Overall

The paper presents a spatially explicit modeling approach for human-caused wildfire ignitions across various European regions. The study applies machine learning techniques, specifically Random Forest models, to analyze historical fire records and environmental variables such as land cover, population density, accessibility, and dead fine-fuel moisture content (DFMC). The results highlight that the most influential variables in predicting ignition probability are DFMC anomalies, proximity to the Wildland-Urban Interface (WUI), and road accessibility.

The study emphasizes the role of anthropogenic factors in fire ignition and provides valuable insights into human-caused wildfire ignitions. However, it faces challenges related to model generalizability, temporal dynamics, and policy application.

Major comments

- One major comment regarding this paper is that a big portion of the core of this study has been already published as a conference paper https://doi.org/10.23919/SpliTech58164.2023.10193249 with a high degree of overlap in content, methodology, and key findings of the current article under review. Here it seems that there is an extended and refined version of the previously published material, thus I leave it up to the editors to make a decision about that.



- Although the study acknowledges spatial autocorrelation effects,  does not fully resolve them, leading to reduced model performance in regions with fewer fire records (e.g., Attica). This undermines the reliability of the model when applied to areas with limited historical data, reducing its effectiveness for wildfire prediction.



- In this study the authors develop separate models for different pilot sites and then compare them to a full model. While this approach helps capture local variations, it may lead to overfitting within specific regions, limiting the model’s ability to generalize ignition likelihood across broader areas​. Although the authors discuss some of these aspects (e.g., Section 4) they could provide a more detailed discussion and clearly state all the limitations of their approach.

- For further improvement: Although the authors state some of these issues in Section 4.4, their study focuses on static environmental and anthropogenic variables but does not incorporate seasonal or real-time human activity variations (e.g., increased tourism in summer, agricultural burning periods). Since human behavior significantly influences fire ignition, integrating temporal dynamics would improve model accuracy​.



- Although the temporal coverage is short in most areas, did the authors consider any temporal trends in DFMC?

Minor comments

- Line 28: AUC abbreviation is not introduced earlier.

- Line 105: Needs to be revised-error message.

- Line 115: Maybe “seasonal” instead of “annual”?

- Lines: 104-116: Some references to the related statements are necessary here.

- Lines 128-132: The native resolution of the fuel type is missing here.

- Lines 183-185: Could the authors be more specific about the terms reclassifying and merging? Does this also involve any regrid method and if yes, which one?

- Line 191-194: Is this daily-mean or daily-max DFMC? Could you please clarify what do you mean by aggregating daily values to annual products? Furthermore, could the authors specify the time scale of the 5th percentile and the anomalies? Are these multi-year daily climatological values or something else?

- Lines 239-240: Needs to be revised-error message.

- Line 258: needs to be revised.

- Lines 259-276: Could authors provide some further explanation for the limited importance of DFMC in PS4 and especially in PS5? Is this related only to the more frequent low DFMC conditions compared to the northern sites?

- Could the authors provide some explanation for the limited role of fuel type as a predictor? The study finds that fuel type is not a significant factor in human-caused ignitions, which contradicts existing research. This could indicate potential data quality issues or model design limitations. A sensitivity analysis on fuel-related variables would clarify this discrepancy​.

- The study uses multiple terms for similar human-related ignition factors (e.g., "human pressure on wildlands," "accessibility," "population influence"). Standardizing terminology throughout the paper would improve clarity and coherence​.



- Figures illustrating ignition probability (Fig. 4) distributions lack sufficient annotation or explanation (e.g., annotations of subfigures). Enhancing the clarity of these visuals would make the findings more accessible. Furthermore, the colobar  for the probabilities could be revised to better communicate the results.

Reviewer Comments

The authors in this paper present a comprehensive and methodologically robust assessment of human-caused wildfire ignition probability across diverse European landscapes. By combining machine learning techniques, specifically Random Forest models, with high-resolution geospatial and socio-environmental data, they deliver both localized and regionally integrated ignition probability models. The overall quality of writing is good, with a clear structure, appropriate referencing, and a sound methodological framework.

The topic is highly relevant and timely, particularly given the increasing wildfire risk under changing climate and land use dynamics in Europe. Importantly, the authors' focus on the interplay between local ignition drivers and their generalization into a full model provides valuable insight into the complexity and variability of fire ignition processes. This shift from local to pan-European modelling is of great significance for the development of integrated fire management strategies at the EU scale.

