Air Pollution From Forest and Vegetation Fires in Southeast Asia Disproportionately Impacts the Poor

Abstract Forest and vegetation fires, used as tools for agriculture and deforestation, are a major source of air pollutants and can cause serious air quality issues in many parts of Asia. Actions to reduce fire may offer considerable, yet largely unrecognized, options for rapid improvements in air quality. In this study, we used a combination of regional and global air quality models and observations to examine the impact of forest and vegetation fires on air quality degradation and public health in Southeast Asia (including Mainland Southeast Asia and south‐eastern China). We found that eliminating fire could substantially improve regional air quality across Southeast Asia by reducing the population exposure to fine particulate matter (PM2.5) concentrations by 7% and surface ozone concentrations by 5%. These reductions in PM2.5 exposures would yield a considerable public health benefit across the region; averting 59,000 (95% uncertainty interval (95UI): 55,200–62,900) premature deaths annually. Analysis of subnational infant mortality rate data and PM2.5 exposure suggested that PM2.5 from fires disproportionately impacts poorer populations across Southeast Asia. We identified two key regions in northern Laos and western Myanmar where particularly high levels of poverty coincide with exposure to relatively high levels of PM2.5 from fires. Our results show that reducing forest and vegetation fires should be a public health priority for the Southeast Asia region.

Vietnam; also referred to as the Indochina Peninsula or Peninsula Southeast Asia) and south-eastern China, which have been much less studied.
In Southeast Asia, fires are used as a tool for agricultural management, for example, to remove agricultural residues (mainly from rice and sugarcane cultivation) and weeds, and for forest clearance for agricultural purposes (Biswas et al., 2015;Chen et al., 2017;Phairuang et al., 2017). Fires in Mainland Southeast Asia mainly occur during the premonsoon season (roughly February to April), due to widespread forest fires and crop residue burning in preparation for planting at the Asian summer monsoon onset (W.-R. Huang et al., 2016;Phairuang et al., 2017). The increased fire activity coincides with a widespread stable temperature inversion layer over Thailand, Vietnam, Laos, and Southern China (Nodzu et al., 2006) with hot, dry, and stagnant air over northern Thailand (Kim Oanh & Leelasakultum, 2011) promoting haze conditions. During the burning season, long-range transport of smoke from fires in Mainland Southeast Asia has been observed in Southwest China (Zhu et al., 2016), south-eastern Tibetan Plateau (Sang et al., 2013), Southern China, Taiwan, and Hong Kong (K. Huang et al., 2013). Fires reduce substantially after the onset of the summer monsoon rainfall (in late April) and are minimal until around the start of the dry season (in November). Fires in this region display a degree of interannual variability linked to changes in atmospheric circulation features, such as the India-Burma Trough (W.-R. Huang et al., 2016).
Several studies have used a mix of models and observations to explore the impacts of fire on atmospheric aerosol properties, visibility, and/or air quality in Mainland Southeast Asia (Duc et al., 2016;K. Huang et al., 2013;Lee et al., 2017Lee et al., , 2018Li et al., 2017;Lin et al., 2013;Vongruang & Pimonsree, 2020;Yin et al., 2019). However, studies quantifying the contribution of fire to particulate air pollution, population exposure and public health are lacking in this region (H. J. Johnston et al., 2019;Yadav et al., 2017), compared in particular to the large number of studies focused on Equatorial Asia (e.g., Crippa et al., 2016;Kiely et al., 2020;Koplitz et al., 2016;Marlier et al., 2012). Recent studies show that fire is the dominant cause of the variation of local ambient air quality in Mainland Southeast Asia (Yin et al., 2019); contributing 49% of ambient PM 10 (particulate matter with aerodynamic diameter ≤ 10 μm) concentrations during peak open burning in March 2012 (Vongruang & Pimonsree, 2020) and 70%-80% to aerosol optical depth in source regions during March-April 2013 . Preventing fire could yield substantial reductions in population-weighted PM 2.5 (particulate matter with aerodynamic diameter ≤ 2.5 μm) concentrations across Mainland Southeast Asia (Reddington, Conibear, et al., 2019). There are large uncertainties associated with quantifying and simulating particulate emissions from fire in tropical regions (Reddington et al., 2016). In Mainland Southeast Asia, there is a large range in emissions estimates (Kaiser et al., 2012;Lasko et al., 2017;Shi & Yamaguchi, 2014;Sornpoon et al., 2014;Phairuang et al., 2017;van der Werf et al., 2017;Wiedinmyer et al., 2011) and varying performance when tested in models against observations (Fu et al., 2012;Lee et al., 2017;Pimonsree et al., 2018;Reddington et al., 2014Reddington et al., , 2016Takami et al., 2020;Vongurang et al., 2017). Emissions from the Fire Inventory from NCAR (FINN; Wiedinmyer et al., 2011) have been used widely in models over this region; with simulated PM concentrations showing good agreement against observations in some studies (Reddington et al., 2014(Reddington et al., , 2016Takami et al., 2020), but overestimation by a factor of ∼2 in others Pimonsree et al., 2018;Vongurang et al., 2017). Fires also impact ozone concentrations, being a source of ozone precursors and altering photochemistry, impacting ozone production (Jaffe & Wigder, 2012). The efficacy of photochemical ozone production in fire plumes is highly variable and uncertain, and is affected by nonlinear ozone dependence on changes in precursor concentrations, and high particulate loadings, which affect photochemistry (Jaffe & Wigder, 2012). Fires have been shown to enhance regional ozone concentrations in Mainland Southeast Asia (Pochanart et al., 2001) and aloft over southern China (C. Y. Chan et al., 2003;L. Y. Chan et al., 2000;Kondo et al., 2004), although fires have also been implicated in suppressed ozone in some situations (Deng et al., 2008).
Links between socioeconomic factors, population exposure to ambient air pollution, and associated health effects have been well documented in parts of North America and Europe (e.g., Fairburn et al., 2019;Hajat et al., 2015). However, few studies have focused on countries in Southeast Asia, with some demonstrating strong connections between ambient air pollution and poverty, for example, in urban areas of Laos (S. Dasgupta et al., 2005), rural areas of Vietnam (Narloch & Bangalore, 2018) and Ho Chi Minh City (Mehta et al., 2014); and others finding only weak connections, for example, in Cambodia and Vietnam (S. Dasgupta et al., 2005) or no connection, for example, in Laos (Pasanen et al., 2017). The majority of these studies REDDINGTON ET AL. explored links between poverty and multiple environmental risks, including ambient air pollution from all sources. To our knowledge, no previous studies have examined the poverty levels of populations exposed to air pollution from fires in this region.
In this work, we use a combination of satellite-derived data sets of fire emissions, models and observations to quantify the contribution of forest and vegetation fires to air quality degradation and disease burden in Mainland Southeast Asia and south-eastern China. We also examine the poverty levels of the Southeast Asian population exposed to PM 2.5 pollution derived specifically from fire emissions.

