Generalizing Human-Driven Wildfire Ignition Models Across Mediterranean Regions Using Harmonized Remote-Sensing and Machine-Learning Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents a comprehensive study within the field of geomatics, addressing a relevant and timely research problem with clear implications for spatial analysis and geospatial applications. The study clearly defines its objectives and situates them well within the current state of research. The authors combine established methodologies with contemporary data sources and analytical techniques, demonstrating a solid understanding of both theoretical foundations and practical implementation. Overall, the paper is well structured, technically sound, and contributes useful insights to the existing literature.

However, some sections of the manuscript could be improved to enhance readability and strengthen the interpretation of the results. Addressing these points would improve the overall clarity and impact of the manuscript.

The abstract is informative but somewhat dense. Consider slightly shortening it by focusing more clearly on the main findings and implications, and reducing methodological detail. This would help the key contributions stand out more clearly. (Page 1).

The introduction provides sufficient background and motivation, outlining the scientific relevance of the work and identifying gaps that the study aims to address. The methodological framework is generally robust and appropriate for the research questions posed, and the data sources are suitable for the analyses conducted. The results are presented in a logical sequence and are supported by figures and tables that effectively illustrate the main findings.

The discussion successfully relates the results back to the stated objectives and to previous studies, highlighting both consistencies and differences with earlier work. The authors demonstrate awareness of the strengths and limitations of their approach, which adds credibility to the study. The conclusions are supported by the results and provide a clear summary of the main contributions of the paper.

There are, however, several areas where the manuscript could be further improved. In the second and third paragraphs of Introduction, the literature review is comprehensive but could be streamlined by reducing repetition and grouping similar studies together. The research gap and novelty could be stated more explicitly toward the end of the Introduction to clearly distinguish this work from previous studies (page 1-2).

The description of the study area and datasets is clear, but some background information (e.g., general characteristics of the data source) could be shortened, as this information is well known in the literature (pages 3–4).

While the methodology is detailed and technically robust, some sections would benefit from streamlining to improve readability. Some technical and procedural descriptions could be shortened or moved to supplementary material without compromising reproducibility (e.g., parameter settings, intermediate processing steps). You may consider moving some of these details to supplementary material to improve readability. In Section 2.3, the rationale for selecting specific model parameters or thresholds could be explained more clearly, particularly for readers less familiar with the technique (pages 4–7).

The presentation of results could be strengthened by providing more explicit interpretation in the main text. Several figures are well designed, but in some cases, figures and tables are presented with limited discussion of their broader implications. Expanding the explanation of key patterns and trends would help readers better understand the significance of the findings.

For example, in the first and second paragraphs of Section 3.2, consider adding more explanation of the observed spatial or statistical patterns and their relevance to the research questions. When presenting quantitative results, it would be helpful to briefly highlight which findings are most important (pages 7–10).

The discussion effectively relates results to previous studies; however, the limitations of the approach could be addressed more explicitly, particularly regarding data uncertainty and potential sources of error, particularly related to data quality, model assumptions, or spatial and temporal constraints.. You may consider adding a short paragraph discussing how the results might change under different assumptions or data resolutions (pages 10–12).

The conclusions are clear and appropriate, but could be strengthened by explicitly stating the main practical implications of the study and potential directions for future research (page 12).

Finally, the manuscript would benefit from minor improvements in language and clarity. Some long sentences and dense paragraphs could be revised for conciseness, and ensuring consistent use of terminology throughout the paper would improve overall readability. A slight tightening of the abstract and conclusion could also help emphasize the main findings and contributions more effectively.

In summary, this is a solid and valuable contribution to the geomatics literature. The study is methodologically sound, well motivated, and relevant to both research and applied contexts. With minor revisions focused on clarity, interpretation, and presentation, the manuscript would be well suited for publication.

Author Response

We thank Reviewer 1 for the positive assessment of the manuscript’s relevance, structure, and methodological soundness. We have carefully addressed all suggestions aimed at improving clarity, readability, and interpretation.

 

Comment: Abstract is dense and could better emphasize main findings

Response:

The abstract has been substantially tightened to reduce methodological detail and to emphasize the study’s novelty and key results, particularly the dominant role of anthropogenic drivers and the cross-regional transferability of ignition susceptibility patterns.

—> Changes made in: Abstract

 

Comment: Methodology and results would benefit from clearer interpretation

Response:

The Results and Discussion sections were expanded to provide more explicit interpretation of spatial patterns, key trends, and their relevance to the research questions.