Overall, the manuscript is a solid contribution to the scientific understanding of ignition patterns and offers operationally meaningful outcomes for fire risk management and prevention across Europe.

Below I provide a series of detailed comments and questions that may help the authors strengthen the manuscript even further:

Specific Comments and Questions

• Section 2.3.3: How did you identify the so-called "mixing areas" algorithmically? Some additional details about the method used would be appreciated.
  
  
  
  
• Section 2.3.4: What are the FirEUrisk fuel classes used in the study? Can these be differentiated enough to capture important distinctions in land cover such as eucalyptus in Portugal, which, despite being a broadleaf, behaves quite differently due to its high flammability? Also, how did you project a 10-meter resolution (CLC+ Backbone) raster to 100 meters, considering the categorical and delicate nature of land use data?
  
  
  
  
• Line 208: What exactly is meant by "null model"?
  
  
  
  
• Line 210: The term number of predictors should be highlighted, perhaps using italics or quotation marks, for clarity.
  
  
  
  
• Line 214: Reference to “Section 0” is likely a formatting or numbering error and should be corrected.
  
  
  
  
• Line 214 (continued): Was the Autocorrelation Control (AC) also used in the full model with all the regions? If so, what was the bounding box adopted?
  
  
  
  
• Line 217: The AC strategy deserves more clarification. It seems to include:
  
  
  
  
• Distance from the center of each PS
  
  
  
  
• Distance from the corners of each PS bounding box
  
  
  
  
• x and y coordinates in the adopted CRS
  
  
  
  

Given this, I would expect that a local PS model relying too heavily on the AC variables (as visible from importance rankings) could be overfitting the ignition patterns of its training set rather than capturing true statistical drivers of ignition. The dummy variable for the PS code used in the full model seems to act similarly to the AC, but at a larger scale. Could you please clarify what is meant by “The AC control successfully alleviated spatial autocorrelation…” (line 217) and confirm whether “disregarding AC control” means simply removing all AC variables from the feature set?

• Figure 2 (Line 250): This figure highlights an extremely important aspect. The choice of presence and pseudo-absence points can shift AUC from 0.4 to 0.9, as in the case of East Attica. This issue is critical and often neglected in wildfire susceptibility literature.
  
  
  
  
• Line 260: Please specify that “dry and warm season” refers to the Swedish climate, which may not be intuitively understood by all readers.
  
  
  
  
• Line 266 and elsewhere: The phrase "chances of ignition" may be misleading. Human sources of ignition (e.g., arson, negligence) are usually orders of magnitude higher than the fires actually recorded. What determines whether a fire is recorded is its success in developing beyond a minimal threshold. It would be more precise to refer to “chances of successful ignition” throughout the manuscript.
  
  
  
  
• Caption of Table 2: The sentence should be revised to: “The top 5 variables for each column are highlighted in grey.”
  
  
  
  
• Figure 4 (Line 310): While the figure is clear and well done, the discussion could be enriched by acknowledging a critical issue in model interpretation: susceptibility values from RF (ranging 0 to 1) cannot be meaningfully compared across pilot sites. A 0.999 value in Sweden does not equate to a 0.999 in Attica. This is where the full model provides value by smoothing across regions. In some of my previous work, I have addressed this using quantile ranking—i.e., describing a pixel as “top 5% susceptibility” within its region, rather than relying on the raw RF voting score. Consider discussing this approach or acknowledging the issue.
  
  
  
  
• Line 325: Since temporal variability is removed by taking average values, I assume your model highlights “spatial areas where extreme dry events tend to occur” rather than correlating specific years with ignition. Is that correct?
  
  
  
  
• Line 339: "Modest climate conditions" is ambiguous. I believe what you mean is that Attica’s fire season is, under climatic/weather conditions,  uniformly and persistently extreme. It would be more accurate to use the phrase “fire-prone climate” here instead of “favourable climate”.
  
  
  
  

Additional Literature Suggestions

1. On RF versus other techniques, and wildfire susceptibility modelling performance  (Section 2.4, line 195):
   
   
   
   

• Trucchia, A., Izadgoshasb, H., Isnardi, S., Fiorucci, P., Tonini, M. (2022). Machine-Learning Applications in Geosciences: Comparison of Different Algorithms and Vegetation Classes’ Importance Ranking in Wildfire Susceptibility. Geosciences, 12(11), 424. https://doi.org/10.3390/geosciences12110424