Forest fire threatens global carbon sinks and population centres under rising atmospheric water demand

Levels of fire activity and severity that are unprecedented in the instrumental record have recently been observed in forested regions around the world. Using a large sample of daily fire events and hourly climate data, here we show that fire activity in all global forest biomes responds strongly and predictably to exceedance of thresholds in atmospheric water demand, as measured by maximum daily vapour pressure deficit. The climatology of vapour pressure deficit can therefore be reliably used to predict forest fire risk under projected future climates. We find that climate change is projected to lead to widespread increases in risk, with at least 30 additional days above critical thresholds for fire activity in forest biomes on every continent by 2100 under rising emissions scenarios. Escalating forest fire risk threatens catastrophic carbon losses in the Amazon and major population health impacts from wildfire smoke in south Asia and east Africa.

1. One of the key strengths of this work is the use of a huge sample of fire events on daily timescales globally. I would suggest really highlighting this given that most global climatefire studies look at monthly or seasonal relationships. However, the sampling approach for quasi-absence data seems off. I would suggest sampling from within the local fire season to avoid conflating weather (i.e., VPD) that is occurring during a different time of the year. 2. The authors should make a better effort to state why they use VPD rather than other variables. For example, Brey et al., (2021) contends that VPD shows a much larger increase in fire potential vs other variables (e.g., fuel moisture, fire danger). It might help to argue why VPD is a better proxy that other factors both in a contemporary perspective -and under future climate scenarios -since VPD is looked at here exclusively. That said, there is precedent for using VPD or temperature in forward looking studies. For example, Gutierrez et al., (2021) found that daily maximum temperature (and VPD) nonlinearly increased fire risk and used such projections to estimate future fire activity. 3. The methods of the paper have many potential holes that need to be addressed including 3.1. Fire-VPD analyses: 3.1.1. The MODIS burned area data provides a daily stamp of when the 500-m pixel burned. In the analysis, how do you deal with data independence? Namely, during large fire events, you may have an entire ERA-5 cell with fire on a given day. Would you count ~3000 cells with identical VPD values? 3.1.2. Is there any way to demonstrate that you are sampling from the same climateniche as your fire pixels? I think you are fine, unless for some reason fire pixels happened in a significant different climatological subregion. You could in a supplemental analysis show that the locations or climate of locations for presence and quasi-absence are effectively similar to counter this. 3.2. Treatment of GCM data: 3.2.1. While I am glad to see the authors used more than 1 GCM, I don't consider 3 GCMs to be a particularly robust result. While there is a statement about "skill selected global climate models", I don't think the reference looked at VPD. Why not use more here? Typically for climate change assessment you want to use at least 10 GCM ensembles. 3.2.2. Each GCM will have biases though, so you'd ideally want to perform bias correction to get comparable VPD. The delta bias-correction approach for treating GCM data would be fine if the daily distributions of VPD from GCMs credibly represented those from ERA-5. It is unlikely that they are though.
While delta bias-correction is OK for many climate change assessments, I am concerned here given that use of thresholds from ERA-5 data were used. One would ideally want to use a more sophisticated BC approach here to account for potential large differences in the distribution. 3.3. Exposure data 3.3.1. Smoke impacts can spread well downwind of fires. In the absence of using a smoke dispersion model, it would be good to have strong justification for a distance from potential fires. There is reference to "GCM allowing smoke transport for tens of kilometers", which doesn't seem very logical as GCM resolutions are often 100- I reviewed the paper titled "Forest fire threatens global carbon sinks and population centres under rising atmospheric water demand" by Hamish Clarke and colleagues. The authors use logistic regression models to relate the probability that a pixel burned to daily vapor pressure deficit (VPD) for 70 different regions across the globe. These models were then used to identify a VPD threshold at which the probability of fire >50%; the number of days exceeding this threshold was then calculated for each of the 70 regions. Next, the authors used climate change models to project the change in the number of days exceeding this threshold in the future. Finally, projected changes in the number of days exceeding the VPD threshold were intersected with maps of carbon and human population density to infer impacts to carbon and humans. This is an interesting paper, but I do have concerns about whether or not the number of days exceeding the VPD threshold is actually related to area burned among the 70 regions analyzed. This is not presented, and when I qualitatively look at the results and compare them to my mental image of fire prone areas, I am not convinced there is a relationship. A little less qualitative: in looking at the animation here (https://earthobservatory.nasa.gov/globalmaps/MOD14A1_M_FIRE), Japan and the Korean peninsula have very little fire, yet according to figure 3a, it has by far the highest number of days exceeding the VPD threshold. The same can be said for parts of Europe such as Scandinavia. A more complete feedback can be found below.  fig. 3a) would be exceptionally fire prone and exhibit high amounts of area burned. However, I don't consider Japan and the Korean peninsula particularly fire prone. Same goes for northern Europe. In the southeastern USA, the frequency of VPD exceedance is high, but most fires are prescribed fires.
2. Related and very important: if the frequency of days exceeding the VP threshold and area burned are not at least moderately correlated, and projections of effects to carbon or people under a future climate are potentially suspect. I guess I'd like to see some sort of analyses that relates the frequency of VPD exceedance to area burned (by sub-continental window, for example). If this relationship is moderately strong, then there is reason to make the projections under climate change.
3. I'm guessing that a lot of the fire seen in some parts of the planet are cultural, agricultural, or prescribed fires. If this is the case, can projections like this even be made? Related, for those fires that are cultural/agricultural/prescribed, they probably serve to stabilize carbon, meaning these fires are generally intended to preserve large trees and not kill them. So many fires in these areas are not necessarily a threat to carbon, now and into the future.
Moderate concerns: 1. Scientists are increasingly being criticized for exaggerating the effects of climate change by, for example, using the most extreme climate change scenarios in their analyses. It is my understanding that RCP 8.5 is unlikely, so it is perhaps more appropriate to use a more relevant emissions scenario for the main findings. I know it is less splashy, but I think it is important to not overexaggerate climate change effects in the abstract and the main figures in the paper. Sure, keep RCP 8.5, but that might be better in the supplemental as opposed to the main findings.
2. The absolute change in the number of days exceeding the VPD threshold under climate change is interesting, but I am wondering about a relative metric of projected change. For example, I can envision some parts of the boreal forest increasing from 10 to 30 days, which is 20 day increase but a three-fold increase. Will the implications of this differ when compared to a projected increase from 100 to 120 days in other parts of the globe?
Response to reviewer comments -Forest fire threatens global carbon sinks and population centres under rising atmospheric water demand Note that line numbers quoted in the response refer to the manuscript with tracked changes showing all markup.