Fire Emissions in GLOMAP
Fire emissions of sulfur dioxide (SO 2 ), black carbon (BC), and organic carbon (OC) were specified using three different data sets: the National Center for Atmospheric Research Fire Inventory version 1.5 (FIN-Nv1.5) (Wiedinmyer et al., 2011), the Global Fire Emissions Data Set version 4.1 with small fires (GFED4s) (van der Werf et al., 2010(van der Werf et al., , 2017 and the Global Fire Assimilation System versions 1.0 and 1.2 (GFASv1.0 and GFASv1.2) (Kaiser et al., 2012); hereafter referred to as FINN, GFED, and GFAS, respectively. The different fire emission estimation methodologies of these data sets are described in detail in their references given above and in our previous work (Reddington et al., 2016;Reddington, Morgan, et al., 2019). We use daily fire emissions from all three data sets (daily GFED emissions are available from 2003 onwards [Mu et al., 2011]). Fire emissions were distributed vertically over six ecosystem-dependent altitudes between the surface and 6 km according to Dentener et al. (2006). Over Mainland Southeast Asia, all emissions were injected below 3 km elevation, which is consistent with satellite observations of the vertical distribution of smoke in this region (Gautam et al., 2013 (NCEP, 2000(NCEP, , 2007 Aerosol size distribution Modal scheme (7 log-normal modes) Sectional scheme (MOSAIC 4-bin; Zaveri et al., 2008) Gas-phase chemistry TOMCAT (Chipperfield, 2006) MOZART-4 (Emmons et al., 2010) Simulation

Description of the WRF-Chem Regional Model
We used the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem; Grell et al., 2005) version 3.7.1, a high-resolution regional model, to simulate air pollutant concentrations for 1 year (2014) and quantify the public health impacts of long-term exposure to fire-derived PM 2.5 and ozone (O 3 ) concentrations. Table 1 summarizes the model setup used for this study; see Section S1.2 for further details.

Fire Emissions in WRF-Chem
Fire emissions were taken from FINN version 1.5 (Wiedinmyer et al., 2011), with a spatial resolution of 1 × 1 km for the year 2014. Fire emissions were included for BC, OC, PM 2.5 , PM 10 , carbon monoxide, ammonia, nitrogen oxides, SO 2 , and nonmethane volatile organic compounds (speciated according to the Model for Ozone and Related Chemical Tracers (MOZART); Emmons et al., 2010). We applied a diurnal factor (Western Regional Air Partnership, 2005) to the daily emissions, which assumes greater emissions during the day (between 10:00 and 19:00 local time, peaking at 15:00-16:00 local time) and minimal emissions during the night. The injection heights of the fire emissions were calculated online in the model using the Freitas et al. (2007) plume-rise parameterization. The plume-rise parameterization applies a 1-D cloud-parcel model to each grid-column within the WRF-Chem model domain that contains a fire.

