Assessing Fire Risk Zones in Phrae Province, Northern Thailand, Using a MaxEnt Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents an application of the MaxEnt model for fire risk mapping, but it suffers from significant technical weaknesses in data preparation, model validation, variable justification, and reproducibility. Critical steps in the MaxEnt pipeline are either omitted or ambiguously described. The lack of robust validation, incomplete methodological transparency, and unclear handling of model uncertainty raise serious concerns about the scientific rigor of the study. Substantial revision is required to meet peer-reviewed publication standards. Following are the specific comments:

1. The study fails to explicitly justify the selection of only seven environmental variables; critical drivers such as wind speed, land use types, and anthropogenic ignition sources are omitted without explanation.
2. The NDVI dataset lacks temporal context. Forest fire susceptibility varies seasonally, yet the study does not clarify whether the NDVI used represents an annual average, seasonal composite, or a specific time frame, which significantly undermines model interpretability.
3. DEM, slope, and aspect are all derived from the same elevation data source, leading to collinearity. The authors fail to address or test multicollinearity among predictors, violating a core assumption of the MaxEnt algorithm’s robustness.
4. The inclusion of 300 "random additional hotspots" is described ambiguously. It's unclear whether these points are synthetic duplicates, spatially smoothed interpolations, or bootstrapped resamples. Artificial replication of presence data can bias model performance by inflating spatial autocorrelation.
5. The study does not implement or report spatial filtering to address sampling bias in presence-only data. This omission violates best practices for MaxEnt modeling and likely results in spatial overfitting.
6. The use of AUC as the sole model performance metric is insufficient. No true absence data are used, making AUC a limited and possibly misleading indicator. Other metrics such as TSS, sensitivity, or Boyce Index should be included.
7. The model validation approach is flawed; a single AUC value of 0.849 is reported without confidence intervals or cross-validation. This hinders assessment of model robustness and generalizability.
8. The statement that the model is “completely reproducible” due to use of a random seed does not account for inherent spatial biases or sampling variance across geographic extents.
9. Although the authors mention jackknife analysis, they present only a summary figure without tabulated permutation importance, percent contribution, or response curves - preventing replication or critical evaluation of variable influence.
10. Response curves are claimed to be complex and nonlinear, but no actual curves or functional forms are shown, depriving the reader of critical insights into ecological interpretation.
11. Reclassification of the MaxEnt output raster into five classes using equal interval bins (0.0–0.2, ..., 0.8–1.0) is arbitrary and not ecologically justified. Alternative thresholds based on statistical optimization or expert knowledge are not considered or tested.
12. The manuscript states that “final predictions are also obscured in flammable areas of the forest” without specifying what “obscured” means—whether this refers to masking, uncertainty weighting, or exclusion. The lack of detail makes this an uninterpretable and methodologically opaque decision.
13. Descriptions of MaxEnt functionality are lifted almost verbatim from foundational literature (e.g., Phillips et al.) without proper contextualization for the current study’s limitations, which borders on formulaic usage.
14. The fire risk mapping (Table 1 and Figure 5) does not display spatial uncertainty, error margins, or validation with independent fire records. The mapping is presented as deterministic, which is inappropriate for a probabilistic model.
15. No comparison with other modeling methods (e.g., Random Forest, Logistic Regression) is made, which is essential for justifying the exclusive use of MaxEnt in a high-dimensional geographic modeling context.