Reviewer 1 Comment 1 (R1.1)
One of the key strengths of this work is the use of a huge sample of fire events on daily timescales globally. I would suggest really highlighting this given that most global climate fire studies look at monthly or seasonal relationships.

Response
Revised. L25-28, L62-65, L167-170. Thank you for raising this point. We agree that this aspect was not emphasised in the original aspect and have added text to the abstract (L25-28), introduction (L62-65) and discussion (L167-170) in order to further emphasise this aspect of the study.
L25-28: "Using a large sample of daily fire events and hourly climate data, here we show that fire activity in all global forest biomes responds strongly and predictably to exceedance of thresholds in atmospheric water demand, as measured by maximum daily vapour pressure deficit." L62-65: "Our use of daily remotely sensed burned area and hourly climate reanalysis data is a key advance on previous studies, which typically focusing on aggregate measures such as total area burnt over a season." L167-170: "Our findings provide new evidence at a high temporal resolution (i.e. daily) of the link between fuel moisture and forest fire activity 30,31 and the potential for fuel moisture-mediated changes -nearly always increases -in risk due to climate change 32-34 ."

R1.2
However, the sampling approach for quasi-absence data seems off. I would suggest sampling from within the local fire season to avoid conflating weather (i.e., VPD) that is occurring during a different time of the year.

Response
Not revised. We understand this comment as many studies have restricted their analysis to 'the fire season'. Notwithstanding the lack of consensus on precisely how to define the fire season, our intention is to model the relationship between atmospheric dryness and fire activity and to do this objectively we do not wish to bias our sample by excluding data unnecessarily and in particular by restricting our analysis to the relatively short satellite record over which fires have been observed.
Additionally, we would like to stress the important point that many studies indicate that the fire season has already lengthened in many areas globally under anthropogenic climate warming (Jolly et al. 2015) and the trend is expected to continue. In other words, the current fire season is a dubious proxy for the future fire season. Our approach does not assume that fire will remain within historical windows and has the further advantage of setting a baseline against which to detect future changes in the timing and duration of fire activity.