—> Changes made in: Results; Discussion

 

Comment: Limitations and practical implications could be strengthened

Response:

A dedicated limitations subsection was added, explicitly addressing data resolution, ignition proxy uncertainty, and static predictors. Practical fire-management implications and future research directions were also expanded.

—> Changes made in: Discussion (Limitations and Implications); Conclusions

Reviewer 2 Report

Comments and Suggestions for Authors

1.  The "Ignition Point" Proxy Issue
    My biggest concern is the use of the centroid of 500m MODIS burned area pixels as a proxy for "ignition points" (Section 2.2).
    While I understand that precise ignition data is hard to obtain for all these countries, using the centroid of a burned scar introduces a significant spatial error. A fire might start at a roadside (high human impact) and burn into a forest (low human impact). By taking the center of the burned pixel, you might attribute the ignition to the forest rather than the road.
    Furthermore, MODIS MCD64A1 is known to miss small fires (under 25 hectares). Since the paper argues that human factors are dominant, missing these small, often human-caused fires in peri-urban areas creates a sampling bias.
    Suggestion: If possible, please validate your proxy against a real ignition database for one of the European regions (e.g., Spain's Egif or similar). If that is not feasible, you need to add a substantial discussion on how this spatial uncertainty and the omission of small fires might be skewing your feature importance results (specifically regarding gHM and NTL).

2.  Ambiguity in Weather Data
    In Section 2.3.3, you mention using ERA-5 Land data (Wind, Temp, RH). However, it is not clear to me if you extracted these variables “dynamically” (i.e., the specific weather at the moment/day of the `BurnDate`) or if you used “climatological averages” (e.g., mean summer temperature).
    Clarification needed: Fire ignition is driven by instantaneous weather. If you used long-term averages, your model is predicting "static fire susceptibility," not "ignition probability." Please clarify this in the methodology. If you did use static averages, the terminology in the title and throughout the paper needs to be adjusted to reflect this limitation.

3.  Sampling Strategy and High AUC
    The AUC scores reported (>0.95) are exceptionally high. In my experience with wildfire modeling, such high scores often indicate that the "background" (non-ignition) samples were too easy to classify. For instance, if the background points were placed in water bodies, bare rocks, or dense urban centers, the model is simply learning to distinguish "fuel" from "no fuel," rather than "ignition risk" within fuel beds.
    Suggestion: Please confirm that your background samples were restricted to burnable vegetation. This is crucial for the validity of the risk maps.

Comments on the Quality of English Language

Typos: The manuscript appears to have been finalized in a bit of a rush, as there are several typos. For example: "Extreem" (Abstract), "indictor" (Sec 2.3.2), "climatique" (Sec 2.3.1), and "desinty" (Sec 2.4.3). Please proofread carefully.
Figures: In Figure 1, the coordinate labels and scale bars are too small to be read comfortably. Similarly, the text inside the ROC curves (Figures 4 & 5) is quite blurry.
Variable Names: There is a small inconsistency in variable naming (e.g., "NTL_1Log" vs "NTL_log1p") in the discussion section.
Population Data: You are using 2020 population data for a 2013-2021 study period. I think this is acceptable, but it would be good to add a brief sentence justifying why this static map is sufficient (e.g., assuming slow urban growth in the specific AOIs).

Author Response

We thank Reviewer 2 for the detailed and technically important comments, which substantially improved the methodological clarity and conceptual framing of the study.

 

Comment: MODIS centroid is an imprecise ignition proxy; small fires are omitted

Response:

Ignition locations are now explicitly defined as ignition proxies derived from MODIS burned-area centroids. We added a detailed discussion of spatial uncertainty (including edge effects and potential misattribution near anthropogenic features) and explicitly acknowledge the omission of small fires (<25 ha), noting that this likely leads to conservative estimates of anthropogenic influence.

—> Changes made in: Methods (Ignition and Background Samples); Discussion (Limitations)

 

Comment: Ambiguity in temporal treatment of weather variables

Response:

We clarified that ERA5-Land variables represent long-term seasonal averages rather than instantaneous conditions. Accordingly, the model is now explicitly framed as estimating long-term ignition susceptibility rather than event-specific ignition probability. Terminology was harmonized throughout the manuscript.

