Large air quality and human health impacts due to Amazon forest and vegetation fires

We used the Weather Research and Forecasting online-coupled Chemistry model (WRF–Chem) version 3.7.1 (Grell et al 2005). The model domain covers most of South America (figure 1) with a horizontal resolution of 30 km, extending vertically from the surface to 10 hPa. Details of model setup are shown in supplementary table 1 is available online at stacks.iop.org/ERC/2/095001/mmedia.

Zoom In Zoom Out Reset image size

Gas-phase chemistry is calculated using the extended Model for Ozone and Related Chemical Tracers, version 4 (MOZART-4) (Emmons et al 2010, Knote et al 2014). Aerosol chemistry and microphysics is simulated using an updated Model for Simulating Aerosol Interaction and Chemistry (MOSAIC) with aqueous chemistry and four sectional discrete aerosol size bins: 0.039–0.156 μm, 0.156–0.625 μm, 0.625–2.5 μm, 2.5–10 μm (Zaveri et al 2008, Hodzic and Knote 2014). An updated volatility basis set mechanism was also included for secondary organic aerosol (SOA) formation (Knote et al 2015).

Microphysics is simulated using the Morrison 2-moment scheme (Morrison et al 2009) and the Grell 3D parameterisation is used for simulating convection (Grell and Dévényi 2002). Initial and boundary chemistry and aerosol conditions were taken from 6-hourly simulation data from the MOZART-4/Goddard Earth Observing System Model version 5 (GEOS5) (NCAR 2019). Initial and boundary meteorological conditions were taken from the European Centre for Medium–Range Weather Forecasts (ECMWF) global reanalysis (Dee et al 2011)

During model simulations, we nudged the meteorological components, horizontal and vertical wind, potential temperature and water vapour mixing ratio, to ECMWF re-analysis in all model levels above the planetary boundary layer over 6 h.

WRF-Chem simulations were conducted for the year 2012 at horizontal resolution of 30 km. We focused on the dry season (defined here as August 1st to October 31st) when vegetation fires are most active. We performed simulations from April to December, discarding the first month as model spin-up. We performed two types of simulations: one simulation excluding fire emissions ('fire_off') and one simulation including vegetation fire emissions ('fire_on'). The contribution of fires to PM_2.5 was calculated as the difference in concentrations between these two simulations.

2.2.1. Non-fire emissions

2.2.2. Fire emissions

We used fire emissions from the Fire Inventory from NCAR (FINN) version 1.5 (Wiedinmyer et al 2011) (see list of emissions species in supplementary table 2). Daily emissions are estimated on a 1 km² grid based on the location and timing of active fires taken from the Moderate Resolution Imaging Spectroradiometer (MODIS) Fire and Thermal Anomalies Product (Giglio et al 2003). Each fire count is assigned a burned area (0.75 km² for grassland and savannah and 1 km² for other land covers). To account for missing fire retrievals due to cloud cover, FINN averages fire emissions over two days, assuming that the detected burned area will be half the original size the following day (Wiedinmyer et al 2011, Pereira et al 2016). Trace gas and aerosol emissions are calculated using emission factors (Akagi et al 2011, 2013, Yokelson et al 2013) in conjunction with MODIS Land Cover Type and Vegetation Continuous Fields. Fire emissions are emitted with a diurnal cycle that peaks in the early afternoon (local-time) based on Giglio (2007).

Buoyancy due to fire plumes can cause rapid injection of fire emissions above the ground surface (e.g., Fromm et al 2000). WRF-Chem can simulate this plume-rise from fires (e.g., Freitas et al 2007). However, previous modelling studies of Amazon fires during the dry season of 2012 shows this method overestimates the height of fire plumes (Archer-Nicholls et al 2015, Archer-Nicholls et al 2016). Observational work also found very little evidence of extensive elevated layers during the same simulation period, suggesting little evidence of pyroconvection (Darbyshire et al 2019). We therefore inject vegetation fire emissions evenly throughout the boundary layer (BL) (e.g., Dentener et al 2006), as supported by analysis of plume heights over the Amazon (Marenco et al 2016, Gonzalez-Alonso et al 2019).

