Introduction

The Arctic has warmed three times faster than the rest of the world during 1971–20191, with parts of Siberia warming even faster, making it one of the most rapidly warming areas in the world2. Warming in this region poses a potential climate risk due to the large stores of carbon contained in its ecosystems, peatlands, and permafrost soils3. During 2019 and 2020 unusually large wildfires in the Siberian Arctic burned almost half of the total area that had burned in the previous four decades4,5, releasing an amount of carbon (in CO2 equivalent terms) comparable with the emissions of France in 20196. Arctic peatlands have historically acted as a carbon sink, but increased fire frequency and permafrost melting could convert them into a source during the coming century7,8.

Globally, the area burned by fires has declined by around 25% over the period 1997–2013 according to satellite observations9. This is attributed to an increase in the proportion of land being used for agriculture, on which fires are more heavily supressed. Despite the burned area decline, the total carbon emitted from fires has remained stable, at around 2 Pg C year−1, due to forest accounting for an increasing fraction of burned area, which has a greater carbon emission intensity than grassland10. Future climate change is likely to exacerbate this trend, especially in the Arctic, where increasing summer temperatures will render areas of boreal forest increasingly vulnerable to fire11,12. The frequency of peatland fires, which have the largest CO2 emission intensity of all fire types13, is predicted to increase in the Arctic boreal region14. Fire-sourced carbon emissions could increase by a factor of 2–3 by 2100 in high latitude regions, including Siberia and Canada14,15,16.

Wildfires in the Arctic are known to worsen air quality both in their immediate vicinity and in distant countries, due to the potential for smoke to be transported large distances in the atmosphere17,18,19. Wildfires produce primary air pollutants including particulate matter (i.e., smoke), carbon monoxide and polycyclic aromatic hydrocarbons, as well as contributing to secondary air pollution by emitting volatile organic compounds and nitrogen oxides, which can lead to tropospheric ozone formation, and the formation of secondary particulate matter20,21,22. Exposure to the fine fraction of particulate matter (PM2.5), has been estimated to cause millions of deaths worldwide each year23,24,25, with epidemiological evidence showing it elevates the risk of acute and chronic cardiovascular and respiratory disease26,27. PM2.5 from different sources varies in toxicity due to difference in particle size distributions and chemical composition28. Exposure to PM2.5 sourced from wildfires has been shown to increase all-cause mortality29,30, and there is some evidence it is more toxic than PM2.5 from other sources31,32.

Figure 1 uses a geographically-weighted regression surface PM2.5 reanalysis dataset33 (GWR PM2.5) to demonstrate the increasing influence of wildfires on summer PM2.5 concentrations in the northern hemisphere high-latitudes between 1998 and 2020. Seasonal mean PM2.5 peaks during the boreal summer across much of Siberia and Canada (Fig. 1a), suggesting that in these areas wildfires make a dominant contribution to annual PM2.5. Figure 1b shows there was a positive trend of around 1.5 µg m−3 year−1 in summer (JJA) PM2.5 concentrations in parts of Siberia over the past two decades. This represents a substantial worsening of summer air quality, with parts of the Irkutsk Oblast and Sakha Republic regions of Russia moving from average summer PM2.5 concentrations of ~10 µg m−3 in 1998 to ~40 µg m−3 in 2020.

Fig. 1: Temporal patterns of PM2.5 using data from the Geographically Weighted Regression PM2.5 dataset33.
figure 1

a Average timing of the annual maximum in PM2.5 during 1998 to 2020 (b) Summer (JJA) PM2.5 trend during 1998 to 2020. Trends are calculated using the Theil-Sen trend estimator69, and insignificant trends (Mann–Kendall test, p > 0.05) are greyed out.

Full size image

We aim to quantify the health impacts of wildfire-sourced PM2.5 emitted from Arctic Council states, the eight countries that have territories within the Arctic Circle (Canada, Denmark, Finland, Iceland, Norway, Russia, Sweden and the United States), on local and distant populations. The Arctic Council is an intergovernmental forum that promotes cooperation among its members on issues such as community health and sustainable development34. Using the Community Earth System Model (CESM) with the Quick Fire Emissions Dataset (QFED)35, we estimate the contribution of wildfires in the Arctic Council states to high-latitude surface PM2.5, and quantify the public health impacts due to long-term ambient PM2.5 exposure.