—> Changes made in: Methods (Weather Variables); Abstract; Discussion; Conclusions

 

Comment: Very high AUC values raise concerns about sampling strategy

Response:

We explicitly state that background samples were restricted to burnable land-cover classes, excluding non-burnable surfaces. In the Results section, we clarify that high AUC values therefore reflect discrimination within vegetated landscapes rather than trivial fuel versus non-fuel separation.

—>Changes made in: Methods (Ignition and Background Samples); Results (Model Performance)

 

Comment: Language issues, typos, and figure readability

 

Response:

The manuscript was thoroughly proofread to correct typographical errors and harmonize variable naming. Figures were revised to improve label and text readability.

—> Changes made in: Entire manuscript; Figures

Reviewer 3 Report

Comments and Suggestions for Authors

This study focuses on the Mediterranean, a global wildfire hotspot, exploring the influence of anthropogenic and natural factors on wildfire ignition risk. The topic selection has clear practical significance. The authors integrated multi-source remote sensing data and employed machine learning methods such as Random Forest and XGBoost to develop an open-source workflow based on Google Earth Engine. The overall predictive performance of the models is good, and this framework provides an operational technical pathway for cross-regional wildfire risk comparison.

1. (Page 3, Lines 99-100) The study selected only one specific area of interest (AOI) as representative for each country (Spain, Italy, France, Morocco). This selective sampling may lead to significant sample bias, neglecting the heterogeneity within different bioclimatic zones inside each country. This is not conducive to supporting the conclusions of "cross-Mediterranean regional comparison" and "the model possesses generalization capability."

2. (Page 4, Lines 132-137) The authors directly defined the centroid of burned-area pixels in the MODIS MCD64A1 burned area product as the "ignition location," which is not sufficiently precise. The 500-meter resolution MODIS pixel records the first detection date of a burned area, and its geometric center may not be the actual ignition point. Fires may start and spread from the edge of a pixel. Furthermore, for large or rapidly spreading fires, the location of the first burned pixel may be far from the actual source of human activity (e.g., roads, agricultural land). This methodological simplification introduces significant localization errors, directly affecting model accuracy.

3. (Page 11, Line 306-307)
   On page 11, Table 9: The Cohen’s Kappa value for Random Forest is reported as "0.780," while for XGBoost it is "0.820 (approx.)". This "(approx.)" notation is unprofessional, and the basis for calculation is not explained. Based on other metrics provided in the context, a precise κ value should be calculable. This notation compromises the rigor of the results.

4. (Page 11, Lines 308-321) Here, NDVI and slope are considered key drivers, but the discussion is limited to the phrase "plays a key role," lacking analysis of their specific mechanisms of action and interactions with anthropogenic factors.

5. (Pages 9-11) A core objective of the study is to evaluate the model's generalization capability in a "cross-country" setting. However, the results section only reports the overall model performance trained on data pooled together from all regions. It is not revealed how the model performed when trained on data from Morocco and tested on data from France, for example. The authors mention "strong cross-country generalization (mean transfer AUC ≈ 0.85)" in the abstract and conclusion, but the main results section completely lacks presentation, tables, or analysis of the specific data from these transfer validation tests.

Author Response

We thank Reviewer 3 for the careful review and for raising important points regarding spatial representativeness, reporting rigor, and cross-country validation. These comments substantially improved the robustness and clarity of the manuscript.

 

Comment: Use of a single AOI per country may introduce sample bias and weakens cross-Mediterranean generalization claims

Response:

We acknowledge this limitation and have revised the manuscript to clearly frame each Area of Interest as a representative test site rather than a nationally exhaustive sample. The Study Area section now explicitly states that the selected AOIs do not capture the full heterogeneity within each country. Accordingly, claims of cross-Mediterranean generalization were tempered throughout the manuscript to reflect partial transferability across selected regions rather than basin-wide applicability.

—> Changes made in: Study Areas; Discussion

 

Comment 2: Cross-country transfer performance is mentioned but not presented in the Results section

Response:

To address this concern, we added a dedicated Results subsection describing the cross-country transfer evaluation using a leave-one-country-out (LOCO) validation strategy. In this approach, models were trained on data from three countries and evaluated on the held-out country. A summary table reporting the mean transfer performance (AUC = 0.85) is now included in the Results section, providing explicit evidence of model transferability.

—> Changes made in: Results (Cross-Country Transferability); Table added

 

Comment: Variable importance results lack mechanistic interpretation

Response:

The Discussion section was substantially revised to provide mechanistic explanations for the role of key predictors, including NDVI, slope, and anthropogenic proxies. We explain how vegetation condition, accessibility, and human activity interact to shape ignition susceptibility, while avoiding causal overinterpretation of machine-learning outputs.