WRF-Chem Model Simulations
The model domain is located over East Asia, using a Lambert conformal conical projection with a horizontal resolution of 30 × 30 km (covering a 130 × 124 grid) and 33 vertical levels up to a minimum pressure of 10 hPa. We regridded the model output, using linear interpolation, onto a regular latitude-longitude grid at 0.25° × 0.25° resolution. We performed three model simulations with WRF-Chem: one simulation excluding fire emissions ("WRChem_nofire"); one simulation including FINN fire emissions ("WRFChem_ FINN"); and one simulation where FINN fire emissions of OC and BC were scaled upwards by a factor 1.5 ("WRFChem_FINNx1.5"). The simulation period was for 1 year from January 9, 2014 to January 9, 2015, with the first 8 days of January 2014 run as spin-up. We selected 2014 for our simulation year since both PM and O 3 measurements are available for this year (Section 2.5).

Description of the Public Health Impact Assessment
We estimated the disease burden attributable to ambient PM 2.5 exposure (simulated by WRF-Chem) using population attributable fractions of relative risk. The relative risk of disease at a specific ambient PM 2.5 exposure was estimated through the Global Exposure Mortality Model (GEMM) (Burnett et al., 2018). We calculated the disease burden due to long-term exposure to ambient O 3 (simulated by WRF-Chem) using the exposure-response function from Turner et al. (2016). Uncertainty intervals at the 95% confidence level (95UI) were estimated through using the derived uncertainty intervals from the exposure-outcome functions, baseline mortality and morbidity rates, and population age fractions. See Section S1.3 for further details.
The mortality due to fire emissions (M FIRE ) was calculated using the "subtraction" method (Conibear et al., 2018); calculating the difference between the premature mortality from all sources (M ALL ) and the premature mortality when fire emissions have been removed (M FIRE_OFF ) as in Equation 1:

Description of the Poverty Proxy Data
As a proxy for population poverty levels, we used gridded subnational Infant Mortality Rate (IMR) estimates from NASA Socioeconomic Data and Applications Center for 2015 (Center for International Earth Science Information Network (CIESIN), 2018b). For further details see Section S1.5 and Figure S8.
The IMR is defined as the number of children who die before their first birthday for every 1,000 live births in a given year. For context, previous studies have defined populations with IMR < 15 to be not poor; 15 ≤ IMR < 32 to be moderately poor; 32 ≤ IMR < 65 to be poor; and 65 ≤ IMR < 100 to be very poor (De Sherbinin, 2008); and populations with a high IMR as having > 32 deaths per 1,000 live births (Barbier & Hochard, 2019).
Subnational IMR estimates have been used as a proxy for poverty indicators in a range of previous studies (Barbier & Hochard, 2018Barlow et al., 2016;De Sherbinin, 2008;Hauenstein et al., 2019). A strong correlation between IMR and other poverty-related metrics, including population income, education and health (De Sherbinin, 2008;Fritzell et al., 2015;O'Hare et al., 2013;Reidpath & Allotey, 2003;Sartorius & Sartorius, 2014), justifies the use of IMR as a proxy for overall poverty levels. In addition, it is difficult to obtain alternative poverty measures at sub-national levels for multiple countries (CIESIN, 2018a; P. Dasgupta, 1993). Other advantages of this data set over alternative poverty measures include its highly standardized nature and availability for ≥90% of medium-income and low-income country populations (Balk et al., 2006;CIESIN, 2018a).

