Seasonal Driving Mechanisms and Spatial Patterns of Danger of Forest Wildfires in the Dongjiang Basin, Southern China

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors!

I think that this article is in the scope of Forest journal, but at present time article has some disadvantages.

You presented jointed Introduction and Background section. It is necessary to add works on wildly used systems to predict forest fires like NFDRS in the USA, CFFDRS in Canada, EFFIS in EU, ISDM-Rosleskhoz in the Russian Federation. Also you must add information on alternative methods to predict forest fires like deterministic, probabilistic, deterministic-probabilistic, regressions, Artificial neural networks, etc.

Why did you use term "Susceptibility"? This wrong terminology widely spread in current published articles, but this terminology wrong. You must use term "danger" instead of "susceptibility". Forest fire occurrence prediction based on two general terms of solid theory of risk analysis - "danger" and "risk". Probability of disaster occurrence in risk analysis is danger, while consequences is risk. Risk is equal probability multiplied by waiting damage. There are no such term "susceptibility" in strong mathematical theory of risk analysis. You write scientific publication and must use strong terminology. Term "susceptibility" shown general degradation of scientific community in the field of forest fire danger and risk prediction and evaluation.

2 Study area and methodology

You wrote that MODIS device is used in the article, in particulare, high level products MOD14 and MYD14. Why did you choose MODIS device? Why did you not choose, for example, Suomi NPP VIIRS?

You used population density and distance to objects in you research.How did you take into account definite sources of high temperature in your study? This is very important to estimate forest fire danger.

Clarify, why did you generate 4629 pseudo-non-fire points in distance of 1 km from actual forest fire points?

Please, provide brief description with mathematical basis of Random Forests algorithm. How did you implement this algorithm? What GIS-system was used with built-in program instruments? Or did you write you own program code? What development system was used for this purposes in case of your own program code? What programming language was used in this case?

There are many similar works in scientific periodics using Random Forests algorithm. Differences lay in the area of study only. What the real novelty of you scientific work? Or you used well known algorithm only to another geographic area?

3 Results

You wrote about potential fire risk. This is wrong terminology. Risk is probability of forest fire multiplied to waiting damage. What damage from forest fires is considered in your work? What consequences for economy, ecology or population did you consider in your work? I pretty confident, that you must use term "danger" instead of term "risk".

5 Conclusion

This section must be completely reworked. Please, provide a numbered set of 3-5 key findings with corresponding conclusions.

References

Please, extend this reference list according remarks to Introduction section.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript explores the seasonal driving mechanisms and spatial susceptibility of forest wildfires in the Dongjiang Basin. It is novel in its integration of modern machine learning (random forest) with SHAP value interpretation, providing insights into the contributions of variables such as NDVI and anthropogenic factors. The study' s breakdown by season and its focus on the subtropical monsoon zone of southern China offer fresh insights compared to global- scale analyses and commonly used linear approaches.The utilisation of multi- source data (MODIS, remote sensing, socio- economic) and advanced methods like kernel density analysis enhances the understanding of spatial and temporal wildfire patterns. The research is advanced, combining robust geospatial data processing, machine learning techniques, and model interpretation tools to reveal the complex interactions of wildfire drivers.   Although the concepts are sophisticated, the manuscript communicates methods and results with clarity, making it accessible to readers with an intermediate level of background knowledge. However, non- specialists might require simplified explanations or analogies, especially concerning modelling details and SHAP interpretations.Language

The language throughout the manuscript is generally clear, concise, and well- organised. It avoids excessive technical jargon and provides definitions, such as for the SHAP value and kernel density estimation, assisting in reader comprehension. At certain points, details on multicollinearity tests and variable screening could be simplified further to aid understanding for readers without a technical background .

Overall, the paper maintains a conversational tone in sections such as the introduction and management implications, helping to bridge the gap between technical details and practical applications.

Data presentation is methodical, with clear references to seasonal trends and spatial patterns using figures (e. g., kernel density maps, susceptibility maps), even though specific figures and tables are alluded to rather than displayed in the parsed text. Figures are well-organised and informative. However, there is a lack of explanations. I’d recommend adding more detailed explanations in the figure captions, e.g. Fig. 3. The structured approach to model performance evaluation using metrics such as AUC, precision, recall, and F 1- score offers a comprehensive view of the model' s effectiveness across seasons. Results from the SHAP analysis are discussed clearly with examples of how various factors influence wildfire risk in different seasons.

However, there are weak points in this manuscript. The spatial resolution of MODIS data (1 km) may not capture fine-scale wildfire events, potentially underestimating frequency. Reliance on static environmental data limits the dynamic understanding of climate change and the long-term effects of vegetation succession. The “black-box” nature of the random forest model, despite explanatory SHAP values, could still pose interpretability challenges for some decision-makers. The summer model shows lower precision and higher false positives, suggesting the need for additional refinement.

In conclusion, the manuscript presents a solid and innovative approach to understanding forest wildfire risks in a critical region of China. Integrating multiple data sources, advanced modelling, and clear presentation of seasonal and spatial drivers significantly contributes to the field. Despite some methodological limitations, such as data resolution and model interpretability challenges, the study meets its objectives and offers valuable insights.

Overall, with minor revisions to address these weaknesses, the manuscript appears to be acceptable for publication in the Forests.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 3 Report

Comments and Suggestions for Authors

My comments are included in the main document. 
The work is interesting but needs to be improved.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Please, revise some terms.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors!

Thank you for your revision.

I accepted all your improvements and explanations.

I only recommend to add response 7 in section 2. It will depends on your own opinion make this addition or not.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf