Projections of wildfire risk and activities under 1.5 °C and 2.0 °C global warming scenarios

Wildfires are important ecosystem processes that have a significant impact on terrestrial vegetation, environment, and climate. This study investigates how future wildfire risk and activities could change under 1.5 °C and 2.0 °C warming scenarios relative to pre-industrial levels using a modified McArthur Forest Fire Danger Index (FFDIn) and the CLM4.5-BGC land surface model. Sixteen Earth System Models (ESMs) from CMIP5 and CMIP6 were employed to supply the variables of climate change under low, middle, and high greenhouse emission scenarios in the 1.5 °C and 2.0 °C scenarios. The ensemble means from the FFDIn and results from the CLM4.5-BGC with multiple forcings show that the dry areas in the southwestern US, Brazilian Highlands, and Arabian islands are projected to face higher wildfire risk with larger burned areas and more carbon emissions under a warmer climate. The Congo Basin and part of the Amazon could have a lower wildfire risk with smaller burned areas and less carbon emissions. The absolute changes in the projected FFDIn are small, although large increases are observed in boreal areas, particularly in the winter and spring. Burned area and carbon emissions are projected to increase in general in the boreal area but decrease in northeastern Asia. Compared to the 1.5 °C scenario, the wildfire risk and burned area levels are projected to increase under the 2.0 °C scenario except in the western Amazon. However, fire carbon emissions are projected to decrease more in tropical areas under the 2.0 °C scenario. The different change directions in eastern North America and eastern China produced by the FFDIn and CLM4.5-BGC suggest the potential effect of non-meteorological elements on fire activities.


Introduction
The IPCC's Fifth Assessment Report indicates that the increased concentration of anthropogenic greenhouse gases has led to a significant increase in global temperature since the industrial revolution, which is particularly prominent in the mid-to-high latitudes of the Northern Hemisphere. These changes could further affect ecological degradation and biodiversity reduction, thereby posing a great threat to the sustainable development of global and regional economies and societies [1]. To address with the emergence of climate future change, the Paris Climate Agreement adopted in 2015 commits to limiting the global average temperature to values below 2.0°C above pre-industrial levels, while preferably pursuing efforts to limit the temperature increase to 1.5°C [2]. A series of studies have investigated the possible future responses in atmospheric circulation, freshwater systems, biomass energy, global mean sea level, precipitation, and temperature extremes under the 1.5°C and 2.0°C warming scenarios [3][4][5][6][7][8]. Schleussner et al [9] used CMIP5 data to assess the changes in extreme weather events, water availability, agricultural yields, sea-level rise, and coral reef loss risk at warming levels of 1.5°C and 2.0°C and showed that an increase of 0.5°C would lead to more frequent and intense changes, such as an N to 90°N; and (2) investigate the possible future changes in wildfire risk and fire-related carbon emissions under 1.5°C and 2.0°C warming scenarios.

Data
The daily output derived from the Earth System Models (ESMs) participating in the Coupled Model Intercomparison Project Phase 5 and Phase 6 (CMIP5 and CMIP6. https://esgf-node.llnl.gov/) are used in this study. Considering the availability of ESM daily or sub-daily output, 16 ESMs were selected from CMIP5 and CMIP6 under the different scenarios [60][61][62][63][64][65][66][67][68][69]. Model results were derived from historical simulations and future simulations under representative concentration pathways (RCP2.6, RCP4.5, and RCP8.5) in CMIP5 and shared socioeconomic pathways (SSP126, SSP245, and SSP585) in CMIP6. The 20-year periods were chosen for the ESMs under different scenarios based on the 20-year running averages of temperature when 1.5°C and 2.0°C warming levels were reached (figure 1). Bilinear interpolation was used to unify the spatial resolution of the model results as 2.8125°× 2.8125°. The 20 years from 1986 to 2005 was the baseline period, when the temperature was 0.61°C higher than the pre-industrial level [1]. Therefore, increases of 0.89°C and 1.39°C above the baseline level are equivalently defined as 1.5°C and 2.0°C above the pre-industrial level, respectively. Version 4.1 of the Global Fire Emissions Database (GFED4.1 s [70]), which supplies the observed burned area fractions and biomass burning carbon emissions between 2007 and 2016, is applied to assess the performance of the FFDI in representing real fire occurrence. Daily observations were obtained from the reanalysis of agrometeorological indicators from 1979 to the present [71]. This dataset was based on hourly ECMWF ERA5 data at the surface level.
The atmospheric forcings for the baseline simulation during 1986-2005 were obtained from the Global Soil Wetness Project Phase 3 (GSWP3) [72]. For the two warming simulations, the atmospheric forcing data were derived by adding the ensembled future change in forcings from the ESMs to the forcings during the baseline period. The CRUNCEP reanalysis during 1997-2006 was the atmospheric data that drove the present-day simulation.