—> Changes made in: Discussion (Interpretation of Drivers)

 

Comment: Imprecise reporting of Cohen’s Kappa values

Response:

All performance metrics are now reported precisely. The previously imprecise notation for Cohen’s Kappa was corrected, and a brief description of its calculation was added to improve transparency and rigor.

—> Changes made in: Results (Model Performance)

Reviewer 4 Report

Comments and Suggestions for Authors

This study develops a reproducible modeling framework for human-driven wildfire ignition risk across Mediterranean regions and demonstrates clear publication potential. However, further improvements are needed in clearly articulating methodological novelty, justifying the chosen spatial scale, and defining the applicability boundaries of the results.

(1) The proposed framework mainly integrates established data sources and commonly used machine-learning models; the core added contribution relative to existing wildfire ignition studies should be more clearly articulated.

(2) The use of 500 m spatial resolution and MODIS burned-area products as ignition proxies may smooth local ignition processes, and the potential impact of this choice on modeling results warrants further discussion.

(3) Strong correlations among night-time lights, population density, and the human modification index may affect the interpretation of variable importance; this limitation should be explicitly acknowledged.

(4) The cross-country transfer analysis is based on a limited number of study areas; therefore, conclusions regarding generalization across the entire Mediterranean Basin should be stated with caution.

(5) The discussion primarily emphasizes model performance metrics, while further elaboration on variable mechanisms and their implications for fire management would strengthen the manuscript.

Author Response

We thank Reviewer 4 for the constructive and insightful comments highlighting the need to clarify the methodological contribution, applicability boundaries, and interpretation of model outputs. These suggestions substantially improved the framing and interpretability of the manuscript.

 

Comment 1: Methodological novelty and contribution are not sufficiently articulated

 

Response:

We clarified that the contribution of this study lies not in algorithmic novelty, but in the harmonized cross-regional integration of multi-source data, the explicit assessment of cross-country model transferability, and the scalable implementation of the workflow using Google Earth Engine. The manuscript now clearly distinguishes between methodological innovation and implementation value.

 

—> Changes made in: Abstract; Discussion

 

Comment 2: Use of 500 m MODIS data may smooth local ignition processes and affect interpretation

 

Response:

We explicitly acknowledge the limitations associated with using MODIS burned-area products at 500 m resolution, including spatial smoothing and uncertainty in ignition location. The potential impact of this choice on modeling results and variable interpretation is now discussed in detail, and conclusions are framed accordingly.

 

—> Changes made in: Methods (Ignition and Background Samples); Discussion (Limitations)

 

Comment 3: Strong correlations among anthropogenic predictors complicate variable-importance interpretation

 

Response:

We explicitly address multicollinearity among night-time lights, population density, and the Global Human Modification index. These variables are now interpreted collectively as proxies for human pressure rather than as independent causal drivers, and this limitation is clearly stated in the Discussion.

—> Changes made in: Discussion (Interpretation of Drivers)

 

Comment 4: Cross-country transfer analysis is based on limited study areas and should be interpreted cautiously

 

Response:

We revised the manuscript to emphasize that cross-country transfer results reflect partial transferability across selected Mediterranean regions rather than full basin-wide generalization. Applicability boundaries are now clearly stated, and claims were tempered accordingly.

—> Changes made in: Abstract; Discussion; Conclusions

 

Comment 5: Discussion emphasizes model performance more than variable mechanisms and management implications

 

Response:

The Discussion section was expanded to provide mechanistic interpretations of key predictors and to strengthen links between model findings and practical fire-management and policy implications, particularly regarding prevention strategies targeting human activity patterns.

—>Changes made in: Discussion (Interpretation and Implications)

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have made substantial improvements in the revised manuscript (v2), particularly with the addition of Section 3.4 (Cross-Country Transferability). By implementing the "Leave-One-Country-Out (LOCO)" validation strategy, the authors have not only addressed previous concerns regarding model generalization but have also provided solid quantitative evidence (mean AUC ≈ 0.85) to support the central claim of "Generalizing" in the title. This addition significantly enhances the scientific value of the paper.However, while the scientific content has been strengthened, the manuscript still suffers from noticeable deficiencies in copy-editing and proofreading. Several spelling errors present in the methods section of the first version remain uncorrected in the second version. To ensure publication quality, the authors must perform a thorough linguistic revision before final submission.