Table 1 shows annual and dry season organic carbon (OC) and black carbon (BC) emissions from FINN v1.5 summed over our modelling domain. Emissions are dominated by OC, which contributes 90% of OC + BC emissions. In 2012, domain wide annual OC + BC emissions were 4 Tg a⁻¹, approximately 17% greater than the 11-y (2008 to 2018) average (3.4 Tg a⁻¹). The domain contributes 17% of global fire BC + OC emissions during 2008 to 2018 (figure 2(d)), highlighting the importance of fires in this region. Emissions are dominated by dry season fires, which account for 65% of annual OC emissions over the 11-y average. Fire emissions are greatest in southern areas of the Amazon (figures 2(a), (b)) which are undergoing land-use change. During 2008 to 2018, the Amazon accounted for 66% (72%) of annual (dry season) domain total OC emissions (table 1). Compared to the 11-y average, fire emissions in 2012 were enhanced across the Brazilian and Peruvian Amazon, but were lower across the Bolivian Amazon (figure 2(c)). Emissions were also higher in the eastern Cerrado region of Brazil in 2012 compared to the 11-y average, as supported by observed enhancement in Cerrado active fires in this region (supplementary figure 1). Previous work comparing FINN (v1.5) to the Global Fire Assimilation System (GFAS) (Kaiser et al 2012) and Global Fire Emissions Dataset (GFED) (v4s) (Van Der Werf et al 2017) in 2012, showed that emissions datasets had broadly similar spatial patterns of BC and OC emissions over the Amazon region. Emissions in FINN were greater compared to GFED and GFAS over western regions and lower compared to other datasets over eastern regions (Reddington et al 2019b). Comparison against measured aerosol concentrations suggested that GFED and GFAS emissions may be underestimated in the western Amazon and all emission datasets may underestimate in the eastern Amazon (Reddington et al 2019b).

Table 1. Domain wide and Amazon Basin annual and dry season (August–October) total organic carbon (OC) and black carbon (BC) emissions from FINN (v1.5). Emissions in 2012 are compared against the 11-y (2008 to 2018) average in parentheses.

Emission species	Annual domain (Tg a−1)	Dry season domain (Tg a−1)	Amazon Basin annual	Amazon Basin dry season
OC	3.6 (3.1)	2.6 (2.0)	2.4 (2.1)	1.8 (1.5)
BC	0.4 (0.3)	0.3 (0.2)	0.3 (0.2)	0.2 (0.2)

Zoom In Zoom Out Reset image size

We evaluate our WRF-Chem simulations against a comprehensive set of observational datasets from surface and aircraft collected as part of the South American Biomass Burning Analysis (SAMBBA) field campaign, which took place over the southern Amazon in September and October 2012 (Johnson et al 2016, Reddington et al 2019b). We compliment SAMBBA observations with additional surface and satellite measurements.

2.3.1. Statistical methods

In order to compare WRF-Chem to measurements, we linearly interpolated model output to the time and location of measured data. Comparison to aerosol mass measurements was conducted using simulated mass within the instrument detection ranges. Model evaluation was quantified using Pearson correlation coefficient (r), mean bias (MB) and normalised mean bias factor (NMBF) (Yu et al 2006):

$$\begin{eqnarray}&&\begin{array}{l}MB=\displaystyle \frac{1}{N}\displaystyle \sum _{i=1}^{N}\left(Mi-Oi\right)\,\\ NMBF=\displaystyle \frac{\displaystyle \sum \left(Mi-Oi\right)}{\displaystyle \sum Oi}\,if\,\overline{M\,}\geqslant \,\overline{O}\,or\\ \,NMBF=\displaystyle \frac{\displaystyle \sum \left(Mi-Oi\right)}{\displaystyle \sum Mi}\,if\,\overline{M\,}\lt \overline{O\,}\end{array}\end{eqnarray} \tag{ 1 }$$

where M and O are the model and observation value at location and timestep i. MB shows the deviation of the model to observation in the same units. NMBF is unitless and is interpreted as a factor NMBF + 1 by which the model under or overestimates the observation.