16. Fire hotspot data is stated to come from a 10-year database, yet the authors do not clarify spatial resolution, detection sensitivity, or whether MODIS or VIIRS data is used. These details are critical given the known biases in thermal anomaly detection in forested terrain.
17. The distance layers (to roads, water, communities) are poorly described: interpolation methodology, buffering thresholds, or spatial resolution are not defined, leaving substantial ambiguity in data preparation.
18. The reported area percentages of each risk class (e.g., 52.99%, 13.68%) are overly precise and statistically misleading, especially when derived from coarse raster reclassification.
19. District-level risk tabulation (Table 2) is descriptive but uncorrected for district size. The table should express risk proportionally (e.g., percentage of district area in “high risk” category) to allow meaningful spatial comparisons.
20. Several contradictory and redundant sentences exist throughout the manuscript, such as “destroy forest areas but also destroy forest areas” in the introduction. These affect clarity and reflect poor proofreading.
21. Several references cited in support of methodological decisions are general forest fire literature, not MaxEnt-specific or geographically relevant to Southeast Asia. Literature synthesis lacks critical specificity.
22. Environmental drivers are ranked by “importance” (e.g., pDEM = 81%), but the unit or method of contribution estimation (e.g., gain, permutation importance) is not disclosed, rendering the interpretation of those values ungrounded.
23. Claims such as “this analysis is extremely extensive and accurate” in the conclusion are unsupported by rigorous sensitivity analysis, independent validation, or comparative benchmarks.
24. The study uses a 12.5m resolution DEM, which is exceptionally fine for regional fire modeling. No computational or resampling details are given, and this likely introduces overfitting in the presence-only MaxEnt context.
25. The manuscript includes verbose, overly didactic sections (e.g., MaxEnt model theory) while omitting essential implementation details (e.g., number of background points, regularization multiplier settings), reflecting an imbalance in technical depth.
26. Textual inconsistencies such as “Logistics Criteria of the 10th percentile training state” vs “logistical requirement for the training state's sensitivity” confuse the reader and reflect unclear understanding of model thresholds.
27. The authors assert “ecological gap extends beyond the geographical bounds” but fail to conduct or present any transferability test or extrapolation diagnostic, contradicting their claim of general model performance.
28. The use of public GIS layers (e.g., roads, communities) without detailed citation, preprocessing description, or error analysis introduces untraceable uncertainties in spatial predictors.
29. The manuscript lacks any visualization of variable contribution or response across space (e.g., variable influence maps), making it impossible to discern where and how individual drivers manifest in the landscape.
30. Figure captions (e.g., Figure 5) are minimal and non-informative (“This is a figure. Schemes follow the same formatting”), suggesting an incomplete or automated document assembly.
31. The GIS-based data preparation (GRID format) is not reproducibly described - software used, projections, masking steps, and spatial alignment procedures are omitted.
32. Ethical or practical limitations of using remote sensing to infer fire risk in inhabited or indigenous areas are not acknowledged. This omission is particularly problematic given Thailand’s diverse socio-ecological contexts.
33. No attempt is made to ground-truth or validate model outputs using independent fire datasets, local knowledge, or historical records, severely limiting credibility.
34. Multiple grammar and syntax issues, such as “burning to prepare agricultural land, and intentionally burning forests make it critical,” create semantic ambiguity and hinder scientific communication.
35. The conclusion section restates results without critical reflection. No acknowledgement of methodological limitations or model uncertainties is made, which weakens scientific integrity.
36. No code, workflows, or model configuration files are made available, despite using publicly accessible datasets and open-source MaxEnt software - this severely limits reproducibility.