R1.3
The authors should make a better effort to state why they use VPD rather than other variables. For example, Brey et al., (2021) contends that VPD shows a much larger increase in fire potential vs other variables (e.g., fuel moisture, fire danger). It might help to argue why VPD is a better proxy that other factors both in a contemporary perspective -and under future climate scenarios -since VPD is looked at here exclusively. That said, there is precedent for using VPD or temperature in forward looking studies. For example, Gutierrez et al., (2021) found that daily maximum temperature (and VPD) nonlinearly increased fire risk and used such projections to estimate future fire activity.

Response
Revised. L169, L170. We have strengthened the case for VPD by adding references in the discussion to Gutierrez et al. (2021)

R1.4
The MODIS burned area data provides a daily stamp of when the 500-m pixel burned. In the analysis, how do you deal with data independence? Namely, during large fire events, you may have an entire ERA-5 cell with fire on a given day. Would you count ~3000 cells with identical VPD values?

Response
Revised. L498-500. Thanks for raising this point. We have clarified in the methods that this is indeed the case.
"Due to a mismatch between the spatial resolution of fire and climate data, the same VPD value may be assigned to multiple burned area grid cells within a single climate grid cell." Ultimately one of the constraints on our analysis is the spatial and temporal resolution of the climate and fire data. While this results in unsampled variation for the averaging region and time period (as in your example) our results show that there are still strong regional patterns in the relationship between VPD and fire activity.

R1.5
Is there any way to demonstrate that you are sampling from the same climate niche as your fire pixels? I think you are fine, unless for some reason fire pixels happened in a significant different climatological subregion. You could in a supplemental analysis show that the locations or climate of locations for presence and quasi-absence are effectively similar to counter this.

Response
Revised. L502-504, Supplementary Fig. 8. We have added a summary of the correlation between the climate of presence and absence data using independent very high resolution climate data (Fick & Hijmans 2017).
The high correlations provide confidence that these data have been sampled from the same climate niche. Please refer to Supplementary Fig. 8, reproduced below. Text has been added to the methods as follows: "A supplemental analysis confirms that presence and absence data points are drawn from the same climate zone ( Supplementary Fig. 8)."

R1.6
While I am glad to see the authors used more than 1 GCM, I don't consider 3 GCMs to be a particularly robust result. While there is a statement about "skill selected global climate models", I don't think the reference looked at VPD. Why not use more here? Typically for climate change assessment you want to use at least 10 GCM ensembles.