2.3.2. Surface measurements

2.3.3. Aircraft measurements

2.3.4. Aerosol optical depth measurements

2.3.5. Radiosonde measurements

We used simulated annual mean surface PM_2.5 concentrations to quantify the health impact due to fires through the disease burden attributable to air pollution exposure. To estimate annual mean PM_2.5 we assumed simulated concentrations in May and December are representative of January–April, when fire emissions are also low.

Using population attributable fractions of relative risk taken from associational epidemiology, intervention–driven variations in exposure (i.e., population exposure including and excluding vegetation fires) were used to predict associated variations in health burden outcomes. The population attributable fraction (PAF) was estimated as a function of population (P) and the relative risk (RR) of exposure:

$$\begin{eqnarray}&&PAF=P\times (R{R}_{EXP}-1/R{R}_{EXP})\end{eqnarray} \tag{ 2 }$$

The RR was estimated through the Global Exposure Mortality Model (GEMM) (Burnett et al 2018). We used the GEMM for non–accidental mortality (non–communicable disease, NCD, plus lower respiratory infections, LRI), using parameters including the China cohort, with age–specific modifiers for adults over 25 years of age in 5–year intervals. The GEMM functions have mean, lower, and upper uncertainty intervals. The theoretical minimum-risk exposure level for the GEMM functions is 2.4 μg m^–3. The toxicity of PM_2.5 was treated as homogenous with no differences for source, shape, or chemical composition, due to a lack of associations among epidemiological studies.

The effect of air pollution is known to be significantly different for morbidity and mortality from cardiovascular outcomes (IHD and STR) (Cohen et al 2017), and the relative risks from equation (1) were adjusted by equation (2) (RR_adjusted) when estimating years lived with disability (YLD). The ratio for IHD and STR morbidity were 0.141 and 0.553 from the GBD2016 (Cohen et al 2017), based on data from three cohort studies (Miller et al 2007, Puett et al 2009, Lipsett et al 2011)

$$\begin{eqnarray}&&R{R}_{EXP,adjusted}=ratio\times R{R}_{EXP}-ratio+1\end{eqnarray} \tag{ 3 }$$

Premature mortality (MORT), years of life lost (YLL), and years lived with disability (YLD) per health outcome, age bracket, and grid cell were estimated as a function of the PAF and corresponding baseline mortality (I_MORT), YLL (I_YLL), and YLD (I_YLD) following equations (3)–(5), respectively.

$$\begin{eqnarray}&&MORT=PAF\times {I}_{MORT}\end{eqnarray} \tag{ 4 }$$

$$\begin{eqnarray}&&YLL=PAF\times {I}_{YLL}\end{eqnarray} \tag{ 5 }$$

$$\begin{eqnarray}&&YLD=PAF\times {I}_{YLD}\end{eqnarray} \tag{ 6 }$$

Disability–adjusted life years (DALYs), i.e. the total loss of healthy life, was estimated as the total of YLL and YLD:

$$\begin{eqnarray}&&DALYs=YLL+YLD\end{eqnarray} \tag{ 7 }$$

The rates of MORT, YLL, YLD, and DALYs were calculated per 100,000 population. Mean estimates were quantified in addition to upper and lower uncertainty intervals at the 95% confidence level. The United Nations adjusted population count dataset for 2015 at 0.25° × 0.25° resolution was obtained from the Gridded Population of the World, Version 4 (GPWv4) (Doxsey-Whitfield et al 2015). Population age composition was taken from the GBD2017 for 2015 for early–neonatal (0–6 days), late–neonatal (7–27 days), post–neonatal (8–364 days), 1–4 years, 5 to 95 years in 5–year intervals, and 95 years plus (Global Burden of Disease Study 2017, 2018). Shapefiles were used to aggregate results at the country and state level (Hijmans et al 2012).