Author Response

the points to points response is attached below.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

As interesting as this paper is, and it currency given wildfires globally, the approach is quite basic and pedestrian and in fcat mapping and GIS are not new to this field. The combination of approaches may be useful in the context of GIS. There is little or no reference to existing fire modelling approaches and software and the literature review is really quite narrow given the work already done in this field. The English also needs a fair amount of work through careful proof-reading, editing etc. In fact, much of the paper could do with a re-write as there seem to be alot of sentences that start and go nowhere in the paragraphs. Many paragraphs are very short and say little especially in the methodology section. Maps are inconsistent in content and layout. What about some photographs of the study site? What about flow diagram? What GIS was used? ArcGIS I guess from the map style. The analysis and interpretation of the study data and method is weak and the results look like a dissertation study? Some captions are strange? Hot spots > better refer to as fires ....and so on. Interesting but this needs work before it can get to the stage of publication. It needs more detail, more illustrations, a wider literature, rationale..... as well as proof-reading, editing etc.... I also notice the 22% Similarity Index which needs to be reduced to to 2-3%.

Comments on the Quality of English Language

This paper needs careful editing and proof-reading as well as a rewrite in places as it does not read well in many places.

Author Response

the points to points response is attached below.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors state that this study aims to investigate the physical factors that influence wildfire occurrence and create a fire risk map, applying Remote Sensing and Geographic Information System (GIS) technology for analysis. Seven factors were considered: Digital Elevation Model (DEM), slope, Normalized Difference Vegetation Index (NDVI), and aspect. Distance from people, water, and roads are examples of geographic factors that can affect wildfires. Importantly, they used the MaxEnt (Maximum Entropy) model with an AUC (Area Under the Curve) of 0.849. The findings revealed that variables that influence wildfire incidence include Digital Elevation Model (DEM), Normalized Difference Vegetation Index (NDVI), slope, distance from roads, distance from water, distance from communities, and aspect. This work has been replicated in different parts of the world; however, it represents an effort in the search for elements that allow us to mitigate the incidence of fires that have devastating effects on humans and ecosystems.
To improve the publishable version of this work, we need to make some important adjustments:
The Introduction
This needs to be reformulated and completed; the state-of-the-art of knowledge is not thoroughly established.
The advantages of using a GIS in this type of work need to be established.
The research questions need to be clearly stated.
Materials and Methodology
Figure 1, the study area, requires larger elements that allow for proper location, showing at least a continental location.
It would be advisable to include a climate classification diagram for the study area.
A schematic of the methodology used is required; this is essential to identify the procedure used. The use of a methodological schematic is essential to establish the working mechanism. This, in turn, allows this work to be replicated by other researchers or to change and/or improve the procedures used.
Results and Discussion
The discussion regarding the validation of the results needs to be improved. Use more impactful references.
The findings revealed that the variables that influence the incidence of wildfires include: Digital Elevation Model (DEM), Normalized Difference Vegetation Index (NDVI), slope, distance from roads, distance from water, distance from the community, and orientation. In this sense, it is important to validate this finding bibliographically. This finding must be consolidated and use bibliographic references to help validate it.
Formal Aspects
Substantially improve the resolution of the figures.
Improve the description of the figures; for example, Figure 4 requires more explanation in its descriptive section, included at the bottom of the page.
Indicate the licenses for the software used, especially GIS.

Author Response

the points to points response is attached below

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript shows progress, but several critical issues remain unresolved and need further revision before publication. DEM, slope, and aspect are included without tests for redundancy; a correlation matrix or Variance Inflation Factor (VIF) analysis should be reported to confirm that collinearity does not bias results. The study still relies mainly on AUC for model evaluation; although TSS is mentioned, no results are presented. Sensitivity, specificity, or Boyce Index should also be reported to give a more balanced assessment of performance. Only a single AUC value is provided, with no confidence intervals or variability estimates. K-fold cross-validation or bootstrapping would strengthen claims of robustness.

The risk classification remains unclear, with inconsistency between equal-interval binning and the 10th-percentile logistic threshold. The manuscript must clarify which method was used and justify its ecological or statistical basis. In the conclusions, terms such as “extremely extensive and accurate” are overstated and should be moderated, with explicit acknowledgment of methodological limitations. No independent validation is attempted with external fire records or expert knowledge, which limits confidence in the model’s generalizability. Finally, while data sources are now described, no code, configuration files, or input datasets are provided. Supplementary material would greatly enhance reproducibility and align the work with open-science standards.

Author Response

Comments 1: The revised manuscript shows progress, but several critical issues remain unresolved and need further revision before publication. DEM, slope, and aspect are included without tests for redundancy; a correlation matrix or Variance Inflation Factor (VIF) analysis should be reported to confirm that collinearity does not bias results. The study still relies mainly on AUC for model evaluation; although TSS is mentioned, no results are presented. Sensitivity, specificity, or Boyce Index should also be reported to give a more balanced assessment of performance. Only a single AUC value is provided, with no confidence intervals or variability estimates. K-fold cross-validation or bootstrapping would strengthen claims of robustness.