Results

Fire emissions spatial patterns and temporal trends

Spatial and temporal patterns of wildfires and other fires (e.g., agricultural crop burning) in the Arctic Council states for the period 2001–2020 are shown in Fig. 2, based on QFED emissions data and satellite-based land cover classification (see Methods). In the Arctic Council nations, wildfires account for 89% of total carbon released by fires during 2001–2020, which varies annually from 82% to 94%. Most agricultural fires occur south of 55°N, while wildfire emissions extend to 68°N. The majority of agricultural fires occur at the southern perimeter of Arctic Council states, primarily around the Canada-contiguous US border and the Russia-Kazakhstan border. Agricultural fires also appear widespread in southwestern regions of Russia (North Caucasus, Southern and Central Federal districts), which may be attributable to prescribed crop residue burning36.

Fig. 2: Arctic Council wildfire emissions scenarios using data from the Quick Fire Emissions Dataset (QFED, version 2.5)35.
figure 2

a Total fire emissions during 2001–2020 from the QFED. Purple shades represent emissions from land classed as agricultural in MODIS, while orange shades represent emissions from non-agricultural land, assumed to be wildfires. b Time series showing the total fire emissions (Tg C) per year.

Full size image

Despite the wildfire emissions across the Arctic Council remaining relatively stable during 2001–2020 (Fig. 2), some regions in Russia, i.e., the Sakha Republic, Irkutsk Oblast and Krasnoyarsk Oblast, have a positive trend in annual total wildfire carbon emissions of >4% year-1, though only Krasnoyarsk Oblast has a statistically significant trend (Supplementary Fig. 1).

Wildfire-sourced PM2.5

As expected, most of the areas of elevated Arctic wildfire-sourced PM2.5 lie within the borders of the Arctic Council states (Fig. 3). On average, wildfire PM2.5 accounts for 21% of annual mean PM2.5 from all sources within Arctic Council states. In Russia and Canada, wildfires contribute 22.8 and 27.1% of annual PM2.5 on average respectively, while in Scandinavian countries wildfires contribute only ~2%. The areas with the largest wildfire-attributed PM2.5 occur in Russia, especially around the Chinese and Mongolian borders, where wildfires add more than 8–10 µg m−3 (40–50%) to annual mean PM2.5 (Fig. 3). In more remote parts of Siberia and Canada, wildfire-attributed PM2.5 accounts for over 70% of annual mean PM2.5. China and Mongolia also contain regions where mean PM2.5 concentrations are elevated by Arctic wildfires, suggesting that they regularly suffer from episodes of enhanced pollution due to transboundary transport of Arctic wildfire plumes.

Fig. 3: Average (during 2001 to 2020) PM2.5 increase attributed to wildfires in Arctic Council states.
figure 3

a showing the percentage of PM2.5 attributed to Arctic wildfires and (b) showing the absolute concentration of wildfire-sourced PM2.5. In both panels data is only shown where there is a significant increase in PM2.5 attributed to Arctic Council wildfires.

Full size image

In most countries, the proportion of population-weighted mean PM2.5 attributed to wildfires is lower, as wildfires tend to occur in less densely populated areas, while the majority of PM2.5 exposure occurs in urban areas. Canada and Russia have the highest proportion of their population-weighted mean PM2.5 attributed to Arctic wildfires, at 8.7 and 8.2% respectively. Mongolia has the highest population-weighted mean PM2.5 attributed to wildfires outside the Arctic Council (Table 1), as its capital city is close to the area of regular Siberian wildfires.