The health impacts of PM_2.5 depend non-linearly on exposure, with impacts starting to saturate at high PM_2.5 concentrations. We estimate the health benefits that would arise if fires were prevented, as the health burden from a scenario with fires (fire_on) minus the health burden from a scenario without fires (fire_off) (but including other emission sources). This is described as the 'subtraction' method (Kodros et al 2016, Conibear et al 2018).

Figure 3 shows measured and simulated surface aerosol mass concentrations at Porto Velho, a location heavily influenced by vegetation fires. Before the dry season (May to July), measured PM_2.5 concentrations are typically less than 3 μg m⁻³, peaking at 30–50 μg m⁻³ in August and September, followed by a decline in early October to less than 10 μg m⁻³ (figure 3(a)). The model captures this seasonal cycle relatively well (r = 0.57), but overestimates concentrations (NMBF = 0.72, MB = 7.6 μg m⁻³) largely due to an over prediction from mid-August to September. Simulated PM_2.5 concentrations are underestimated in early August, possibly due to the missing fires at the start of the dry season in the FINN dataset (Reddington et al 2019b). Vegetation fires contributed ∼86% to simulated PM_2.5 concentrations in August and September. In the simulation without fires, PM_2.5 concentrations remain below 3 μg m⁻³ throughout May to December. At urban locations far from the fires, the model simulates annual mean surface PM_2.5 concentrations to within 25% (supplementary figure 3; NMBF = −0.26, MB = −3.47 μg m⁻³).

Zoom In Zoom Out Reset image size

During the SAMBBA campaign, measured concentrations of OA, BC and total aerosol mass show a decline in mass concentrations towards the end of the dry season, from 13th to 30th September 2012 (figure 3(b)). Overall, the model also overestimates OA (NMBF = 0.87), BC (NMBF = 0.2) and total aerosol mass (NMBF = 0.87) in September (figure 3(b)). The model performs better during SAMBBA Phase 1 (OA NMBF = 0.36; BC NMBF = 0.03; total aerosol mass NMBF = 0.37), compared to Phase 2 (OA NMBF = 3.51; BC NMBF = 0.60; total aerosol mass NMBF = 3.05). Measured and simulated aerosol mass (not shown) are dominated by OA (measured: 78.7%, model: 83.6%), with smaller contribution from BC (measured: 11.5%, model: 5.8%), and inorganics (NH₄ + NO₃ + Cl + SO₄; measured: 9.8%, model: 10.6%). Analysis of OA:CO ratios in biomass burning smoke during the SAMBBA campaign suggests limited net gain of OA mass from secondary processes within fire plumes (Brito et al 2014, Morgan et al 2019a). We also find limited additional mass due to SOA formation, with SOA contributing <15% to simulated OA mass at Porto Velho.

Figure 4 shows simulated and measured vertical profiles of OA and BC mass taken during the SAMBBA flight campaign. Following previous studies (e.g., Archer-Nicholls et al 2015, Reddington et al 2019b), we separate analysis into SAMBBA Phase 1 and Phase 2, and into western and eastern regions (section 2.3.3). Measured aerosol mass concentrations are greater in the lower 2.5 km in the west and the lower 4 km in the east, with concentrations rapidly declining above. Simulated concentrations capture the shape of the observed vertical profile (r ≥ 0.9), showing that distributing fire emissions through the BL is a reasonable assumption.

Zoom In Zoom Out Reset image size

In the western Amazon, simulated BL concentrations of both BC and OA are in generally good agreement with observations (NMBF ≤ 0.15). Measured OA and BC concentrations show a reduction between Phase 1(OA: 15 μg m⁻³; BC: 1 μg m⁻³) and Phase 2 (OA: 7 μg m⁻³; BC: 0.6), consistent with the decline in measured surface concentrations at Porto Velho (figure 3). The model simulates this reduction in OA (observations: −56%; model: −58%) and BC (observations: −38%; model: −30%) concentrations, coinciding with the dry-to-wet-season transition. Previous modelling studies have struggled to simulate this observed reduction in the western Amazon, potentially due to poor simulation of wet removal of aerosols (Archer-Nicholls et al 2015, Reddington et al 2019b).