Modified mcarthur forest fire danger index
The FFDI was calculated using daily meteorological variables according to equation (1) [35,73]. It builds the relationship between the wildfire index and daily maximum temperature, minimum relative humidity, mean 10-meter wind speed, and drought factors. Before applying the FFDI to project future changes in wildfire risk, we tried to improve the performance of the FFDI to represent the possibility of fire occurrence in mid-to-high latitudes. The non-parametric Spearman correlation test was used to measure the degree of similarity between two variables [74]. Although a higher FFDI does not necessarily mean a larger burned area, the correlation coefficient between these two variables can be used as a measurement to some extent. Assuming that there is a linear relationship between the fire risk index and fire activities, we used observations of burnt areas to quantify fire occurrence following Bett et al [43].
To improve the representativeness of the FFDI in mid-to-high latitudes, we first identified the meteorological and environmental variables that had the greatest effect on monthly fire activities from 2007 to 2016 in the area from 45°N to 65°N and 60°E to 135°E. The examined variables included soil humidity, maximum near-surface temperature, minimum relative humidity, precipitation, 10-meter wind speed, soil temperature of the top layer at a depth of 0-7 cm, dew point, and solar radiation. In addition to these variables, the relationship between VPD and fire activity was explored. The VPD was calculated using the near-surface temperature and relative humidity according to the FAO-56 Penman-Monteith method [75]. Because drought correlates closely with fire activities, the drought effect in the original equation is maintained, that is, the computed value of D 1.275 0.987 (D is the drought factor and is calculated using equation (3)) is retained in the modified formula (equation (2)). We then scrutinized the linear relationship between the natural logarithm of the observed burned fraction Burned Fraction K K D ln , 1.275 0.987 ( (( ) ) ) / = and the meteorological variable because the exponential relationship is more reasonable than the linear relationship between fire activities and meteorological variables [76,77]. The most influential variables in the mid-to-high latitudes were the VPD, precipitation, wind speed, and minimum relative humidity. The modified FFDI is shown in equation (2) where T is the maximum temperature (°C), H is the minimum relative humidity (%), W is the mean 10-meter wind speed (km h −1 ), VPD is the vapor pressure deficit (KPa), PR is the daily precipitation (mm day −1 ), and D is the drought factor calculated by equation (3).
where N is the number of days since the last day with rainfall no less than 2 mm day −1 ; P is the amount of rainfall on that day (mm day −1 ), which is different from PR; and I is the amount of rain needed to restore the soil's moisture content to 200 mm. Considering the difficulty in obtaining global soil moisture observations at daily frequency, I was set to a constant value of 120 mm [42,43]. Furthermore, the FFDIn is set to 0 when the near-surface temperature is less than -10°C according to the setting in CLM4.5 [78,79], and it represents a dimensionless variable. A series of threshold values was used to determine wildfire danger ratings, as previously used by Luke and McArthur [80], as shown in table 1.