Table 1 Mean and standard deviation (2001–2020) of PM2.5 concentrations (µg m−3) averaged by country
Full size table

As shown in Fig. 1b, summer (JJA) PM2.5 concentrations have been increasing over the period 1998–2020 across much of Siberia, while slightly decreasing in other parts of the Arctic. During this 22-year period, some areas of Siberia show a summer PM2.5 trend of >1 µg year-1. Our simulations show that the proportion of wildfire-sourced PM2.5 has not changed on average across the Arctic Council states, but has increased by >3% year-1 in parts of Siberia (Supplementary Fig. 2). The area in Fig. 1b that shows the most rapid increase in wildfire PM2.5 is to the north of Lake Baikal, where a significant increase in fire burned area has been detected during the past two decades37.

Arctic Council states show different time-evolution of wildfire-sourced PM2.5 in populated regions. Figure 4 shows the change in the area-weighted mean and population-weighted mean concentrations of wildfire-sourced PM2.5 across each of the Arctic Council states. Where the population-weighted mean PM2.5 is larger than the area-weighted mean (e.g., for Norway and Greenland), this indicates the wildfire smoke has a disproportionately elevated impact on populated areas. When the opposite is true (e.g., for Sweden, Russia and Canada), wildfire smoke increases PM2.5 concentrations more in less densely populated regions. In Russia, the population-weighted mean PM2.5 decreases, while the area-weighted mean PM2.5 increases over time. This indicates that despite wildfire PM2.5 concentrations increasing on average during 2001–2020, the impact on populated areas in Russia has decreased.

Fig. 4: Wildfire-sourced PM2.5 averaged across each of the Arctic Council states.
figure 4

The PM2.5 average weighted by area (blue) and population (red) is shown. The thin lines show the annual means and the thick lines are a smoothed fit.

Full size image

Our analysis indicates a shift in the latitude of wildfire emissions and PM2.5 over time in some regions of the high latitudes. Figure 5 shows that areas of wildfire-sourced PM2.5 have shifted northwards in much of eastern Russia. As this area has a high proportion of wildfire-sourced PM2.5 (Fig. 3a), this indicates a shift in the wildfire regime, with wildfires occurring at higher latitudes during the 2011–2020 period compared to 2001–2010, which has been noted in ref. 38.

Fig. 5: Comparison of 2001–2010 and 2011–2020 Arctic wildfire-attributed PM2.5.
figure 5

a, b show the mean Arctic wildfire-attributed PM2.5 during 2001–2010 and 2011–2020, respectively. Green circles are added to show populated places data from Natural Earth (https://www.naturalearthdata.com/downloads/10m-cultural-vectors/10m-populated-places/ (last accessed 13/05/2023)), with circle size determined by relative population count. In (c), the wildfire-sourced PM2.5 averaged by latitude for each time period are shown.

Full size image

Wildfire-attributed mortality

Our modelling attributes 21,100 (95% confidence interval (CI): 15,800–27,700) excess deaths to Arctic Council wildfire-sourced PM2.5 on average each year, with over a third of the health impact (8100 (95% CI: 5900−10,800) mortalities) occurring in nations outside the Arctic Council (Table 2; Fig. 6). The country that experiences the single largest mortality burden is Russia, with 9900 (7300–13,100) mortalities on average per year (Table 2). In per capita terms, Russia’s health impact is also the largest, with around 69 (51–92) annual excess deaths per million people attributed to wildfire PM2.5 on average. Among the Arctic Council states, Canada also has a high per capita impact, with 38 (31–46) Arctic wildfire-attributed mortalities per million. Other countries with a high per capita health impact are those bordering or close to Russia, including Mongolia (39 (26–57)), Belarus (26 (18–38)) and Georgia (19 (14–25) mortalities per million).

Table 2 Annual PM2.5-attributed excess deaths averaged by country averaged over the 2001–2020 period
Full size table
Fig. 6: Time series of the additional mortalities attributed to Arctic Council wildfires and Arctic wildfire-sourced annual mean PM2.5 concentrations.
figure 6

The time series are divided into Arctic Council states (blue) and the rest of the northern hemisphere (orange). The excess deaths are shown as coloured bars with grey error bars showing the upper and lower confidence intervals estimated from the health impact model. The Arctic wildfire-sourced annual mean PM2.5 concentrations are shown as lines.