The risk classification remains unclear, with inconsistency between equal-interval binning and the 10th-percentile logistic threshold. The manuscript must clarify which method was used and justify its ecological or statistical basis. In the conclusions, terms such as “extremely extensive and accurate” are overstated and should be moderated, with explicit acknowledgment of methodological limitations. No independent validation is attempted with external fire records or expert knowledge, which limits confidence in the model’s generalizability. Finally, while data sources are now described, no code, configuration files, or input datasets are provided. Supplementary material would greatly enhance reproducibility and align the work with open-science standards.

Response 1: Thank you for pointing this out. We agree with this comment.

• Collinearity Analysis:Variance Inflation Factor (VIF) analysis. Edited in section 2.2.2, page 4, lines 120-121.
• Model Evaluation:The study still relies mainly on AUC for model. "The use of AUC as the primary evaluation metric is supported by several studies, such as (Fielding & Bell, 1997), which highlighted AUC as a reliable metric for evaluating imbalanced classification models like MaxEnt.”
• Risk Classification Inconsistency: "Using the logistic threshold method at the 10th-percentile training state is a widely accepted approach in several studies, such as (Phillips et al., 2008), which showed that this method helps reduce errors in classifying risk areas. Additionally, this approach has proven to be highly accurate in distinguishing risk zones in model outcomes." Edited in section 2.3.1, page 6, lines 197-200.
• Conclusions Overstatement: We have made the necessary adjustments. Edited in section 4, page 12, lines 335-336.
• Independent Validation: "Using external data to validate model accuracy is recommended in (Swets, 1988), which noted that this method is essential for assessing the robustness and accuracy of models across different environments. While our current work does not include external validation, we will consider this approach in future research once suitable data becomes available."
• Reproducibility and Open Science: The website provided can search for files. The website provided allows you to search for files, and each website has instructions for downloading the files, but most are in Thai, so you can't insert them into code.

Reviewer 2 Report

Comments and Suggestions for Authors

My main feeling is that the English has been addressed and the authors have made some effort to do this. However, I still think that some other aspects require a little more input to provide the context to this 'experiment' (see below).

What about some photographs of the study site?  What about flow diagram? What GIS was used? ArcGIS I guess from the map style. The analysis and  interpretation of the study data and method is weak and the results look like a dissertation study? Some captions are strange? Hot spots > better refer to as fires ....and so on. Interesting but this needs work  before it can get to the stage of publication. It needs more detail, more illustrations, a wider literature, rationale..... as well as proof-reading, editing etc.... I also notice the 22% Similarity Index which needs to be reduced to to 2-3%. ..... not all addressed in revised version.

Furthermore there are some loose statements such as this one: The area with the lowest risk was Nong Muang Khai District. ....... what is the relevance of this? I also think that whilst the method is well documented, the interpretation is a little limited.... for example: This study found that geographical factors were the most important environmental components for forest fires, .... is this not a little self evident..... why? explain....

Author Response

Comments 1: My main feeling is that the English has been addressed and the authors have made some effort to do this. However, I still think that some other aspects require a little more input to provide the context to this 'experiment' (see below).

What about some photographs of the study site?  What about flow diagram? What GIS was used? ArcGIS I guess from the map style. The analysis and  interpretation of the study data and method is weak and the results look like a dissertation study? Some captions are strange? Hot spots > better refer to as fires ....and so on. Interesting but this needs work  before it can get to the stage of publication. It needs more detail, more illustrations, a wider literature, rationale..... as well as proof-reading, editing etc.... I also notice the 22% Similarity Index which needs to be reduced to to 2-3%. ..... not all addressed in revised version.

Furthermore there are some loose statements such as this one: The area with the lowest risk was Nong Muang Khai District. ....... what is the relevance of this? I also think that whilst the method is well documented, the interpretation is a little limited.... for example: This study found that geographical factors were the most important environmental components for forest fires, .... is this not a little self evident..... why? explain....