Model and simulation setup
Because the selected ESMs supply limited output on fire carbon emissions, the Community Land Surface Model, version 4.5 coupled with the biogeochemistry model (CLM4.5-BGC) [79] was employed in this study to simulate future changes in burned fraction and carbon emissions. CLM4.5-BGC has been widely used in the simulation of coupled carbon, nitrogen, water, and energy cycles on the land surface [81][82][83]. It can simulate fires due to natural and anthropogenic well-being [84]. The daily output of maximum 2-meter temperature, minimum 2-meter temperature, precipitation, surface downwelling shortwave radiation, surface downwelling longwave radiation, 10-meter wind speed, specific humidity, and monthly average surface air pressure from eight CMIP5 ESMs and eight CMIP6 ESMs (figure 1) were selected to be interpolated into 3 hourly data through the algorithm in MetSim [85]. The diurnal cycles of the air temperature and shortwave radiation were processed as follows: (1)Based on the latitude, longitude, and calendar of each grid, we determined each day's Sunrise and Sunset times and the day length of each grid; (2)The daily maximum and minimum temperatures were assumed to occur at 14:00 and Sunrise local time, respectively, and then the hourly air temperature values were computed using cubic spline interpolation; (3)Based on the Sunrise and Sunset times obtained in step (1), the hourly surface downwelling shortwave radiation is evenly assigned throughout the day. Before conducting the baseline and two warming simulations, a present-day simulation was run for 80 model years. The output for the last 10 years was compared to satellite observations from GFED4.1 s to assess the model performance in simulating fire activities. The baseline and two warming simulations were initiated from initial data provided by the NCAR, and the cycles were run for 100 and 200 model years, respectively, to obtain their quasi-equilibrium states in the terrestrial ecosystem. We simulated both the present-day and baseline simulations because of the availability of CRUNCEP and GSWP3 forcings. The output for the last 20 years was used to derive the future changes in fire activities. All simulations were run at a 1.9°× 2.5°spatial resolution. Figure 2 shows the structure of the study.

Assessment of the FFDIn
The VPD has a significant influence on the fire area during ignition, initial ignition rate, daily burning area, and extinction in boreal forest ecosystems [86]. Figure 3 shows that the burned fraction and biomass burning carbon emissions were positively correlated with the VPD during 2007-2016. The correlation coefficients (R 2 ) were 0.159 and 0.275 in the selected area (45°N-65°N, 60°E-135°E) and 0.422 and 0.463 in the areas between 45°N and 90°N, respectively, with p-values less than 0.01. Thus, it is reasonable to incorporate the VPD into the wildfire risk formula to improve the representativeness of the FFDI, especially in mid-to-high latitudes. Figure 4 shows that the burned fraction in the mid-to-high latitudes does have higher correlation coefficients with the FFDIn than the original model, especially in the boreal region of Asia. However, in the low latitudes, particularly in eastern Asia, the effect is limited or even worse than the original FFDI. This may suggest that the major climate elements affecting the wildfire occurrence could be different in different climate zones or different vegetation types. Therefore, we used the FFDIn formula in the mid-to-high latitudes of the Northern  . Interestingly, the model may overestimate fire activities in areas where there could be massive human interventions. This may suggest that the model should appropriately reduce fire emissions in these areas by considering the influences of human activities on wildfires, such as pre-prevention and firefighting.