Full size image

Although only 0.2% of China’s population-weighted mean PM2.5 is attributed to Arctic wildfires, it has the second largest mortality burden of 4800 (3500–6400) mortalities attributed to Arctic wildfires on average per year. This is due to its proximity to Siberia; wildfire smoke plumes are advected over north-eastern China, which is a relatively densely populated region, resulting in a larger exposed population and a larger associated mortality burden.

The number of mortalities attributed to Arctic wildfires declines during the period of our study (Fig. 6), with 22,900 (17,300−30,000) excess deaths on average during 2001–2010, and 19,200 (14,200-25,400) during 2011–2020 (Table 2). This is despite the positive trend in summer PM2.5 concentrations in some parts of the Arctic shown in Fig. 1. The majority of the decreased health burden occurs in non-Arctic Council states, where Arctic wildfire-sourced PM2.5 concentrations have declined, whereas within Arctic Council states annual wildfire-sourced PM2.5 concentrations and wildfire-attributed mortalities remain stable over time. Years with larger health impacts outside the Arctic Council are caused by wildfires occurring in areas bordering populous regions of non-Arctic Council states.

Discussion

Wildfires account for the majority of annual mean PM2.5 across much of the Canadian and Siberian Arctic (Fig. 3a). Due to the sparsely populated nature of these areas, the health impact is relatively low, accounting for <2% of PM2.5 attributed excess deaths in all countries except Canada (9%), Russia (6%) and Mongolia (3%). The largest health impacts in absolute terms occur when plumes of wildfire pollution are advected over densely populated areas. However, the populations who experience the largest impacts on a per capita basis from Arctic wildfires live in the remote areas immediately around fires (Supplementary Fig. 3). In some countries, especially Russia, our modelling indicates that the health impact is exacerbated further due to the high underlying baseline rates of mortality.

Our results demonstrate that PM2.5 pollution sourced from wildfires inside the Arctic Council states have a substantial transboundary health impact on remote populations, with over a third of Arctic wildfire PM2.5-attributed excess deaths occuring in non-Arctic Council states. However, the transboundary impacts appear to have decreased during the last two decades, as Siberian wildfires shift to higher latitudes, away from the Mongolian and Chinese borders with Russia.

Burned area in boreal forests is positively correlated with fire weather index39. Fire weather frequency and severity has increased in boreal regions during recent decades40. It is not clear to what extent climate change has driven this apparent trend, and therefore whether it could be expected to persist in the future, especially as the response of wildfires and underlying vegetation to climate are highly uncertain41,42. It remains to be seen whether the northwards shift in burning will continue over the next years and decades. More work is needed to establish which drivers have led to the strong positive trend in wildfire-sourced air pollution in some parts of Siberia during the past two decades, and to establish how likely the anomalously large Siberian fires of 2019–20 are to be repeated in the context of a shifting climate.

Our health calculations use the GEMM23, based on from data 41 epidemiological studies to establish exposure-response curves between annual mean PM2.5 concentrations and mortality. These exposure-response curves likely best represent the regions from which most of epidemiological data is collected, i.e., urban regions outside the Arctic. Arctic populations likely have different lifestyle factors (e.g., time spent outdoors, activity rates, public transport use) that may result in a different relationship between ambient air pollutant concentrations and personal exposure. Our population age structure and baseline mortality data are at the national level, so any regional variation in population demographics within countries is not included in this analysis. Sub-national level baseline health data and Arctic wildfire specific PM2.5 exposure-response health functions could improve our estimates in future work. Despite these potential limitations, our study provides the first estimate of wildfire health impacts across the Arctic Council states, to the best of our knowledge, based on available information and modelling analyses.