Description of the Particulate Matter and Ozone Measurements
We used 2003-2015 monthly mean PM 10 concentrations measured at air quality monitoring stations located in fire-influenced regions of Thailand ( Figure S1a) from the Pollution Control Department (PCD) of the Thailand Government Ministry of Natural Resources and Environment. The fire-influenced stations were selected using GLOMAP or WRF-Chem model data where fire emissions contributed 20% or greater to the simulated annual mean PM 10 . We used surface O 3 concentration measurements from air quality monitoring stations located in China and surrounding countries ( Figure S1b) from the Berkley Earth China Air Quality Data Set (Rohde & Muller, 2015). See Section S1.4 for further details on the measurements.
To evaluate model-simulated surface PM 10 concentrations due only to the influence of fire, we calculated and compared simulated and measured fire-derived (smoke) PM 10 concentrations. The simulated and measured fire-derived PM 10 concentrations were estimated for each year separately, by subtracting the minimum monthly mean PM 10 concentration from all monthly mean concentrations for that year. A similar approach has been used in previous modeling studies (e.g., Kiely et al., 2020) to isolate enhancements in surface PM concentrations due only to fires.
To quantify the agreement between model and observations, we used the Pearson correlation coefficient (r) and normalized mean bias factor (NMBF) as defined by Yu et al. (2006). A positive NMBF indicates the model overestimates the observations by a factor of NMBF + 1. A negative NMBF indicates the model underestimates the observations by a factor of 1 -NMBF.  Across Mainland Southeast Asia, fire emissions are predominantly from deforestation/degradation fires (accounting for 31%-57%) and savanna-type fires (accounting for 35%-55%) (Figure 2b). A detailed analysis of forest fires in Myanmar confirms that most are of anthropogenic origin (Biswas et al., 2015). Vadrevu et al. (2019) found that most fires occurred in forests as opposed to cropland across much of Mainland Southeast Asia including Myanmar, Laos, Cambodia, and Vietnam. In regions with both deforestation and savanna fires, deforestation fires emit a greater amount of particulate emissions, due to a combination of larger fuel loads/biomass consumption and emission factors, and thus tend to dominate emissions (Figure 1d). However, savanna fires are more prevalent across the region and so the accumulated emissions from this fire type per country are generally comparable to or greater than deforestation fires. In south-eastern China, OC emissions arise predominantly from fires classified as temperate forest fires (67%). Agricultural fires make up a relatively small fraction of fire OC emissions across the region (1%-14%), but the occurrence of these fires may be underestimated or misrepresented both in GFED (Reddington et al., 2016;T. Zhang, Wooster, et al., 2018), and more widely by satellite-based estimates (Lasko et al., 2017;Shen et al., 2019;Stavrakou et al., 2016;L. Zhang et al., 2016;T. Zhang et al., 2020). 2007, and 2012). The multiyear GLOMAP simulations demonstrate that fires consistently make a substantial contribution to surface PM 10 concentrations in northern Thailand over a 13-year period.   Figure S2). This result is consistent with our previous work (Reddington et al., 2016) that used AERONET aerosol optical depth to evaluate the GLOMAP model over Southeast Asia. Therefore, we use the FINN emissions in our high-resolution regional model simulations in the following sections.   underestimates the magnitude of the observations (NMBF = −0.28) predominantly during January to July. This is consistent with total PM 10 concentrations ( Figure S3).

Evaluation of WRF-Chem Particulate Matter Concentrations
Increasing the particulate fire emissions by a factor 1.5 improves the overall agreement with measured fire-derived PM 10 (Figure 4; r 2 = 0.94, NMBF = −0.10). Specifically, the FINNx1.5 simulation better captures the measured seasonal variation and magnitude of fire-derived PM 10 at 11 out of 12 stations ( Figure S4; FINN: normalized standard deviation (NSD) = 0.55-1.21; FINNx1.5: NSD = 0.64-1.74), with little change in the strong temporal correlation (FINN: r = 0.83-0.97; FINNx1.5: r = 0.87-0.98). The FINNx1.5 simulation also agrees well with PM 2.5 measurements (see Section S2.2 and Figure S5). Previous studies have used similar or larger scaling factors to increase fire emissions in models to better match observations (see Reddington et al., 2016 and references therein). In the following sections, we show results from the FINNx1.5 simulation as it gives the best match to PM observations.

Evaluation of WRF-Chem Surface Ozone Concentrations
Figure 5 compares simulated and measured daily mean surface O 3 mixing ratios averaged over two regions in Southeast Asia during April-July 2014. Regional-average measured O 3 mixing ratios range from ∼10 to ∼60 ppbv. Variability in surface O 3 concentrations over Southeast Asia is driven by a complex mix of factors, including varying precursor gas emissions and concentrations, photochemical production, and meteorological effects (causing accumulation, transport and removal). We evaluate the model against total O 3 rather than fire-derived O 3 , as for total PM 2.5 in Section S2.2, because these quantities are used for the health impact assessment in Section 3.4. Measured surface O 3 mixing ratios in Thailand show a peak during April (Figure 5a), which has been reported to be due to regional scale O 3 production triggered by fires Pochanart et al., 2001). The FINNx1.5 simulation captures this peak and reproduces the general daily variability in measured O 3 concentrations (r 2 = 0.81), while slightly overestimating the magnitude of the measurements (NMBF = 0.19). In south-eastern China (Figure 5b), the model simulates the magnitude and temporal variability of the measured O 3 mixing ratios reasonably well (r 2 = 0.46, NMBF = 0.11). Model-measurement comparisons are shown for separate provinces/regions in south-eastern China in Figure S6. Previous studies have reported increased ozone concentrations aloft (∼2-6 km altitude) over southern China due to fires in Mainland Southeast Asia but show little enhancement at the surface (C. Y. Chan et al., 2003;L. Y. Chan et al., 2000;Kondo et al., 2004), consistent with the model results. Reductions in photochemical ozone production as a result of PM from fires can also act to reduce ozone concentrations (Deng et al., 2008).     Eliminating fire emissions reduces simulated annual mean surface PM 2.5 concentrations by ∼40%-70% in northern Thailand, Myanmar, Cambodia, and Laos, with reductions in south-eastern China ranging from ∼10%to 40% in the region of Mainland Southeast Asia and in Taiwan, to ≤10% in the provinces further east. Population-weighted annual mean PM 2.5 concentrations across Southeast Asia are reduced by 7%, with reductions of 20% in Cambodia, 41% in Laos, 31% in Myanmar, 23% in Thailand, and 7% in Vietnam.