Future changes in wildfire risk and activities
Under warming, future changes in climate-related-wildfire risk will probably vary in different areas and seasons due to inhomogeneous climate changes. Figure 6 shows the ensembled future changes in the FFDIn under the two warming scenarios. To show the solidity or consistency of the results, the signal-to-noise ratio (S/N), defined by the ratio of the mean absolute value of future change (S) and the standard deviation (N) across experiments, is used here. The highest increase in wildfire risk is projected to occur in semi-arid areas, that is, the southwestern US, Arabian Peninsula, Brazilian Highlands, South Africa, Spain, northwestern edge of Africa, Sahel, and western Australia (figures 6(a)-(c)). Although the change in the FFDIn results is comparatively small at high latitudes, the relative change is large in the area north of 60°N, Greenland, and southeastern Asia (figures 6(d) and (e)). The percentage of FFDIn changes increases more under the 2.0°C scenario than the 1.5°C scenario, especially in the area north to 60°N, eastern US, Turkey, Brazilian Highlands, and South Africa (20%, figure 6(e)), which may be caused by the projected decrease in precipitation in these areas (Figures not  shown). The increase in vegetation coverage could contribute to the increase in wildfire risk in eastern Canada  [87] and areas north of 60°N in Eurasia. Wildfire risk is projected to decrease by a very small magnitude only in northwestern South America, Central Africa, and the Indian Peninsula (figures 6(d)-(f)). Generally, areas with high wildfire risk show a northward expansion tendency, which is consistent with the projected northward expansion of fire-adaptive vegetation [88].
Considering that future changes in climate-related-wildfire risk vary from region to region, six areas from the tropical to subarctic zone were selected to investigate the seasonal variations in FFDIn changes (figure 7). The squares in figure 6(d) (d)). This suggests a sudden increase in wildfire risk during the winter season in eastern Europe.
To explore the responses of wildfire risk to the individual climate parameters in the FFDIn formula, the sensitivity levels in the selected areas were calculated. The sensitivity level of the FFDIn to a specific climate parameter is derived by holding the climate parameters at their baseline level, except for the specific one. Figure 8 shows that the cumulative FFDIn in northeastern Asia and eastern Europe is most sensitive to the VPD, suggesting the important effect of an air moisture deficit on wildfires in forest areas at high latitudes. Changes in daily maximum temperature tended to have the largest impact on the cumulative FFDIn in the Congo Basin, eastern US, and Amazon. The cumulative FFDIn in the Brazilian Highlands was most sensitive to daily minimum relative humidity, especially in the 2.0°C scenario.   Figure 9 shows the future changes in burned area and fire carbon emissions simulated by CLM4.5-BGC under the 1.5°C and 2.0°C scenarios. The future changes in burned area fractions and carbon emissions in the Northern Hemisphere are generally larger than those in the Southern Hemisphere, and the differences are more prominent when the temperature rise reaches 2.0°C. The zonal average of fire carbon emissions is projected to decrease in tropical areas but increase in northern high latitudes. The zonal average of the burned area is projected to have a similar tendency to that of fire carbon emissions, although higher variability will occur along the latitudinal direction. This may be due to the distribution of forests and grasslands. Forest areas can emit more carbon during wildfires than grassland areas with the same burned extent [89]. Compared to the present conditions, the future fire burned area and carbon emissions are projected to decrease in the Amazon, Argentina, Congo Basin, Indo-China Peninsula, eastern US, and northeastern East Asia, and increase near 60°N in Canada and Eurasia, especially under the 2.0°C scenario. The projected decrease in fire carbon emissions in some tropical areas is projected to be stronger under 2.0°C warming than 1.5°C, although the decrease in burned area is weaker.

Future changes in burned area and carbon emissions
Except for the Congo Basin and Brazilian Plateau, the burned area fractions and fire carbon emissions in the other selected regions were relatively small (figure 10). The burned area and carbon emissions simulated by the CLM4.5-BGC in the Congo Basin are projected to decrease the most in MAM, followed by SON. The largest increase in burned area fractions occurred in the spring in the Brazilian plateau, while the largest increase in carbon emissions occurred in the summer in eastern Europe. The burned area and carbon emissions in the midto-high latitudes of East Asia and the eastern US are projected to decrease throughout the year, while those in eastern Europe are projected to increase. The future changes in fire activities projected by CLM4.5-BGC generally show common trends with the projected future changes in the FFDIn. In particular, the Congo Basin is projected to have lower wildfire risk and activities and the Brazilian Highlands and eastern parts of Europe are projected to have higher wildfire risk and activities. However, some differences were observed. Northeastern Asia and eastern US are projected to have higher wildfire risks but lower fire activities. The Amazon is projected to have the same trends under the 1.5°C warming level but opposite trends under the 2.0°C warming level. This inconsistency may be caused by the different effects considered in the CLM4.5-BGC model and the FFDIn formula, suggesting a complex effect of biome mass in these areas.