Our simulated PM2.5 concentrations have a negative bias when compared to the few available observations within the Arctic (see Section “Correction of simulated PM2.5 to observations”), which introduces uncertainty into our estimation of the fire fraction of PM2.5, which could further propagate into our health impact estimation. Lack of Arctic PM2.5 observations, particularly those of speciated PM2.5 makes it difficult to quantify uncertainty in our fire fraction estimate that would propagate into our mortality estimates. However, it is well-established that models tend to underestimate aerosol concentrations in areas of high biomass burning43,44,45, it is likely that our mortality estimates are conservative, despite them being larger than those estimated for high-latitude nations in a previous global study, ref. 46. McDuffie et al. 46 provide by-country totals of their mortality attributed to non-cropland fire emissions in 2017 and 2019, estimated using the GEMM model. We exclude the United States from the comparison, as in our study we do not remove wildfire emissions from the contiguous-United States. For the remaining Arctic Council nations, in 2017 and 2019 our mortality estimates are a factor 2–3 times larger, although we note both chemical transport modelling and fire emissions differ between the two studies. McDuffie et al. 46 use the Global Fire Emissions Dataset version 447, which has been shown to underestimate burned area in the high latitudes48.

The northwards shift in fire emissions and air quality impact in Siberia has caused the overall PM2.5-attributable mortality impact of wildfires to decrease, despite the increase in fire emissions in this area4, mainly due to the reduced impact on relatively densely populated regions such as northeast China and Mongolia. Although this work focuses on the human health impact, shifting Arctic fire regimes will also affect ecosystems that are already vulnerable due to climate change49,50, and could accelerate the release of carbon stored in tundra ecosystems.

Methods

Model setup

To estimate the impact of Arctic Council wildfires on human mortality, we simulate global PM2.5 under two scenarios:

  • ‘FIRE ON,’ a control simulation using the Quick Fire Emissions Dataset (QFED, version 2.5)35 fire emission dataset, which includes daily fire emissions from all fires (wildfire and agricultural) and all regions.

  • ‘ARCTIC WILDFIRE OFF,’ a counterfactual scenario without wildfire emissions within Arctic Council states, but including other types of fires in the Arctic, and all types of fire outside the Arctic. In this scenario wildfires resulting from both human and natural ignitions are removed.

The eight Arctic Council member states are; Canada, Denmark, Iceland, Norway, Sweden, Finland, Russia and the United States. As we are focussed on the impact of wildfires in high latitude regions, we do not include the contiguous United States or Denmark (but do include Greenland) as Arctic Council areas.

The simulations are performed with the Community Earth System Model (CESM) version 2.251 in a configuration which includes the Community Atmosphere Model with chemistry (CAM-Chem) to simulate tropospheric and stratospheric chemistry, and simplified representations of other components of the Earth (e.g., oceans, sea-ice etc.). The version of CAM-chem in CESM v2.2 is CAM6-chem, which uses a recent version of the Model for Ozone and Related chemical Tracers (MOZART) chemical mechanism, MOZART-TS152, to represent gaseous species, and a Volatility Basis Set scheme to represent ageing of organic aerosol. The aerosol size distributions are represented using the four-mode version of the Modal Aerosol Model (MAM4)53, which added an additional mode to improve representation of carbonaceous aerosols, black carbon (BC) and primary organic matter (POM). We run CESM at the standard grid resolution of 0.9° by 1.25° (latitude by longitude), with the meteorology nudged using MERRA2 reanalysis data.

Due to coupled chemical and meteorological processes in CESM, changes in the fire emissions between the two scenarios result in small perturbations of meteorological parameters (e.g., temperature, wind speeds), which will in turn affect secondary PM2.5 concentrations, which can propagate to regions distant from the Arctic. These fluctuations could be misinterpreted as being directly caused by wildfire-sourced PM2.5 when comparing the two scenarios. To focus on the direct impacts of wildfire, we calculate wildfire-attributable PM2.5 as the difference in the POM and BC fractions between the two scenarios, which are not affected by secondary aerosol formation or meteorological feedbacks. Since these species represent on average >99.9% of primary aerosol mass emitted by Arctic Council wildfires, this method covers the large majority of the expected driver of PM2.5-attributable health impacts.

