Forest Fire Prediction: A Spatial Machine Learning and Neural Network Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Please find the comments attached.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Author Response

Dear Reviewer,

 

I hope you are doing well!

 

Thank you for the review! I really appreciate the way you help me to increase the efficacy of my work. Following are the response to your review,

 

1 ans) Yes, all factors were considered as a model building, but some were disregarded as they were highly correlated to each other based upon heat map.

 

2 Ans) Yes, based on their relation to the fire occurrence whether the correlation coefficient is very high or very low they are considered as the critical factors.

 

3 Ans) Yes I compared the work with other authors who utilized some machine learning or neural approach here, I included most of them to see the effect.

 

4 Ans) We need a coefficient to do the weighted overlay. In linear regression we can directly get the coefficient but in machine learning or neural network each independent variable gave their own importance after we run the model and were used to do the raster overlay to predict the model.

 

5 Ans) I generated the ROC and AUC later to see the impact. 

 

6 Ans) I did the transformation on the earth engine is inbuilt as a WGS 84. The missing data were handle by removing the row which have missing data.

 

7 ans) The feature importance was not predicting better as compared to the kernel density map, hence correlation coefficients are predicting better based upon the kernel density map and the hazard map obtained by the correlation coefficients.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Is the period at the end of the title necessary?

 

Are Machine Learning and Deep Learning in the title overlapping?

 

Please explain more technical details on limitation of this work and on future work.

 

Clarify originality and advantages against state of art.

 

The paper heads and footers used 2022, why ?

 

Lack 2024, 2023, 2022 references. There is only one 2021 reference item. Please discuss more relevant recent work, such as:

 

 

A Hierarchical Method for Locating the Interferometric Fringes of Celestial Sources in the Visibility Data

Research in Astronomy and Astrophysics

 

Improved Density Peaking Algorithm for Community Detection Based on Graph Representation Learning.

Computer Systems Science & Engineering 43 (3)

 

 

 

 

Explain in good technical details on complexity of the solutions.

Explain the validity and generalisability of experimental results in more technical details.

Need formal proof of claims and more rigour analysis with better technical depth.

Rational choice of evaluation criteria.

Comments on the Quality of English Language

Please check author guide and journal paper samples and proofread the paper to check English and format issues.

Author Response

Dear reviewer,

Thank you very much for your reply. 

I fixed all the issue based upon your comments. I added more validation approach and more accuracy approach based upon the long data from 2000 to 2023. I gave all effort on clarity and better result interpretation. 

Reviewer 3 Report

Comments and Suggestions for Authors

• The paper explores the potential of machine learning techniques, which is a promising 
• Utilizing data from the South Carolina Forestry Commission strengthens the study's foundation.
• In page 5, Table 1: Different sites and methods used for the data collection, text contents are mixed with the table.
•  
•  
• The research utilizes data from 2023, explore the model's performance with data from different years to assess its generalizability over time.
• Mention the model to be adapted for use in other regions with similar fire risk factors.
• Justify whether were there specific factors that led the decision tree to outperform others.
• Highlight the novelty of the proposed method.
•  

Author Response

Dear reviewer,

 

Thank you for your review. I highly appreciate it. I collected the data from South Carolina forestry commission from 2000 to 2023 and then compared the result obtained my all approach then found out the correlation coefficients provided the better result compared to any other approach. I incorporated all approach and for the state of South Carolina the correlation coefficient has the maximum capability to predict compared to feature importance. The multilayer preceptor outperform the other model as it depends on the correlation coefficient.

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript describes the application of decision tree (DT), random forest (RF), logistic regression (LR), artificial neural network (ANN), support vector machine (SVM), and convolutional neural network (CNN) in forest fire susceptibility modeling and analysis. Despite the importance of the topic, I have reservations about the suitability of the manuscript for publication in a high-quality journal like Fire. Here are my concerns that lead me to recommend rejecting this manuscript for publication:

Lack of novelty: The novelty of this study is my primary concern. What new aspect does this study offer in terms of “machine learning modeling of forest fire susceptibility”? What distinguishes it from existing research in this domain? A quick Google Scholar search reveals numerous similar studies sharing common content. The use of these methods seems to be a routine exercise in this field. The manuscript would be more acceptable if it had considered another route for modeling, such as investigating new machine learning methods or optimizing and improving existing methods/models using Artificial Intelligence.

Selection of input factors: The manuscript selects 16 specific factors from South Carolina as input data for the models. What criteria were used in this selection process, and why were these factors chosen over others? While these factors seem to be drawn from existing literature and empirical experience, the discussion on their significance and contribution to forest fire occurrence in South Carolina is lacking scientific depth.

Evaluation method: The authors skipped from describing the evaluation/validation procedure of the models. While the evaluation of models’ performance appears to rely solely on an accuracy analysis, I believe that this evaluation approach is inadequate and that a more comprehensive assessment (e.g., AUC, sensitivity, and specificity) is required.

Results: The results are trivial and do not offer anything new. Additionally, this section introduces new information (Carbon hotspot) that was not previously described in the methodology section. The poor quality of maps and figures further indicates a lack of careful manuscript preparation.