Impacts of Forest and Vegetation Fires on Air Quality
Simulated PM 2.5 concentrations suggest that for 2014, the World Health Organization (WHO) Air Quality Guideline for PM 2.5 (an annual mean of 10 μg m −3 ; WHO, 2006) is exceeded in almost every location in Southeast Asia even when fires are excluded (see Figures S7a and S7b). However, excluding fires substantially reduces the population exposed to levels of PM 2.5 above the WHO Air Quality Interim Target 2 (annual mean of 25 μg m −3 ) in Thailand (by 64%), Myanmar (by 100%), Laos (by 92%), and Cambodia (by 44%), with smaller reductions in Vietnam (by 9%) and south-eastern China (by 3%). Figure 6b shows the relative change in simulated surface annual mean O 3 concentration when fire emissions are excluded from the model (see Figures S7c and S7d for absolute concentrations). The spatial pattern of relative changes in surface O 3 is fairly consistent with the peak and minimum relative changes in surface PM 2.5 concentrations, with largest reductions over northern Thailand,Myanmar,Cambodia,and Laos (up to 20%) and smaller reductions over most of south-eastern China (<15%). When fires are excluded from the model, the annual average daily maximum 8-h (ADM8h) O 3 concentration is reduced by 5% across Southeast Asia, with reductions of 10% in Cambodia, 12% in Myanmar and Laos, 8% in Thailand, 5% in Vietnam, and 2% in south-eastern China. Table 2 shows the averted disease burden due to changes in long-term exposure to ambient PM 2.5 and O 3 from eliminating fire emissions. Eliminating fire emissions reduces the annual disease burden from ambient PM 2.5 exposure by 12% in Mainland Southeast Asia (ranging from 5% in Vietnam to 28% in Laos), averting a total of 27,500 (95UI: 24,700-30,400) premature deaths. In south-eastern China, the disease burden is reduced by 3%, averting 31,400 (95UI: 30,500-32,400) premature deaths. Assuming a low fire scenario (FINN) decreases the averted annual PM 2.5 disease burden from eliminating fire emissions by a factor of 1.3 (Table S2). Figure 7 shows the averted annual premature mortalities and mortality rate by country from eliminating fire emissions. While the number of avoided total premature mortalities is much higher in south-eastern China, due to the high population, the averted mortality rate in this region is smaller than the other countries, due REDDINGTON ET AL.  Eliminating fire emissions reduces the annual disease burden due to long-term exposure to ambient O 3 by 10% in Mainland Southeast Asia (ranging from 4% in Vietnam to 20% in Myanmar), averting a total of 2,250 (95UI: 2,000-2,470) premature deaths (Table 2). In south-eastern China, the annual disease burden is reduced by 1%, averting 1,530 (95UI: 1,380-1,660) premature deaths. In the FINNx1.5 scenario, the reduction in surface O 3 by country is slightly smaller than for the FINN scenario due to nonlinear effects driving O 3 concentrations, resulting in smaller averted disease burdens (Table S2).