We also exclude areas where the increase in the POM + BC PM2.5 in the FIRE ON scenario is statistically insignificant when compared with the ARCTIC WILDFIRE OFF scenario. We do this by performing a one-sided paired samples t test at each CESM grid cell between the monthly mean concentrations in the two scenarios during the 2001–2020 period, with a significance threshold of p = 0.01. If no significant increase is found, we set the ‘FIRE FRACTION’ term to zero in the bias correction equation (Section “Correction of simulated PM2.5 to observations”), resulting in no health impact attributed to wildfires at that location. This ensures that statistically insignificant differences between the two scenarios in more populated regions distant from the Arctic do not unduly contribute to the calculated health impact. The result of the t test are shown in Supplementary Fig. 4.

Developing a ‘wildfire off’ emissions scenario

Both CESM model simulations include fire emissions from the QFED dataset35. In QFED, emissions of gaseous and particulate species released by biomass burning are based on satellite-detected fire radiative power (FRP) observations, along with biome-specific emissions factors that are calibrated by comparing modelled and observed aerosol optical depth (AOD). We chose QFED over other biomass burning emission inventories (BBEIs) as its use of AOD observations can reduce emission underestimation35,45, and it has been shown to have the lowest negative bias against AOD among six commonly used BBEIs44, so it is less likely to underestimate PM2.5 concentrations, which are the primary focus of this analysis.

However, QFED emission estimates do not discriminate between agricultural and wildfires. Therefore, to construct a biomass burning emissions scenario without wildfires, we assume that emissions from agricultural land are not wildfires, and emissions from other land uses are wildfires.

The resolution of the QFED emissions used has been regridded to CESM resolution (0.9° by 1.25°) while the MODIS land use data is available at 0.01° by 0.01°. Using MODIS data at its native spatial resolution (Supplementary Fig. 5), we calculate the proportion of agricultural land in each QFED grid cell, which is used to fractionally portion the QFED emissions between wildfire and agricultural emissions. The wildfire and cropland attributed emissions are shown in Supplementary Fig. 6. Any error introduced due to the resolution mismatch between QFED and MODIS data will have a relatively small impact on our analysis due to the small fraction of land area with cropland cover in the Arctic Council, most of which is grouped into discrete areas.

Correction of simulated PM2.5 to observations

Comparisons with the few available PM2.5 observations in the Arctic region show that simulated air pollutant concentrations are persistently underestimated by both regional54 and global models55,56,57. This is partly due to the difficulty of representing a smoke plume at low resolutions, where the plume will be diluted within a model grid cell. Additionally, biomass burning emissions are often underestimated43,44,47. A particular cause of this in the Arctic region could be the peat soils, which frequently experience higher emission smouldering fires, typically not well represented in emission inventories, as they are difficult to detect via satellites58,59. Additionally, burned area products derived from satellites can often omit fires due to high detection limits and limited overpass times60.

The CESM model was compared with a dataset of reference grade PM2.5 monitors from the AIRNOW network (north of 55°N), which includes monitors in Alaska, Canada and the US embassy in Almaty, Kazakhstan. CESM has a negative Normalised Mean Bias (NMB) of −0.61 whereas Geographically Weighted Regression (GWR) PM2.533 has NMB of −0.05 (Supplementary Fig. 7). A similar result was found when comparing the CESM model with a dataset from low-cost ‘PurpleAir’ monitors that are located north of 55 N. CESM is negatively biased compared with the PurpleAir data (Supplementary Fig. 8), with a NMB of −0.72. The GWR PM2.5 product is in closer agreement with the PurpleAir monitors with a NMB of 0.1. From these results, we conclude that GWR PM2.5 is much better able to represent the magnitude of Arctic PM2.5 concentrations than the CESM-simulated PM2.5 concentrations.