For the eastern region, simulated concentrations in the BL are underestimated (OA NMBF = −0.66; BC NMBF = −6.32), consistent with a previous study (Reddington et al 2019b). Measured BC:OA ratios in the BL are greater in the eastern region (0.16) compared to the western region (around 0.1). This may be caused by higher BC emission factors of flaming savannah fires in eastern Cerrado regions relative to the smouldering fires typical of tropical forest fires in western Amazon (Hodgson et al 2018, Darbyshire et al 2019). The underestimation of BC in the eastern region (observed: 1.4 μg m⁻³; model 0.2 μg m⁻³) may partly be a result of underrepresentation of emissions from Cerrado fires in eastern regions. However, we note the few flights were undertaken in the eastern region compared to western region (section 2.3.3).

Figure 5 shows daily mean simulated and observed AOD at 500 nm (AOD500) between May and December 2012 at five AERONET stations located in western and southern Amazon Basin. Measured AOD500 is typically <0.2 in May and June, increasing to 0.4 to 1.5 in August and September, reducing to <0.2 in December. The model captures this seasonal cycle relatively well (r = 0.61 to 0.88). May to December mean AOD500 is overestimated in western locations (Rio Branco, NMBF = 0.53, Porto Velho UNIR (NMBF = 0.47) and Santa Cruz UTEPSA (NMBF = 0.47), and well simulated compared to more eastern locations: Alta Floresta (NMBF = 0.03) and Cuiabá-Miranda (NMBF = −0.01).

Zoom In Zoom Out Reset image size

Comparison against AOD at 550 nm retrieved by MODIS (AOD550), confirms the model overestimates in the western Amazon and underestimates eastern regions (supplementary figure 4). AERONET AOD500 and MODIS AOD550 are found to compare well (supplementary figure 5) despite the different wavelengths. Evaluating against MODIS separately over evergreen broadleaf forest and savannah (cerrado) biomes and across the western and eastern regions shows an overall very small low bias in western forest bias regions with an equally small high bias in eastern cerrado regions (supplementary figure 6). The underestimation over savannah regions is considerably less than the underestimate against the one aircraft flight in the east, suggesting comparison against this one flight may not be representative. However, Reddington et al (2019b) also found PM and AOD were underestimated over regions with savannah and grassland fires, possibly suggesting FINN underestimates fire emissions in these regions.

Some previous modelling studies of Amazon (Archer-Nicholls et al 2015, Johnson et al 2016) and other tropical fires (see Reddington et al (2019b) for a review) underestimate AOD and scale-up fire emissions to enable the model to match observed AOD. We find a consistent evaluation of regional boundary layer PM concentrations and AOD, with a slight overestimation over forested regions in the western Amazon and underestimation over savanna regions. Overall, simulated PM_2.5 was typically within 25% of measurements both close to fires in the western Amazon and in urban regions far from fires. We therefore chose not to alter fire emissions and we use PM_2.5 concentrations from the model runs with and without fire emissions to estimate impacts on human health.

Figure 6 shows simulated surface dry season PM_2.5 concentrations. Greatest dry season concentrations (≥45 μg m⁻³) occur in the southern and western Amazon. Vegetation fires contribute up to 80%–95% of simulated dry season PM_2.5 concentrations, with contributions >60% over most of the Brazilian Amazon, Bolivia, and much of Peru and Paraguay.