Conclusions and discussion
Wildfires have a significant influence on the evolution of land surface characteristics and carbon cycle processes, and these events are closely correlated with environmental conditions and biomass fuel supplies. Here, we demonstrate that the atmospheric moisture deficit is strongly correlated with observed fire activities. The incorporation of the VPD into the FFDIn formula could improve the applicability of this formula in representing real fire occurrences in mid-to-high latitudes. Both the FFDIn and CLM4.5-BGC were used to investigate future changes in wildfire risk and activities under the 1.5°C and 2.0°C warming scenarios relevant to the Paris temperature targets.
Consistent with previous studies [42,45,57,90], the ensemble results of the FFDIn show that the semi-arid areas in southwestern North America, Arabian Peninsula, Iran, Brazilian Highlands, South Africa, and western Australia are projected to have higher probabilities of wildfire occurrence in the warming scenarios than present-day. The increased future wildfire risk increases the possibility of more intense wildfire activity, thereby increasing human fatalities and property loss. The projected increase in wildfire risk in these areas could become larger in the 2.0°C scenario than in the 1.5°C scenario. Although the absolute change of the FFDIn is small in the high latitudes, its percentage of change is large, particularly in Greenland and areas north of 60°N. This percentage increase of the FFDIn is particularly evident in the winter season followed by spring.
Under the warming scenarios, CLM4.5-BGC projects that the burned area and fire carbon emissions could decrease in the eastern US, the Brazilian Plateau, Argentina, the Congo Basin, the Indo-China Peninsula, and northeastern East Asia, while increasing in area around 60°N in Eurasia and Canada, especially in the 2.0°C scenario. Both the FFDIn results and CLM4.5-BGC simulations indicate that wildfire risk and carbon activities, including burned area and carbon emissions, will decrease in the Congo Basin, especially in the MAM season, but increase in the Brazilian Highlands, particularly in spring. Wildfire risk is projected to increase mostly due to the increase in temperature and thus the atmospheric moisture deficit in the mid-to-high latitudes of Eurasia, where the forest is located. However, the burned area and carbon emissions in northeastern Asia are projected to decrease.
Compared to the FFDIn, which considers the effect of meteorological conditions only, the CLM4.5-BGC considers the effects of lightning, biomass fuel, and human disturbance [79]. Further and additional work should be done in order to raise awareness about the relationships between meteorological and man-made fires. Although the FFDIn formula could improve the applicability of this formula in representing real fire occurrences in mid-to-high latitudes, it would be appropriate to collect more data on fire incidents, including other factors of topography, such as elevations, vegetation type, and fuel load. This would facilitate more robust findings that would help to improve the function of wildfire danger index.
The different future change directions implied by the FFDIn versus the results of the CLM4.5-BGC either illustrate the possible non-meteorological effect on wildfire activities or the inconsistency between the model and the straightforward formula of the FFDIn. Additionally, both the results from the FFDIn and CLM4.5-BGC projects indicate that the area in the northern high latitudes could experience more fire activities. This may further affect climate change locally and remotely, which requires attention.
Results of this study are subject to limitations from several sources. Firstly, the fuel characteristics play important roles in the wildfire risks and activities. Incorporation of the fuel characteristics may improve the performance of FFDIn further. Secondly, human interference could influence the carbon activities particularly in areas with large population distribution. Besides, results here rely much on the performance of the unique wildfire risk index and model. Uncertainties of the results could be reduced by using the ensemble of multiple wildfire risk indexes and models.