Poverty and Smoke Exposure
In this section, we examine the poverty levels of the Southeast Asian population exposed to fire-derived PM 2.5 pollution. Figure 8 shows WRF-Chem simulated annual mean fire-derived (smoke) PM 2.5 and nonfire PM 2.5 concentrations plotted against gridded poverty proxy (IMR) data for the Southeast Asian domain. Populations in regions with relatively high IMRs (>60 deaths per 1,000 births) are generally exposed to higher annual mean PM 2.5 concentrations from fire than populations with relatively low IMRs (<40 deaths per 1,000 births). In areas with IMR ≥ 60, the mean fire-derived PM 2.5 exposure (10.6 μg m −3 ) is significantly greater (at the 99% confidence level) than the mean fire-derived PM 2.5 exposure in areas with IMR ≤ 20 (3.5 μg m −3 ). At the national scale, countries with higher IMRs (Laos, Cambodia, and Myanmar; Figure S8) also experience greater particulate emissions from fires ( Figure 1b) and greater exposure to fire-derived PM 2.5 (Figure 6a) than other countries in Southeast Asia. Also, this result may reflect that rural populations in Southeast Asia, which are generally located closer to forest and vegetation fires, often experience greater IMRs (e.g., Myanmar Ministry of Health, 2003).
When we consider PM 2.5 from all sources other than fires (Figure 8b), we obtain the opposite result, where populations in regions with relatively high IMRs (>60 deaths per 1,000 births) are generally exposed to lower annual mean nonfire PM 2.5 concentrations than populations with REDDINGTON ET AL.
Considering PM 2.5 from all sources ( Figure S9), we find that on average, "not poor" and "moderately poor" populations (with IMR < 32) are exposed to annual mean PM 2.5 concentrations derived predominantly (88%) from nonfire sources. However, for "very poor" populations (with 65 ≤ IMR < 100), fire-derived PM 2.5 makes up a more substantial fraction (41%) of the total PM 2.5 exposure, with 59% from nonfire sources. Figure 9 shows the spatial distribution of relative poverty levels (IMR) and fire-derived PM 2.5 exposure (WRF-Chem-simulated annual mean fire-derived PM 2.5 concentrations) across Southeast Asia. This figure indicates a large region in Southeast Asia (including northern Laos, northwest Vietnam, northern Cambodia, northern and eastern Myanmar, and Yunnan province in China) where populations with medium or high levels of poverty are exposed to medium or high levels of PM 2.5 pollution from fires. In particular, two areas in northern Laos and western Myanmar show relatively high levels of both poverty and PM 2.5 exposure, suggesting populations in these regions may be particularly at risk to health impacts from fires.
Overall, these results suggest that populations with greater levels of poverty are disproportionally exposed to PM 2.5 from vegetation and forest fires in Southeast Asia. For very poor populations, fire-derived PM 2.5 concentrations contribute over a third to the total PM 2.5 exposure.

Discussion of Public Health Impacts and Policy
To put our estimated public health impacts into context, we compare disease burdens due to fire-derived PM 2.5 exposure calculated for other fire-intensive regions. Previous studies have estimated that preventing forest and vegetation fires would avert ∼5,000-16,800 annual premature deaths across South America (Butt et al., 2020;F. H. Johnston et al., 2012;Nawaz & Henze, 2020;Reddington et al., 2015) and ∼6,000-100,300 annual premature deaths across Equatorial Asia (Crippa et al., 2016;Kiely et al., 2020;Koplitz et al., 2016;Marlier et al., 2012). The wide range in estimates reflects differences in the experimental design/methods, for example, time periods (with strong interannual variability in fire emissions in these regions), atmospheric models, and, in particular, exposure-outcome associations (as discussed by Butt et al., 2020;Conibear et al., 2018;Giani et al., 2020;Kiely et al., 2020;Reddington, Conibear, et al., 2019).
Using similar WRF-Chem setups and exposure-outcome association (the GEMM) as used in this study, previous studies found that eliminating fire would avert 16,800 (95UI: 16,300-17,400) premature deaths across South America in 2012 (Butt et al., 2020) and 44,000 (34,700-53,900) premature deaths across Equatorial Asia in 2015 (Kiely et al., 2020). The total averted disease burden for our Southeast Asian domain, 59,000 (95UI: 55,200-62,900) premature deaths, is greater than estimated for the other two fire-influenced regions, despite there being a major drought-induced haze event across Equatorial Asia in 2015. Removing the population size dependence, the per capita averted disease burden estimates for countries in Southeast Asia (10-26 (95UI: 9-29) deaths per 100,000 people) are comparable to those estimated for Bolivia, Brazil and Peru (11-22 (95UI: 10-26) deaths per 100,000 people) in 2012 (Butt et al., 2020) and for Singapore, Brunei and Malaysia (20-33 (95UI: 16-41) deaths per 100,000 people) in 2015 (Kiely et al., 2020). These comparisons indicate that populations in Mainland Southeast Asia, suffer from substantial exposure to smoke from fires with adverse impacts on public health that are comparable to other major fire regions in the tropics.
We also compared the averted disease burden from eliminating fire to those that would be achieved by eliminating other emissions sectors, estimated in Reddington, Conibear, et al. (2019). Using the same health REDDINGTON ET AL.