Zoom In Zoom Out Reset image size

Figure 7 shows the regional distribution of dry season mean simulated PM_2.5 concentrations. Vegetation fires increased regional mean PM_2.5 concentrations by 260% over the Amazon Basin in 2012, exposing 20 million people to dry season mean concentrations above 10 μg m⁻³ and 3.3 million people to concentrations over 25 μg m⁻³. Similarly large increases are also simulated at the national scale for Peru (394%) and Bolivia (509%) where 7% (1.36 million people) and 4% (0.42 million people) are exposed to PM_2.5 levels above 25 μg m⁻³, respectively. Fires have a marked impact on annual concentrations and thus chronic exposure (see also supplementary figure 7), increasing annual mean population-weighed PM_2.5 concentrations by 35% and 137% in Peru and Bolivia, respectively (table 2). By comparison, vegetation fires increased the national regional dry season mean PM_2.5 by a smaller amount in Brazil (148%) in 2012, exposing 14.4 million people to levels above 25 μg m⁻³ in the dry season, increasing annual mean population-weighed PM_2.5 concentrations by 10%.

Zoom In Zoom Out Reset image size

Due to the proximity of fires, western states of Brazil are impacted by fires disproportionately. Figure 8 shows regional distribution of dry season mean simulated PM_2.5 concentrations in four western states of Brazil (Acre, Amazonas, Mato Grosso, and Rondônia) and the state of São Paulo in south-eastern Brazil. Vegetation fires increase regional PM_2.5 concentrations considerably in western states (296%–791%), exposing the majority of state populations to dry season mean PM_2.5 concentrations above 25 μg m⁻³ in Rondônia (53%) and Acre (75%). Fires similarly increase annual mean population-weighted PM_2.5 concentrations considerably in these western states (59%–278%), highlighting the impact on chronic exposures. In contrast, exposure to unhealthy PM_2.5 concentrations in outflow regions such as São Paulo is largely due to other anthropogenic emissions with fires playing a limited intermittent contribution. Nevertheless fires can contribute to very high concentrations in outflow regions, leading to a 39% increase in dry season regional mean concentrations and a 5% increase in annual mean population-weighted concentrations in São Paulo.

Zoom In Zoom Out Reset image size

Figure 9 shows the estimated reduction in the regional burden of disease that would occur if all vegetation fires were prevented. We estimate that the prevention of vegetation fire emissions had the potential to avoid 641 00 (95UI: 551 900–741 300) DALYs and 16 800 (95UI: 16 300–17 400) premature deaths across our South American domain in 2012. We found that approximately 26% of the avoided DALYs (167 900 (95UI: 143 800–194 900)) and deaths (4 300 (95UI: 4 100–4 500 )) due to fire prevention were located inside the Amazon Basin. At the national level, preventing vegetation fires could have prevented 9 770 (95UI: 9 690–9 870) premature mortalities in Brazil, 1 467 (95UI: 1 340–1 590) in Peru and 1 195 (95UI: 985–1 410) in Bolivia (table 3). Per capita health burdens remove population size dependence highlighting the impact of fires on public health. The per capita avoided heath burden is greatest in Bolivia (789 (95UI: 605–979) DALYs per 100 000 people) and Paraguay (644 (95UI: 531–772) DALYs per 100 000 people) followed by Brazil (597 (95UI: 524–680) DALYs per 100 000 people). Brazilian states of Rondônia, Acre, and Mato Grosso benefit the most from fire prevention with 1300–1800 DALYs per 100 000 people avoided (figure 9 and supplementary table 4). High disease burden rates in these western Brazilian states and wider Amazon Basin, highlights the adverse impact of vegetation fires on regional public health.

Zoom In Zoom Out Reset image size

Our estimated health impacts are consistent with previous estimates for South America from global models but provide advancement due to the use of a high resolution regional model and updated exposure-response relationships. Johnston et al (2012) estimated preventing fires would avoid 10 000 premature deaths annually between 1997–2006. Reddington et al (2015) estimated prevention of vegetation fires would avert ∼7000 to 9700 premature deaths annually between 2002–2011.

