Attention-Based Wildland Fire Spread Modeling Using Fire-Tracking Satellite Observations

Round 1

Reviewer 1 Report

Dear Authors,

I prepared a  review report in accordance with the proofreading criteria of the publisher.

The main conclusions of my analytical work are the following:

-          The content of the proposed paper meets the objectives set out in the special issue information letter.

-          The main strengths of this article is that the authors of this article proposed the wildland fire spread risk model. The model is an attention-based deep learning modelling approach, which can provide a foundation for further risk assessment and management applications.

-          The author’s scientific objectives resulted useful scientific achievements.

-          The references used in the main chapters are relevant and assist the reader to understand the authors proposals. The illustrations used are regular and clear.

 Based on the above, I suggest publishing this article without further review.

Author Response

Thank you very much for your constructive comments and recommendation. We appreciate the time and effort you dedicated to reviewing our manuscript.

Reviewer 2 Report

Notes on the article entitled “Attention-based Wildland Fire Spread Modeling Using Fire tracking Satellite Observations”

1. Figure 3 – (a) Highlight the border that was added in some graphic. (b) Highlight the  line on the graph of ;

2. Line 164 – The equation  does not make sense because in this step has not yet been obtained and will only be reconstructed by the following equation . Suggestion to put only  without the rest of the equation;

3. Lines 111 and 119 - Improve the description of the fuel data (enumerating) as it is confusing the reader;

4. Line 183 – I indicate improving the mathematical notation, defining the variables before and writing the equation after. Put the  outside the parenthesis. Avoid mixing mathematical symbols in the text, for example: "dyn=fire or Weather variables..." can be rewritten as dyn means fire or Weather variables... The same must be done for  (line 184);

5. Figure 5 – The colors referring to Farsite Simulation and Barren are very close, I suggest changing one of them. It is important to clarify in the caption the meaning of the abbreviations FBFM(N);

6. Authors need to comment on the possibility of including another regularization method in addition to data augmentation. Check the possibility of performing cross-validation in k folders. These suggestions are motivated by the result of precision equal to 1 in Table 1 for CNN_NonAttn and very close to 1 for CNN_FirePolyAttn, which may be indicating that the model is overfitting;

7. I request the removal of self-citations that have no connection with the central theme of the article, such as [8] and [17];

8. Figure 2 - Even citing the source articles, it is important that the authors detail the functioning of the Channel+Fire Attention blocks;

9. It is important to make a brief explanation about Fire Attention Module and Channel Attention Module;

10. It is important that the authors detail the mathematical model used to compose the weights with the input variables;

11. What activation functions are being used in the neural networks that make up the model?

Comments for author File: Comments.pdf

Author Response

Thank you very much for your valuable comments and suggestions. Please see our point-to-point responses in the attached PDF file.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper presents an attention-based deep learning modeling approach for predicting the spread of wildfires. The authors integrate spatial and channel attention modules with a CNN architecture and train the models using a fire-tracking satellite observational dataset along with relevant environmental data. The evaluation results indicate that the attention-based models outperform other models in both next-step and recursive predictions. The integration of fire-tracking satellite observational data, along with fuel data, terrain, and weather conditions, not only strengthens the data-driven wildfire modeling approach but also enables the fusion of multi-modal data.

While, there are some minor problems should be addressed before publication.

(1) In “Model architectures”, there is no detail about the fire polygon attemtion and fire line attention. The design rationale of the model should be presented in the main body of the paper, rather than in the supplementary materials.

(2) It is necessary to demonstrate that the two attention modules indeed capture the fire perimeter and fire line. The attention map should to be supplemented.

(3) The authors divided 4,788 pairs of consecutive frames of fires from 2012 to 2019 into training and testing sets in a 9:1 ratio, without differentiating the data for each individual wildfire. Due to the continuity of wildfire spread, it would be more reasonable to include certain fire ignition data in the training set and use other fire ignition data for testing, ensuring that the training and testing sets have no overlap.

(4) TP and FN are calculated based on pixels. Please add more details about these matrix scores.

(5) In Table 1, after training the models with data augmentation, all the results except for the recursive prediction of the FireLineAttn model have deteriorated. This is contrary to expectations and requires an explanation.。

No comment

Author Response

Thank you very much for your valuable comments and suggestions. Please see our point-to-point responses in the attached PDF file.

Author Response File: Author Response.pdf