|This is an interesting topic, within the scope of the journal, and especially appealing given that, as stated by the authors, the number of ignitions in Belgium are likely to increase due to climate change. The paper is clearly written and data and methods are clearly described.|
However, from my point of view, there is a certain number of aspects that put at stake the results presented and the conclusions reached in the paper. This is why I recommend that this paper should not be accepted at this stage, although I encourage the authors to look at my criticism (that I intended to be constructive) and resubmit it at a later stage.
1. My first major concern is the choice of a 10m spatial scale for this study (1st paragraph of section 2.4). Why did the authors chose such a small scale? And how did they handle the implications of such choice? Here are some of my concerns:
1) Given that the majority of ignitions were extracted from newspapers, how did the authors assign the location at a 10m scale? and, for those events where there is GPS data, is the precision less than 10m? and even so, is it really required to locate ignitions at a 10m scale?
2) What is the original spatial resolution of the land cover, soil and land use data? How was the rastering performed? With a nearest neighbor method?
3) How was spatial correlation handled, especially when choosing the tests of hypothesis? At 10m resolution, the predictors will certainly be highly correlated.
4) Is equation 5 still applicable with spatially correlated data?
2. My second major concern is the choice of predictors. Are the chosen three predictors (land cover, soil and land use) independent? How did the authors handle the (quite plausible) dependence among predictors?
The authors show (Fig 5a) the existence of a strong annual cycle but they restrict to static predictors. Since ignitions were mainly obtained from newspapers, it is likely that the information is biased towards ignitions associated to larger burned areas. What is then the importance of associated weather conditions? At least, the authors could have presented the distribution of the color code associated to each ignition (see subsection (iv) - Prevention in section 2.1) to show whether meteorological conditions could be (or could not be) disregarded.
And given that most ignitions are originated by man (see subsection Ignition sources, p. 10), then how was this aspect taken into account?
3. My third major concern is the meaning of obtained probabilities. The author express their concern about the meaning of probabilities as found in previous studies based on "data-driven" methods (2nd paragraph of section 2.3). Why so? And why is the approach they are proposing better than the others? Is it useful for operational purposes to know that a 10m cell will have a probability to burn in a given year of the order of one in a million?
And given the spatial correlation is it true that the areal probability can be computed using Equation 5? And is it reasonable to admit that such probability is the same for all years?
In what sense is this information more useful than the one we directly obtain from a visual inspection of the spatial distribution of ignitions? (Figure 1)
4. The authors have adopted the term "wildfire risk". I think the term "wildfire danger" is more appropriate since the study only deals with probability. However, I recognize that the term adopted by the authors is also used in the literature; unfortunately terms related with hazards are sill confusing.
5. p.4, subsection (i) - Prevalence, p. 4: The authors refer to the extreme event in 2011. At what time of the year did that event take place? How long did it last? Where is it precisely located? What were the meteorological conditions? What was the color code? What is the probability computed by the authors for that location? (This specific aspect should have been discussed in section 3.3) Aren't there any other large events worth being discussed?
6. In subsection (iv) - Prevention, the description of color codes for wildfire risk lacks of details. What factors were taken into account to define the color codes? Structural factors (landscape, proximity to roads, etc.)? Meteorological factors (air temperature and humidity, wind)? How were these color codes validated? Why were they not considered in this study?
7. p. 7, section 2.4, 3rd paragraph: How are the different soil types related to soil moisture in Belgium?
8. p. 8, last paragraph of section 2.4: "Furthermore, in most cases the location of wildfire interventions by firefighters is identified by means of a residential address (...), possibly biasing the perception of wildfire occurrence in function of distance to roads". If so, how were then attributed such ignitions to a 10m cell?
9. p. 10, subsection Temporal distribution: what is the impact of 2010-2013 data on the study, especially because of the higher number of recorded wildfires (that, as stated by the authors, may be just a result of better records)? And, respecting to the peak in April (Fig. 5a) it it the result of a specific year?
10. p. 10, subsection Ignition sources. What about negligence and arson? Aren't they important in Belgium? If so, why?
11. p. 11, last paragraph of section 3.2. The description of how environments were merged is vague. This should have been carefully described, since their choice is crucial to the type of results to be obtained.
12. p. 11, 3rd paragraph of section 3.3. Are the gaps in the histograms real? Or could they be artifacts resulting from the small number of cases in the sample? This is also a crucial aspect since the defined risk classes are based on the existence of such gaps.
13. p. 12, last paragraph of section 3.3. The authors state that probabilities using approaches such as the one based on logistic regression "cannot be interpreted as ignition probabilities, but rather as the similarity between the spatial characteristics of a given pixel and the average spatial characteristic of historical wildfires". Even if so, my question is the following: what is more useful for wildfire management and prevention? such probabilities or the ones resulting from the methodology proposed in this study?
14. p. 12, 3rd paragraph of section 4. In what sense are probabilities obtained in this study "meaningful ignition probabilities that can be interpreted as such"? I really cannot understand the meaning of this sentence? Neither can I understand the meaning of "conservative estimate" in the sentence that follows? Why do the authors think the obtained probabilities are conservative?