10.1029/2021GH000418
13 of 20 impact calculation method as Reddington, Conibear, et al. (2019) (the Integrated Exposure-Response function [GBD 2015Risk Factors Collaborators, 2016), the avoided PM 2.5 disease burdens in Mainland Southeast Asia due to eliminating fire emissions (12,200 [95UI: 6,,000] premature deaths) are lower than calculated with the GEMM (Table 2). These values are comparable to eliminating all industrial emissions; a factor 6 greater than eliminating electricity generation emissions; and a factor 10 greater than eliminating land transport across Mainland Southeast Asia. We note that we do not account for toxicity variation within PM 2.5 exposure as it is currently unknown; with disagreement in the literature regarding the toxicity of fire-derived PM relative to ambient PM (Aguilera et al., 2021;H. J. Johnston et al., 2019;Pongpiachan, 2016;Wegesser et al., 2009). The health effects of different sources and components of PM exposure is an ongoing area of research (Adetona et al., 2016;J. C. Liu et al., 2015;Naeher et al., 2007;Reid et al., 2016).
There is considerable uncertainty associated with deriving fire emissions from satellite retrievals (e.g., Pan et al., 2020;Reddington et al., 2016), and previous studies have reported that these emissions, particularly from agricultural fires, may be underestimated in Mainland Southeast Asia (Lasko et al., 2017;Reddington et al., 2016;Sornpoon et al., 2014) and China (Shen et al., 2019;Stavrakou et al., 2016;L. Zhang et al., 2016;T. Zhang et al., 2020). The underestimation of emissions from these fires is likely due to multiple factors, but particularly their small size (difficult for burned area products to detect) and short duration of active burning (a high potential to be missed by polar-orbiting satellites with detection frequencies of only a few times per day) (e.g., T. Zhang, Wooster, et al., 2018). Applying a simple scaling factor to the fire emissions will partly compensate for emissions underestimation, but emissions estimates are still likely to be conservative in regions with a high number of missed detections.
Our analysis shows that a reduction of fire across southeast Asia would have substantial health benefits. Successful fire management requires information about the main types and causes of fire. Across Mainland Southeast Asia, emissions are dominated by forest fires (deforestation, savanna, and temperate forest classes in GFED) which account for 96% of particulate emissions across our domain, with greater contributions in Cambodia, Laos, and Myanmar. A detailed analysis of fires confirms that most fires in the region occur in forest land covers (Vadrevu et al., 2019). A close association between fire and deforestation has also been shown in other tropical regions including the Brazilian Amazon (Reddington et al., 2015) and Indonesia (Adrianto et al., 2019(Adrianto et al., , 2020. In Southeast Asia, fires are lit in forests to clear the land for agriculture (slash and burn, deforestation fires), to induce growth of grass for grazing, and for collection of forest products (Vadrevu et al., 2019). The large contribution of forest fires to particulate emissions suggests that reducing deforestation and associated fires should be a public health priority for the region. In Cambodia, deforestation has been linked to increased incidence of acute respiratory infection in children, likely due to increased exposure to smoke from deforestation fires (Pienkowski et al., 2017). Future work exploring the relative contributions of different fire types to air pollution in Mainland Southeast Asia would be useful to inform policy options to improve air quality.
Several policies have already been implemented to reduce agricultural fires in Southeast Asia, for example, an Alternative Energy Development Plan and a zero-burning policy for sugarcane in Thailand (Kumar et al., 2020). However, challenges remain with regards to the enforcement of these policies and their practicality, particularly for farmers that rely on manual harvesting practices (Adeleke et al., 2017;Kumar et al., 2020). Recent research shows the most effective solutions for reducing agricultural residue burning and its associated air pollution, are to encourage residue use for other purposes, for example, bioenergy, livestock feed/bedding, composting, green harvesting and so on (Kumar et al., 2020) and to apply coherent policies across multiple provinces and countries in Southeast Asia (Moran et al., 2019).
Discussion of the implementation and benefits of policies addressing deforestation and/or savanna-type fires in Southeast Asia are lacking in the literature. However, a number of policies and projects have been developed and implemented to address forest loss and conversion, many of which are related to UNFCCC REDD+ (reducing emissions from deforestation and forest degradation and the role of conservation, sustainable management of forest and enhancement of forest carbon stocks) (e.g., Kissinger, 2020). Key drivers of deforestation are expansion of cropland and commercial agriculture (Lim et al., 2017;Y. Zhang, Prescott, et al., 2018), for example, conversion of forest to coffee and/or rubber plantations (Fox & Castella, 2013;Kissinger, 2020). There is evidence that protected areas and community-protected forests can play an important role in protecting forests from large-scale burning and deforestation fires (Biswas et al., 2015;Singh et al., 2018).