Response
Revised. L471-481, Supplementary Fig. 7. Thanks for raising this point and highlighting the need to more clearly present our rationale. We share Reviewer 1's concerns for maximising the robustness of the GCM data but we disagree that simply adding more GCMs will achieve this. We briefly discuss our approach in the context of climate model ensemble design to explain why.
Using a full ensemble of all available models is the gold standard for comprehensive projections (particularly where the quantity of interest is directly simulated by the model) and model evaluations, as found in IPCC reports, but is exceedingly rare in wildfire impact assessments. As an example there are 36 CMIP5 models which include historical, RCP4.5 and RCP8.5 data for air temperature, precipitation and specific humidity (https://esgf-node.llnl.gov/projects/cmip5/). An informal survey of around 20 wildfire projection studies yielded a maximum ensemble size of 23 (Hanan et al. 2021), with about half having 11-20 models and the other half having 1-10 models.
In the absence of a full ensemble, the most defensible approach in our view is using an objectively designed ensemble. This involves explicit selection of a subset of models from the full ensemble, following objective and transparent criteria that allow users to understand the implications of what has and has not been included. Objectively designed ensembles are commonly selected on the basis of model skill, model independence, or the ability of models to span the range of projected future change in climate -ideally all three (Evans et al. 2014;. There is greater confidence in projections based on such ensembles due to the omission of poorly performing models, the lack of duplication of models with similar biases and the focus on the outer bounds of the distribution of likely changes in future climate, which are of great interest to decision makers including fire managers. A third, less robust approach to climate projections is to use 'ensembles of opportunity'. Such ensembles vary widely in size and can be quite large (18, 19 and 23 member ensembles featured in three wildfire studies published in 2021), but generally lack a rationale for why some models are included and others excluded. Nevertheless, confidence in ensembles of opportunity can be improved by providing information on things like the skill, independence and fractional range of coverage of future climate space for those models included with respect to the full ensemble Frieler et al. 2017). The number of GCMs used in a study is thus not by itself a reliable indicator of the robustness of results.
Regardless of which approach is taken, a necessary final step in any model selection process is restricting the ensemble to those models for which the required output data are available i.e. particular CMIP experiments, climate variables, time resolution, etc..
For our study we developed an objectively designed ensemble following the comprehensive global evaluation of CMIP5 models by . We began with six models with a performance ranked as 'satisfactory' and output ranked as 'outliers' i.e. spanning the range of future changes in climate. We cross-referenced this with 12 GCMs for which 3-hourly data was available for historical, RCM4.5 and RCM8.5 experiments and for the specific humidity, air temperature and surface pressure variables required to calculate VPD. This resulted in a two member ensemble (ACCESS1-0 and CNRM-CM5). To increase the ensemble size we then expanded the search to a second set of six models from McSweeney et al. (2015) whose performance was ranked 'satisfactory' and whose output was not considered an outlier. Required output data were available for three of these models, however they were all variants of the GFDL model and were thus not independent. We selected the newest of the three models (GFDL-CM3), which has an improved representation of a number of atmospheric and land surface processes compared to the other two GFDL models.
To improve the robustness of our findings we see three main possibilities. One, relax the skill requirement and include all models for which 3-hourly data is available. This would generate a broader range of output, albeit a range featuring multiple models that have been flagged by  as sufficiently implausible as to undermine confidence in their projections. Two, relax the 3-hourly data requirement and use coarser resolution data. As noted by Reviewer 1, the use of daily data is a key strength of this work (we have revised the manuscript to emphasise this) and using the highest possible time-resolution data (hourly data for reanalysis and 3-hourly data for GCMs) provides the most confidence in the robustness of our estimates of maximum daily VPD. Three, provide further information to help assess the robustness of the selected models in the context of the full set of CMIP5 models. This information shows that in most cases, the models we have included span most or all of the range of projected changes in climate. There is therefore reasonable confidence in the robustness of our projections. Based on this systematic review of model ensemble characteristics we concluded that the third option is preferable to the first two and refer to Supplementary Fig. 7 below, reproduced here. We have added selected items from this response to the Methods section to clarify model selection and provide further information about its robustness as follows: "For climate change analyses we selected three global climate models from the CMIP5 dataset on the basis of skill, independence and the ability to span the range of future changes in climate: ACESS1.0, CNRM-CM5 and GFDL-CM3 (Supplementary Table 5). These models were among the best performing compared to other CMIP5 models in a comprehensive evaluation for the purposes of downscaling over multiple regions, which included annual cycles of rainfall and temperature, general circulation patterns, teleconnections and the south east Asian monsoon. Of the highly performing models evaluated, these three models generally spanned all or most of the range of projected future seasonal and regional changes in climate ( Supplementary Fig. 7). We avoided models from the same model family to avoid duplication of models with similar biases."

R1.7
Each GCM will have biases though, so you'd ideally want to perform bias correction to get comparable VPD. The delta bias-correction approach for treating GCM data would be fine if the daily distributions of VPD from GCMs credibly represented those from ERA-5. It is unlikely that they are though. While delta bias-correction is OK for many climate change assessments, I am concerned here given that use of thresholds from ERA-5 data were used. One would ideally want to use a more sophisticated BC approach here to account for potential large differences in the distribution.

Response
Revised. L105-109, L110-113, L485-487, Fig. 3-5, Supplementary Fig. 4-6. We have bias corrected all global climate model data using a quantile mapping approach (Cannon et al. 2015) commonly used in wildfire research (Abatzoglou et al. 2019;Kirschmeier-Young et al. 2019;Ruffault et al. 2020). After bias correction the spatial pattern of results is broadly similar, the magnitude of increases is generally greater (particularly for the ACCESS1-0 model) but the overall conclusions are unchanged.

R1.8
Smoke impacts can spread well downwind of fires. In the absence of using a smoke dispersion model, it would be good to have strong justification for a distance from potential fires. There is reference to "GCM allowing smoke transport for tens of kilometers", which doesn't seem very logical as GCM resolutions are often 100-200km. Again, it might be useful to point to literature on wildfire smoke impacts here to justify choices for population exposure.

