, as well as the applicability, limitations, and future research directions of the SAM-Net model.

Responses to Reviewer #3

[Comment 1]

I would like to thank you for submitting your article entitled "SAM-Net: Spatio-Temporal Sequence Typhoon Cloud Images Prediction Net with Self-Attention Memory." After carefully reading your paper, I wish to congratulate you on the innovation in your work. You propose a novel model that integrates a self-attention memory module (SAM-Net) to solve the complex problem of predicting cloud images in spatio-temporal sequences, particularly for typhoon forecasting….

Response:

Thank you for your affirmation of our work.

[Comment 2]

However, a few aspects of your work require adjustments to strengthen the impact of your contribution, or better justification on why they are not addressed in this paper. First, it would be beneficial to explore optimization methods, such as model compression, to reduce resource consumption while maintaining an acceptable level of performance. This optimization would make your model more accessible and usable in practical contexts.

Response:

Thanks for your valuable comments. We added a section on Discussion in the paper, discussing the direction of future model optimization. (cf. [Major change 5])

Additionally, although you have evaluated SAM-Net on diverse datasets, it would be relevant to extend testing to other meteorological phenomena, such as precipitation or strong winds. This would help assess the robustness and generalization of your model in various conditions beyond typhoons. SAM-Net's ability to adapt to other types of climate data would further enhance its utility to the scientific community.

Response:

Thanks for your valuable comments. Due to the difficulty in obtaining datasets for rainfall and strong winds in a short period of time, we conducted an experiment on a general cloud map of the Northwest Pacific and analyzed the experimental results to demonstrate the generalization of our model. (cf. [Major change 4])

[Comment 3]

Moreover, a more detailed analysis of prediction errors in complex cases would be beneficial. A deeper exploration of situations where the model fails to accurately predict certain irregular cloud movements could provide insights into potential biases and offer directions for further performance improvements. This could include a qualitative study of errors in specific cases, and you might also consider enhancing your discussion on this topic.

Response:

Thanks for your valuable comments. We have added a section on Discussion in our paper, discussing further directions for improving the performance of SAM-Net. In the future, we will consider adjusting the weights of the model from different stages of typhoon generation to achieve more accurate predictions of typhoon cloud images (cf. [Major change 5])

[Comment 4]

Finally, while SAM-Net appears to be effective on smaller images, you mention a degradation in results on larger-scale images. You should address this scenario and propose image processing techniques, such as as super-resolution or dimensionality reduction methods, to improve the clarity of predictions on large-scale images while minimizing information loss.

Response:

Thanks for your valuable comments. We have added a section on Discussion in the paper, where we will use loss functions or incorporate super-resolution methods into the model to make the images predicted by our model clearer. (cf. [Major change 5])

[Comment 5]

Overall, your work provides a valuable contribution to meteorological forecasting based on spatio-temporal sequences, and I encourage you to consider these suggestions to enhance the reach and accuracy of your model.

Response:

Thank you again for your affirmation of our work.

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have made some responses to my comments, but it appears that the revisions are superficial and the essential changes have not been substantially implemented, such as the responses to the original comments (3) and (4).

1. When it comes to typhoon forecasting, whether the result is good or not depends on three aspects: location, time and intensity. I still think that the various evaluation indicators mentioned by the authors may not be sufficient to evaluate the quality of typhoon forecasting.

2. The accuracy of deep learning is closely related to the amount of training data used; if the amount of data used for training is too small, the prediction accuracy is difficult to guarantee. In addition, the temporal resolution of Himawari satellite images is 10 minutes, why do the authors think that choosing 1 hour is not a further reduction in the number of datasets?

3. Line 280, the spatial resolution is 2km. According to this resolution, the spatial extent of one Himawari satellite image is about 256km*256km. The horizontal scale of the typhoon is usually between several hundred kilometres to thousands of kilometres, can one image cover the typhoon completely?

4. Figure 8, the image of the typhoon at moments t=1 and t=10 looks like it changes very little. If the time interval is 1 hour, it's incredible that the development of the typhoon doesn't change nearly as much within 10 hours.