We use the GWR PM2.5 to gain a more accurate estimate of PM2.5 concentrations in areas affected by wildfires. This dataset uses satellite AOD observationsalong with model-derived vertical profiles and surface monitoring measurements to estimate monthly mean PM2.5 at a resolution of 10 km. However, the FIRE ON simulation underestimates compared with GWR PM2.5 in most areas and seasons, not just those affected by wildfire plumes. Therefore, using the ARCTIC WILDFIRE OFF CESM simulation as a counterfactual scenario to be compared with GWR PM2.5 would overestimate the fraction of PM2.5 that can be attributed to Arctic wildfires.

We use our CESM simulations to estimate the proportion of PM2.5 that is attributable to Arctic Council wildfires, which we term the ‘FIRE FRACTION’ (Eq. 1). We apply the FIRE FRACTION to the GWR PM2.5 data to estimate a counterfactual PM2.5 scenario without Arctic wildfires, that we term the ‘BIAS CORRECTED WILDFIRE OFF’ scenario (Eq. 2).

$${{{{{\boldsymbol{FIRE}}}}}}\,{{{{{\boldsymbol{FRACTION}}}}}}=\frac{{{{{{\boldsymbol{FIRE}}}}}}\,{{{{{\boldsymbol{ON}}}}}}-{{{{{\boldsymbol{ARCTIC}}}}}}\,{{{{{\boldsymbol{WILDFIRE}}}}}}\,{{{{{\boldsymbol{OFF}}}}}}}{{{{{{\boldsymbol{FIRE}}}}}}\,{{{{{\boldsymbol{ON}}}}}}}$$
(1)
$${{{{{\boldsymbol{BIAS}}}}}}\,{{{{{\boldsymbol{CORRECTED}}}}}}\,{{{{{\boldsymbol{WILDFIRE}}}}}}\,{{{{{\boldsymbol{OFF}}}}}}= {{{{{\boldsymbol{GWR}}}}}}\,{{{{{{\boldsymbol{PM}}}}}}}_{{{{{{\bf{2.5}}}}}}}\\ \times (1-{{{{{\boldsymbol{FIRE}}}}}}\,{{{{{\boldsymbol{FRACTION}}}}}})$$
(2)

The GWR PM2.5 reanalysis only extends to 68°N, so at more northerly latitudes the FIRE ON CESM simulation is used. To avoid a hard boundary, we linearly interpolate the GWR PM2.5 with the FIRE ON simulation north of 68°N.

Health impact assessment

The health impact of long-term (chronic) exposure to ambient PM2.5 concentrations is estimated using the Global Exposure Mortality Model (GEMM)23 following the method used in previous work61,62,63, and other assessments of wildfire smoke health impacts46,64. Briefly, using data from 41 epidemiological cohort studies, the GEMM estimates increased relative risk of health impacts from chronic ambient PM2.5 exposure above a counterfactual level of 2.4 µg m−3 for adults aged 25 years and older in five-year age groups. We used the relative risk function for non-accidental mortality (including non-communicable disease and lower respiratory infections) and used the parameters that include the China cohort23. We use the GEMM because a model that specifically considers the chronic health impact of wildfire smoke is currently not available in the literature. We consider chronic health impacts because across much of the high-latitudes wildfire smoke regularly degrades regional air quality and accounts for a high proportion of annual mean PM2.565,66.

The Gridded Population of the World (GPW) dataset version 4 was used for population count and distribution67. The GPW dataset contains 5-yearly population estimates, which were linearly interpolated to provide the intervening years. The Global Burden of Disease 2019 study data68 was used for annual population age structure and baseline mortality rates from 2001–2019, with 2020 health impacts calculated using 2019 data.

We use the GEMM to calculate the excess deaths due to chronic PM2.5 exposure in both the control scenario (GWR PM2.5) and the counterfactual scenario with Arctic Council wildfire pollution removed (“BIAS CORRECTED WILDFIRE OFF”, Eq. 2). The difference between the excess deaths under these two scenarios is the mortality burden we attribute to Arctic wildfire smoke. Trend estimations

Trends are calculated using the Theil-Sen trend estimator69, which is a non-parametric trend estimator that is robust to outliers. Trends were tested for significance using the Mann–Kendall test that detects monotonic increasing or decreasing trends70.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.