Response
Revised. L524-526. Thanks for raising this point. Population exposure to wildfire smoke is a complex function of many factors including fire intensity, weather conditions and plume injection height (

R1.9
Did you use projected population density data or leave things at the 1990-2000 levels?
Response Revised. Fig. 5, L32-34, L124-129, L205-207, L526-528. Thanks for raising this. We originally used current population levels but have updated our plots using "middle of the road" projections for 2090 drawn from the same dataset (Jones et al. 2016). The use of these projections led to some changes in the pattern of future population exposure to forest fire smoke as follows: L32-34: "Escalating forest fire risk threatens catastrophic carbon losses in the Amazon and major population health impacts from wildfire smoke in south Asia and east Africa." L124-129: "Substantial increases in the number of days over the VPD threshold -and hence days of elevated probability of fire and smoke emissions -are projected to occur by 2081-2100 near major population centres in south Asia and east Africa by all three models (Fig. 5). Two of three models also suggest considerable population exposure to smoke from increased forest fire activity in parts of central America, west Africa and east Asia." L205-207: "We also show that increases in forest fire activity are projected to occur near major population centres in east Africa and south Asia, and possibly central America, east Asia and west Africa." We have updated Fig. 5 and the methods text as follows: L526-528: "Population projections for 2090 were based on a "middle of the road" scenario in terms of expected population growth, urbanization, and spatial patterns of development." Response Revised. Fig. 3-5, Supplementary Fig. 4-6. White added to zero region of colormap.

Reviewer 2 Comment 1 (R2.1)
This is an interesting paper, but I do have concerns about whether or not the number of days exceeding the VPD threshold is actually related to area burned among the 70 regions analyzed. This is not presented, and when I qualitatively look at the results and compare them to my mental image of fire prone areas, I am not convinced there is a relationship.
A little less qualitative: in looking at the animation here (https://earthobservatory.nasa.gov/globalmaps/MOD14A1_M_FIRE), Japan and the Korean peninsula have very little fire, yet according to figure 3a, it has by far the highest number of days exceeding the VPD threshold. The same can be said for parts of Europe such as Scandinavia.
It is not clear how to interpret the results pertaining to the number of days exceeding the VPD threshold. On the surface, I'd think that areas of the globe with high values (in fig. 3a) would be exceptionally fire prone and exhibit high amounts of area burned. However, I don't consider Japan and the Korean peninsula particularly fire prone. Same goes for northern Europe. In the southeastern USA, the frequency of VPD exceedance is high, but most fires are prescribed fires.
Related and very important: if the frequency of days exceeding the VP threshold and area burned are not at least moderately correlated, and projections of effects to carbon or people under a future climate are potentially suspect. I guess I'd like to see some sort of analyses that relates the frequency of VPD exceedance to area burned (by sub-continental window, for example). If this relationship is moderately strong, then there is reason to make the projections under climate change.