Conclusions
In this study, we explored the impact of forest and vegetation fires on air quality and public health across Southeast Asia. We used a combination of two air quality models: a global aerosol model, GLOMAP, to test three different satellite-derived fire emission data sets (FINN, GFED, GFAS); and a high-resolution, regional air quality model, WRF-Chem, to quantify the air quality and public health benefits of eliminating fire emissions. Simulating the elimination of all fires across the region, rather than fires specifically identified to be human-caused, illustrates the maximum possible public health benefit achievable (within uncertainties) and provides an upper bound for policy makers.
We found that GLOMAP was better able to reproduce measurements of fire-derived PM in Thailand across multiple years with the FINN data set compared to the GFAS or GFED data sets. This result is consistent with findings in our previous work (Reddington et al., 2016). PM emissions across Southeast Asia in FINN are a factor 4 greater than GFED or GFAS. WRF-Chem using FINN best simulated measured PM concentrations when particulate fire emissions were scaled upwards by a factor 1.5. Our analysis suggests fire emissions in this region are underestimated, particularly in the GFED and GFAS data sets.
Overall, we found that preventing fire could substantially improve regional air quality in Mainland Southeast Asia with a more limited benefit to air quality in south-eastern China. Population-weighted annual mean PM 2.5 concentrations were reduced by 16% in Mainland Southeast Asia and by 2% in south-eastern China. ADM8h O 3 concentrations were reduced by 9% in Mainland Southeast Asia and by 2% in south-eastern China. Eliminating fire emissions substantially reduced populations exposed to PM 2.5 concentrations above WHO AQ Interim Target 2 in Thailand, Myanmar, Laos, and Cambodia (by 44%-100%).
We found a considerable public health benefit of eliminating fire emissions across the region, largely due to reductions in PM 2.5 exposures. The annual disease burden due to PM 2.5 exposure was reduced by 12% in Mainland Southeast Asia, averting 27,500 (95UI: 24,700-30,400) premature deaths, and by 3% in south-eastern China, averting 31,400 (95UI: 30,500-32,400) premature deaths. The annual disease burden due to O 3 exposure was reduced by 10% in Mainland Southeast Asia, averting 2,250 (95UI: 2,000-2,470) premature deaths, and by 1% in south-eastern China, averting 1,530 (95UI: 1,380-1,660) premature deaths.
Using subnational poverty proxy data, we found that poorer populations in Southeast Asia are disproportionally exposed to PM 2.5 from vegetation and forest fires; with significantly higher average fire-derived PM 2.5 exposure for populations with relatively high infant mortality rates.
Our analysis suggests that exposure to fire-derived PM 2.5 is associated with a greater annual disease burden in Southeast Asia than in both the Amazon region in 2012 and Equatorial Asia in 2015, with similar per capita averted disease burdens to those estimated for heavily fire-impacted countries in South America. Furthermore, preventing fires across Mainland Southeast Asia would yield a public health benefit comparable to that achieved by eliminating all industrial emissions across the region, and considerably larger than achieved by eliminating emissions from either the electricity generation or land transport sectors.
In summary, forest and vegetation fires are important to consider in addition to more traditional emission sectors (e.g., industry, transport, and residential solid-fuel combustion) when assessing causes of air quality degradation in Southeast Asia and for developing emission control policies to improve air quality across this region. These policies should focus on reducing deforestation and savanna-type fires in addition to agricultural fires in order to effectively address the regional air quality issues. Previous work in Equatorial Asia (Reddington et al., 2014) demonstrates the need to understand the effectiveness of regional emission control strategies and how they will reduce population exposure. Future work is required to identify the regions where emission controls would most effectively reduce exposure, especially for the poorest populations.

Data Availability Statement
The air pollution and health impact assessment data per country/region that support the findings of this study are available at the Research Data Leeds Repository (https://doi.org/10.5518/968). Code to setup and run WRFChem (using WRFotron version 2.0) is available through Conibear and Knote (2020). search Council (NE/S006680/1). This work was undertaken on Advanced Research Computing, part of the High Performance Computing facilities at the University of Leeds, UK. This work used WRFotron version 2.0, a tool to automatize WRFChem runs with reinitialized meteorology (Conibear & Knote, 2020). The authors acknowledge the use of WRFChem preprocessor tools mozbc, fire_emiss, anthro_emiss, bio_emiss provided by the Atmospheric Chemistry Observations and Modeling Laboratory of the National Center for Atmospheric Research. The authors acknowledge the use of Model for Ozone and Related Chemical Tracers version 4 (MOZART-4) global model output available at https://www.acom. ucar.edu/wrf-chem/mozart.shtml. The boundaries shown on any maps in this work do not imply any judgment concerning the legal status of any territory or the endorsement or acceptance of such boundaries.