In contrast to the Amazon Basin and the wider South American region, numerous air quality health assessments due to vegetation fires have been conducted across Equatorial Asia. Marlier et al (2019) found the prevention of fires across Equatorial Asia (Indonesia, Malaysia, and Singapore) would avert 24 000 premature deaths annually in the present-day. Using a similar WRF-Chem setup as used in this study, Reddington et al (2019a) estimated the prevention of fires would avoid 8,000 premature deaths annually across Southeast Asia (Myanmar, Thailand, Laos, Cambodia, and Vietnam) in the present-day. Using a similar WRF-Chem setup as used in this study, Kiely et al (2020) found that the prevention of vegetation and peat fires would avert 15 000 premature deaths and 500 000 DALYs annually in 2004, 2006, and 2009 across Equatorial Asia. A number of studies focused on 2015, where drought induced fires resulted in a major haze event across Equatorial Asia, and caused an estimated 44 000–100 300 premature mortalities (Crippa et al 2016, Koplitz et al 2016, Kiely et al 2020).

Calculated health burdens are sensitive to both simulated PM exposure and exposure-response relationships (Ostro et al 2018, Burnett and Cohen 2020). For the 2015 haze event in Equatorial Asia, Kiely et al (2020) found that exposure-outcome associations have the largest influence on health impact estimates. Kiely et al (2020) used the same exposure-outcome association as this study (GEMM NCD + LRI) and estimated 44,000 premature deaths in 2015. In contrast, Koplitz et al (2016) used an exposure-outcome association based on work from Schwartz et al (2008), Anenberg et al (2012), Lepeule et al (2012), with a 1% in all-cause mortality per 1 μg m⁻³ increase in annual average PM2.5 concentrations. Koplitz et al (2016) estimated 100,300 premature deaths in 2015. When Kiely et al (2020) used the same exposure-outcome association they estimated a similar premature death estimate of 106,000 deaths in 2015. Marlier et al (2019) also used a similar exposure-outcome association to that used in Koplitz et al (2016), based on work from Vodonos et al (2018).

Both the Vodonos et al (2018) exposure-outcome association and the GEMM NCD + LRI rely on epidemiological studies from ambient PM_2.5 exposure only, including some from high-exposure locations, and include all non‐accidental causes of death. The current integrated exposure-response (IER) association from Global Burden of Disease GBD2017 is based on studies of ambient and household air pollution, passive smoking, and active smoking exposures and is cause-specific for six causes of death (Burnett et al 2014). The risk responses of Vodonos et al (2018) and the GEMM NCD + LRI are similar for exposures up to 50 μg m⁻³. At exposures greater than 50 μg m⁻³ these functions diverge, with the GEMM NCD + LRI risk flattening off at higher concentrations. The IER has lower risks than either Vodonos et al (2018) and the GEMM NCD + LRI, and flattens off at lower concentrations. For example, Nawaz and Henze (2020) estimated 4 407 premature deaths per year on average between 2016 and 2019 in Brazil from biomass burning derived ambient PM_2.5, which is approximately half of our estimate for Brazil in 2012, primarily due to their use of the IER from GBD2016 which has approximately half the attributable risks of the GEMM NCD + LRI that we used in this study.

These large differences emphasise the need to reduce these uncertainties and for further epidemiological studies from highly polluted regions of the world (Burnett and Cohen 2020, Pope et al 2020). We used the GEMM NCD + LRI here to be consistent with the latest exposure-outcome associations for ambient PM_2.5 exposure only, to include causes beyond that considered by the current IER in the GBD2017, and to be conservative of risk estimates at higher exposures.

Fire emissions in the tropics depend on land-use change and climate conditions and can exhibit strong interannual variability. Reddington et al (2015) found greater health impacts in years with greater fires due to drought or deforestation. Fire emissions in the Amazon in 2012 were comparable to the 11-y average (figure 1). Years with more fires, due either to drought conditions or greater deforestation and land-use change, would have greater emissions, PM concentrations and likely greater associated public health impacts. Future work is needed to understand the year to year variability in health impacts due to PM from fires in the region.