Discussion: The discussion section appears to have several issues that need addressing. Currently, it reads more like an introduction, primarily focusing on the literature review and problem description. In a robust discussion, it is essential to engage with interpreting the results obtained in your study. What do these results signify, and how do they contribute to the existing body of knowledge? How might the results be applied in real-world scenarios? What were the advantages and disadvantages of your methodology? While the discussion should not merely reiterate the literature review, it should involve a critical engagement with existing research. How do your findings align with or deviate from established knowledge? Most importantly, what does it add to the field, and why should the reader care about the results presented?

Comments on the Quality of English Language

For the current text, minor editing of English language is required.

Author Response

Dear reviewer,

 

Thank you very much for the comments. In this article I Collected the almost all approaches to predict the forest fire. The feature importance was utilized on many other region however in the state of South Carolina it is not generating the goof hazard map compared to the data from South Carolina forestry commission from 2000 to 2030. The kernel density map and the map generated from the multilayer preceptor are very similar compared to all other models which utilized the feature importance. Hence, this study identifies the capability of correlation coefficients in predicting the forest fire then the feature importance though the accuracy was high. In the result I put the AUC and ROC curve to validate the result. 

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

Authors addressed all my comments and suggestions.

Author Response

Dear Reviewer,

 

I hope you are doing well!

 

Thank you for your time and comments!

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

I cannot see any significant improvements in this version of the manuscript. The authors have resubmitted the original submission with only a few revisions. Importantly, they have not marked, highlighted, or tracked changes in the manuscript and did not provide any rebuttal to the reviewer’s comments. I am afraid that I must maintain my initial recommendation to reject this manuscript for publication.

Comments on the Quality of English Language

For the current text, minor editing of the English language is required.

 

Author Response

Dear Reviewer,

 

I hope you are doing well!

 

• Lack of novelty: The novelty of this study is my primary concern. What new aspect does this study offer in terms of “machine learning modeling of forest fire susceptibility”? What distinguishes it from existing research in this domain? A quick Google Scholar search reveals numerous similar studies sharing common content. The use of these methods seems to be a routine exercise in this field. The manuscript would be more acceptable if it had considered another route for modeling, such as investigating new machine learning methods or optimizing and improving existing methods/models using Artificial Intelligence.

Ans = This study compared the feature importance and correlation coefficients. Though the accuracy was high in the decision tree model, the feature importance is not giving the precise map while doing raster overlay. The correlation coefficient did well in precise map development. Most of the research done previously was for the small area. This research used google earth engine for the whole state. Nobody has compared the carbon hotspot with the fire location, which might be a better concept for mitigation of carbon loss due to fire. Most of the past studies which utilized feature importance did not have the long data they just predicted the map but did not compare it with the actual fire density map because of lack of data. Hence, this research distinguishes the different between the past studies and the current studies.

• Selection of input factors: The manuscript selects 16 specific factors from South Carolina as input data for the models. What criteria were used in this selection process, and why were these factors chosen over others? While these factors seem to be drawn from existing literature and empirical experience, the discussion on their significance and contribution to forest fire occurrence in South Carolina is lacking scientific depth.

Ans =All of them were selected based upon the past studies. We considered most of the factors which past studies did not. We incorporated the NDVI, and national land cover dataset which most of the past studies did not include. This variable has a major role in predicting the fire.

• Evaluation method: The authors skipped from describing the evaluation/validation procedure of the models. While the evaluation of models’ performance appears to rely solely on an accuracy analysis, I believe that this evaluation approach is inadequate and that a more comprehensive assessment (e.g., AUC, sensitivity, and specificity) is required.

We considered the ROC and AUC for the different models and later we developed the Kernel density estimation of all the fire points to identify the major areas of fire and compare the result with the output from all model and correlation coefficients.

• Results: The results are trivial and do not offer anything new. Additionally, this section introduces new information (Carbon hotspot) that was not previously described in the methodology section. The poor quality of maps and figures further indicates a lack of careful manuscript preparation.

Ans = We included the carbon data in methodology section. This result shows the different approach of comparison of the final model with the kernel density estimation and ROC. The past studies did not use kernel as they lack data. Here, I used both the approach and identify that to predict the fire we have to find other coefficient estimation technique other then feature importance based upon the final hazard map and the kernel density estimation.

• Discussion: The discussion section appears to have several issues that need addressing. Currently, it reads more like an introduction, primarily focusing on the literature review and problem description. In a robust discussion, it is essential to engage with interpreting the results obtained in your study. What do these results signify, and how do they contribute to the existing body of knowledge? How might the results be applied in real-world scenarios? What were the advantages and disadvantages of your methodology? While the discussion should not merely reiterate the literature review, it should involve a critical engagement with existing research. How do your findings align with or deviate from established knowledge? Most importantly, what does it add to the field, and why should the reader care about the results presented?

Ans = We now focused on interpreting the result. We compared our result with the result from other and we explained how it differs. The result obtained helped in reducing the use of feature importance in predicting the forest fire. The correlation coefficients though have linear relationship but help in predicting in better way as verified by the kernel density map. The result can be applied to prioritize the management to mitigate the carbon loss due to fire. The finding deviate from the existing knowledge as the feature importance is not a great way to incorporate in raster overlay to predict the hazard map. Hence, new approach should be considered to predict the fire hazard better. Even the accuracy score is high the feature importance do not give the better result.

 

Have a wonderful day!

 

Thank you!

 

 

Author Response File: Author Response.docx

Round 3

Reviewer 4 Report

Comments and Suggestions for Authors

-

Comments on the Quality of English Language

-