Specific Comments
1. Scientific Content and Discussion
Suggestion for Section 3.4 (Transferability):
The addition of the LOCO validation is excellent. However, Table 10 currently only presents a summarized mean AUC value (0.85). To increase the depth of the analysis, I suggest providing a 4x4 Transfer Matrix. It would be very insightful to see specific cross-country transfer performance: for instance, does the model transfer better between biogeographically similar countries (e.g., Spain → Morocco) compared to socially distinct ones (e.g., France → Morocco)? This would provide more granular guidance for fire risk management in the Mediterranean basin.

Dominance of the gHM Index:
The results show that the "Global Human Modification (gHM)" index is the most important predictor in both Random Forest and XGBoost models. Given that the vast majority of fires in the Mediterranean are human-caused, this result is somewhat expected. I suggest the authors discuss this further: Does gHM mask more specific drivers? Alternatively, a brief sensitivity analysis discussing model performance without gHM (relying instead on NTL or population density) would help address potential concerns about tautology in using a composite human impact index to predict human-caused fires.

2. Language and Editing (Mandatory Corrections)
There are several basic spelling and grammatical errors, particularly in the Methods section, which have persisted from the previous version. These must be corrected:
- Section 2.3.1 (Topography): "Topography influences climatique..." Please correct to the English spelling "climatic".
- Section 2.3.2 (NDVI): "NDVI was computed direclty..." Please correct to "directly".
- Section 2.3.2 (NDVI): "...haromnized dataset..." Please correct to "harmonized".
- Section 2.3.2 (NDVI): "...vegetation health and the greeness. This greness..." Please correct to "greenness" in both instances.
- Section 2.4.1 (gHM): "Various studies suggest that the ghM is considered..." Please ensure consistent capitalization as "gHM".
- Section 2.4.2 (NTL): "Studies has found..." Please correct the subject-verb agreement to "have found".

Figure Quality:
The text in Figure 2 (Correlation Matrix) appears blurry in the PDF. Please ensure high-resolution vector graphics are used for the final submission.

Comments on the Quality of English Language

Typos: The manuscript appears to have been finalized in a bit of a rush, as there are several typos. For example: "Extreem" (Abstract), "indictor" (Sec 2.3.2), "climatique" (Sec 2.3.1), and "desinty" (Sec 2.4.3). Please proofread carefully.
Figures: In Figure 1, the coordinate labels and scale bars are too small to be read comfortably. Similarly, the text inside the ROC curves (Figures 4 & 5) is quite blurry.
Variable Names: There is a small inconsistency in variable naming (e.g., "NTL_1Log" vs "NTL_log1p") in the discussion section.
Population Data: You are using 2020 population data for a 2013-2021 study period. I think this is acceptable, but it would be good to add a brief sentence justifying why this static map is sufficient (e.g., assuming slow urban growth in the specific AOIs).

Author Response

We thank the reviewer for the insightful comments and constructive suggestions. We have addressed each point in the revised manuscript as detailed below.

1. Scientific Content

a) Suggestion for Section 3.4 (Transferability)

We agree that reporting only the mean LOCO AUC masks important cross-country variability and limits interpretability for operational fire-risk management. In response, we have expanded Section 3.4 by introducing a 4×4 cross-country transfer matrix, reporting AUC values for each train–test country pair and discussing cross-regional transfer patterns.

b) Dominance of the gHM Index

We agree that the strong importance of the Global Human Modification (gHM) index is expected in a region where most ignitions are human-caused and therefore warrants careful discussion. We have expanded the Discussion to clarify that gHM is not fire-specific but represents cumulative human pressure (e.g., built-up areas, infrastructure, agriculture). We further discuss that gHM aggregates multiple correlated anthropogenic drivers and therefore acts as an integrative proxy rather than a tautological predictor of human-caused fires.

2. Language Editing

The manuscript was proofread using British English conventions. This included consistent spelling and hyphenation (e.g., modelling, night-time, behaviour, spatio-temporal), correction of typographical errors, and harmonization of terminology throughout the text.

In addition, the acronym gHM was standardized across the manuscript, and Figure~2 (correlation matrix) was replaced with a higher-quality version (300 dpi).

Reviewer 4 Report

Comments and Suggestions for Authors

I agree to accept this version of the manuscript.

Author Response

N/A