Response
Revised. L182-185, L515-517, L517-519. We share Reviewer 2's concern that our results are robust and clearly communicated, so we have pooled several related comments about the relationship between the number of days that empirically-derived VPD thresholds are exceeded, and fire activity.
The heart of our paper is logistic regression modelling of the daily maximum VPD value (plus confidence intervals) at which the probability of fire -the probability of observing a burned area detectable on MODIS imagery -is greater than 0.5.
We show there is a strong relationship between VPD and fire activity in a wide range of fire-prone forest biomes around the world ( Supplementary Fig. 1, Supplementary Table 1). Although there are some regions where this relationship is weak, the worst performing models represent about one seventh of the mean annual area burnt of the best performing models (Supplementary Note 1).
This relationship is derived on an individual forest biome / subcontinental window basis. There is considerable variation in the value of the VPD threshold within and between biome groups (subtropical to tropical, temperate and boreal, and mediterranean), as well as in the number of days over this threshold.
As suggested by the reviewer, we calculated the correlation between days over VPD threshold and total burned area for each combination of forest biome and sub-continental window, based on monthly area averages rather than individual daily pixels ( Supplementary Fig. 9, reproduced below). These results are consistent with our main analysis i.e. the correlation is generally strong (r averaged across biomes is 0.39), but varies within biomes.
Importantly, it does not follow from any of our analyses that area burnt scales with the total number of days over VPD threshold. The strength of the relationship is not necessarily correlated with the magnitude of area burnt or the number of days over threshold for any given forest biome and sub-continental window.
On a related note, our analysis is guided by the four switch model (Archibald et al. 2009;Bradstock 2010), which argues that the fundamental biophysical preconditions for landscape fire are 1) sufficient fuel, 2) sufficiently low fuel moisture, 3) sufficient fire weather conditions and 4) an ignition source. A corollary of this model is that fire regimes vary in the identity of the switch that acts to limit overall fire incidence overall. As noted in the introduction, there is good evidence that fuel moisture plays an important role in limiting fire behaviour in forest biomes globally, given their abundance of fuel. It is likely, however, that in some circumstances ignition will limit forest fire activity even in the presence of sufficiently dry fuel. High population density, high detection rates and high suppression capacity are all known to lower the effective ignition rate (Collins et al. 2018;Clarke et al. 2019).
In summary, we have modelled fire activity as a function of VPD, but 1) there is not necessarily a relationship between the absolute number of days over VPD thresholds and the absolute amount of fire activity, 2) there are areas where the model does not perform well, and 3) there are areas where fuel moisture may not be the limiting factor on overall fire activity. We have made the following additions to the manuscript: L182-185: "High population density, high fire detection rates and high suppression capacity are all known to lower the effective ignition rate and could weaken the link between VPD and fire activity in some regions." L515-517: "Note that the strength of the relationship between VPD and fire activity in any given region does not imply a particular magnitude of burnt area for a given number of exceedances of daily VPD threshold values." L517-519: "A supplemental analysis examined the relationship between area averaged monthly days over VPDP=50 and burnt area, with broadly similar findings to the main analysis ( Supplementary Fig. 9)."

R2.2
I'm guessing that a lot of the fire seen in some parts of the planet are cultural, agricultural, or prescribed fires. If this is the case, can projections like this even be made? Related, for those fires that are cultural/agricultural/prescribed, they probably serve to stabilize carbon, meaning these fires are generally intended to preserve large trees and not kill them. So many fires in these areas are not necessarily a threat to carbon, now and into the future.

Response
Revised. L462-465. Thanks for raising this point. We do not consider agricultural burns, cultural burns or prescribed burns to have accounted for a significant proportion of our fire activity data. Prescribed burns are generally of far lower size and intensity than wildfires and are frequently undetected by MODIS (Chuvieco et al. 2020). Cultural burns are of even smaller magnitude than prescribed burns.
Given our focus on forests, we have used a forest mask (Schepaschenko et al. 2015; see Methods section L450-451) and consequently it is unlikely that agricultural fires represent a strong contribution. While cultural and prescribed burning may serve to stabilise carbon in some cases, the burns associated with logging in the Amazon and other tropical rainforest regions are less likely to do so. We have added the following text: "Although they are of great interest, prescribed and cultural burns are not likely to have accounted for a significant proportion of the fire activity data as they are generally of far lower size and intensity than wildfires and are frequently undetected by MODIS 61 ."

R2.3
Scientists are increasingly being criticized for exaggerating the effects of climate change by, for example, using the most extreme climate change scenarios in their analyses. It is my understanding that RCP 8.5 is unlikely, so it is perhaps more appropriate to use a more relevant emissions scenario for the main findings. I know it is less splashy, but I think it is important to not overexaggerate climate change effects in the abstract and the main figures in the paper. Sure, keep RCP 8.5, but that might be better in the supplemental as opposed to the main findings.

Response
Not revised. We agree that care needs to be taken in presenting results from realistic emissions scenario. We acknowledged in the original manuscript the increasing plausibility of RCP4.5 (L115-117: "Under a lower and increasingly more plausible emissions scenario (RCP4.5) the magnitude of change is smaller but still features widespread increases in the annual frequency of days of elevated probability of fire ( Supplementary  Fig. 5 and 6)").
However, global CO2 emissions rebounded by nearly 5% in 2021 (IEA, 2022) suggesting that a trajectory closer to RCP4.5 than RCP8.5 is not a fait accompli. In addition, it has recently been argued that worst case scenarios are underexplored (Kemp et al. 2022). On balance we argue that understanding the potential effects of RCP8.5 on forest fire risk is important enough to place this scenario in the main text, with further results in the Supplementary Information.