5. There are several other minor writing errors in the article, such as line 287, line 297, etc.

Author Response

Response to Reviewers Comments

We would like to thank Reviewer and the Editors for the valuable and constructive comments on our manuscript. We especially appreciate your conformation of the reviews about our contribution. We have revised the manuscript accordingly. We have submitted both the revised manuscript and the “Summary of changes” for your reference. If any further information is required, please feel free to contact us.

Summary of Changes

According to the reviewers’ comments, we have made two major improvements in the revised version:

[Major Change 1]

As suggested by reviewer, we have added the Pearson spatial correlation coefficient index to calculate the correlation between the predicted cloud map and the actual typhoon cloud images.

[Major Change 2]

As suggested by reviewer, we have corrected the writing errors in the text.

Responses to Reviewer

[Comment 1]

When it comes to typhoon forecasting, whether the result is good or not depends on three aspects: location, time and intensity. I still think that the various evaluation indicators mentioned by the authors may not be sufficient to evaluate the quality of typhoon forecasting.

Response:

Thanks for your valuable comments. This study primarily focuses on forecasting future typhoon cloud maps to provide reference for downstream operations.

Therefore, it does not involve the evaluation of typhoon intensity at present. In addition, the structural similarity index SSIM cited in this paper is calculated by comparing each pixel in the predicted cloud image and the corresponding real cloud image according to time series. Therefore, the assessment of the cloud map prediction results includes considerations of both time and space. Furthermore, we have introduced the Pearson spatial correlation coefficient as an index to measure the correlation between the predicted and actual cloud images. This coefficient ranges from -1 to 1, where a value closer to 1 signifies a stronger positive correlation, a value near 0 indicates a weak or no correlation, and a value approaching -1 denotes a stronger negative correlation.(cf. [Major change 1])

[Comment 2]

The accuracy of deep learning is closely related to the amount of training data used; if the amount of data used for training is too small, the prediction accuracy is difficult to guarantee. In addition, the temporal resolution of Himawari satellite images is 10 minutes, why do the authors think that choosing 1 hour is not a further reduction in the number of datasets?

Response:

Thanks for your valuable comments. The purpose of our study is to provide some assistance to disaster prevention and mitigation departments in understanding future typhoon changes.  Selecting a dataset with a ten-minute interval could yield minimal cloud image changes, potentially rendering it less valuable for practical reference in disaster management. At the same time, the data we use is typhoon cloud map data that passes through the same area at different time periods, which can meet the feature extraction requirements for the cloud map prediction model in this area. Considering this, we have selected cloud maps with an interval of one hour. In addition, we are also applying for cloud map data with longer time spans and other areas for the optimization of the model in the next step.

[Comment 3]

Line 280, the spatial resolution is 2km. According to this resolution, the spatial extent of one Himawari satellite image is about 256km*256km. The horizontal scale of the typhoon is usually between several hundred kilometers to thousands of kilometers, can one image cover the typhoon completely?

Response:

Thanks for your valuable comments. Indeed, typhoons can have an impact range spanning several hundred kilometers. In the context of this study, which is focused on the downstream task of typhoon disaster assessment, our research specifically concentrates on forecasting cloud changes within the core region of typhoons. The current dataset provides ample coverage to effectively capture the essential structures and dynamic alterations within the typhoon's core area.

[Comment 4]

Figure 8, the image of the typhoon at moments t=1 and t=10 looks like it changes very little. If the time interval is 1 hour, it's incredible that the development of the typhoon doesn't change nearly as much within 10 hours.

Response:

Figure 8 illustrates the model's predictive capabilities across various time intervals. The velocity and intensity of typhoons fluctuate based on the developmental stage and prevailing environmental conditions. As typhoons traverse diverse terrains, they may induce rapid changes or exhibit a gradual, stable progression, influenced by a multitude of factors. The dataset we employed encompasses the full life cycles of multiple typhoons, and the randomly selected test dataset may include the types of variations mentioned earlier. Consequently, the short-term, frame-by-frame predictive outcomes for some of the test data may not manifest changes in an immediately apparent manner.

[Comment 5]

There are several other minor writing errors in the article, such as line 287, line 297, etc.

Response:

Thanks for your careful review. We have corrected the writing errors in the text. (cf. [Major change 2]).

 

 

Author Response File: Author Response.docx