Response 1: Thank you for pointing this out. We agree with this comment.

• Photographs of the Study Site:"Thank you for the suggestion to include photographs of the study site. We acknowledge that visual context is essential for understanding the geographical characteristics of Phrae Province. We will include high-resolution photographs. Edited in Figure 1., page 3, lines 96.
• GIS Software As specified in Section 2.2.6, page 5, lines 151.
• Clarification of "Hot Spots":"Thank you for pointing out the use of the term 'hot spots.' We agree that referring to these as 'fires' would be clearer and more accurate. We will replace 'hot spots' with 'fires' throughout the manuscript." Edited in section 2.2.4, page 4, lines 139-242.
• Relevance of Nong Muang Khai District:"We understand your concern about the mention of Nong Muang Khai District. The reference to Nong Muang Khai was made to highlight the district with the lowest fire risk, which serves as a contrast to the high-risk areas and provides context for our risk classification. In the revised version, we will clarify the relevance of this district by explaining that it represents an outlier with minimal fire risk, reinforcing the overall findings about the varying levels of risk across Phrae Province." Edited in section 3.4, page 11, lines 326-327.
• Similarity Index:"We have noted the concern regarding the 22% Similarity Index. We understand the importance of reducing this to ensure originality, and we are in the process of revising the manuscript by paraphrasing sections that may have contributed to this high similarity. We will aim to reduce the index to 2-3% in the final version."

Reviewer 3 Report

Comments and Suggestions for Authors

I consider the improvement work carried out by the authors acceptable, its publication is possible, after some revisions so that the format reviewer notices.

Author Response

Comments 1: I consider the improvement work carried out by the authors acceptable, its publication is possible, after some revisions so that the format reviewer notices.

Response 1: Thank you for pointing this out. We agree with this comment. Therefore, We have completed the necessary corrections.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

Figures should be checked for consistent font size and clear labeling to ensure they are easily interpretable. A brief note on the potential application of this model for other provinces or cross-border fire management would further strengthen the conclusion. Article can be accepted after these revisions.

Author Response

Comments 1: Figures should be checked for consistent font size and clear labeling to ensure they are easily interpretable. A brief note on the potential application of this model for other provinces or cross-border fire management would further strengthen the conclusion. Article can be accepted after these revisions.

Response 1: Thank you for pointing this out. We agree with this comment. 

• We examined and changed the font size and labeling in all figures to ensure consistency and clarity throughout the book. Each figure now has a consistent style for easier comprehension and understanding.
• We have included a brief note in the conclusion to address the model's possible application to other provinces and cross-border fire management. The MaxEnt model's versatility allows for adaption to different places with similar climatic conditions, and it could be beneficial for assessing wildfire risk in surrounding areas with similar topography and vegetation types.Edited in section 4, page 12, lines 360-362.

Reviewer 2 Report

Comments and Suggestions for Authors

You have improved the manuscript, which is good, but in the revised version I do not see the photographs, what is the Similarity Index now? Also I do not see a response to:  I also think that whilst the method is well documented, the interpretation is a little limited.... 

What are the findings and what can you interpret from these?

Author Response

Comments 1: 

You have improved the manuscript, which is good, but in the revised version I do not see the photographs, what is the Similarity Index now? Also I do not see a response to:  I also think that whilst the method is well documented, the interpretation is a little limited.... 

What are the findings and what can you interpret from these?

Response 1: Thank you for pointing this out. We agree with this comment.

• The changed figure remains the same because the research area is in the central part of the northern region, which the supplied boundary clearly covers, so the image from the article is used.
• The similarity index ​​were examined by the journal. The authors have carefully written the article to minimize the value of the article.
• This new edition includes more information about how to use the MaxEnt model to assess wildfire risk. To improve comprehension, the findings are related to other study topics in Section 4. Section 4, page 12, lines 360–362.