The GEOS-Chem atmospheric chemistry model (http://www.geos-chem.org/, last access: 19 August 2026), first described by Bey et al. (2001), is an open-source global 3-D atmospheric chemical transport model driven by assimilated meteorological observations from the Goddard Earth Observing System (GEOS) of the NASA Global Modeling and Assimilation Office (GMAO) (https://gmao.gsfc.nasa.gov/geos-systems/, last access: 19 August 2026). GEOS-Chem includes detailed gas-phase mechanisms for NOx–VOC–O3 chemistry and aerosol chemistry for sulfate–nitrate–ammonium aerosols, black carbon (BC), and primary and secondary organic aerosols.

In this study, we used GEOS-Chem version 13.2.1 in its high-performance implementation (GCHP). The flexibility and scalability of high resolution simulations in the standard offline version of GEOS-Chem (“GEOS-Chem Classic”) are limited, as it relies on shared-memory parallelization and a rectilinear longitude–latitude grid (Martin et al., 2022). GCHP version 11, developed by Eastham et al. (2018), shares the same source code for physical and chemical mechanisms with GEOS-Chem Classic. It operates on a cubed-sphere grid with Message Passing Interface (MPI) distributed memory framework for massive parallelization by coupling the Model Analysis and Prediction Layer (MAPL) (Suarez et al., 2007) of the NASA GMAO with Earth System Modeling Framework (ESMF) (Hill et al., 2004). Together with the development of Finite­-Volume Cubed-Sphere Dynamical Core (FV3) from Geophysical Fluid Dynamics Laboratory (GFDL) (Harris et al., 2016), GCHP version 13 could operate on a stretched cubed-sphere grid to enhance grid resolution in a cutomised region with smooth, gradual changes in resolution (Bindle et al., 2021).

Fire emission inventories were developed to convert satellite detection of fires into numerical emission inputs for chemical transport models (CTMs) or Earth system models with atmospheric chemistry to simulate and analyze the emission patterns and impacts (Liu et al., 2020). In this study, we employed the Global Fire Emission Dataset version 4 (GFEDv4s), the built-in fire emission inventory in GCHP. GFEDv4s utilizes the Moderate Resolution Imaging Spectroradiometer (MODIS) Collection 5.1 MCD64A1 burned area product at 500 m spatial resolution to estimate the burned areas. It also incorporates 1 km thermal anomalies data from MODIS Terra and Aqua and 500 m surface reflectance observations to account for the burned area of small fires in 0.25° (BAsf) (van der Werf et al., 2017). This treatment of small fires improves accuracy, making GFEDv4s one of the best inventories to capture the magnitudes of observed smoke PM2.5 in a regional case study in Indonesia (Liu et al., 2020). Additionally, GFEDv4s introduces a new burned fraction equation to estimate subgrid burned fractions in frequently burned landscapes, which are commonly found in the southeastern part of the Amazon. This modified approach adjusts the fuel load in grid cells that burn in previous months and avoids the underestimation of emissions in frequently burning regions, especially toward the end of the fire season (van der Werf et al., 2017).

Moreover, GFEDv4s adjusts the fuel consumption rates based on the field measurement database developed by van Leeuwen et al. (2014) and Scholes et al. (2011). The detailed fuel consumption database is available from https://www.geo.vu.nl/~gwerf/FC/ (last access: 19 August 2026). The satellite-measured and ground-observed information is then provided to a biogeochemical model based on the Carnegie-Ames-Standford Approach to generate the dry matter burned (DMB) data (van der Werf et al., 2017). By applying the new set of emission factors (in g species per kg dry matter burned), which are species-specific and fire type-specific, we are able to resolve trace gas and aerosol emissions from different fire types in Amazon for our study (van der Werf et al., 2017).

Recent research on GFED time series data found that 55 % of the smoke emission from 1997–2015 originated from deforestation fires (van Marle et al., 2017). The Amazon deforestation data used in this project were retrieved from TerraBrasilis, which relies on Landsat and MODIS satellite data to monitor and analyze land cover changes in the Amazon rainforest through The Amazon Deforestation Monitoring Project (PRODES) and The Real Time System for Detection of Deforestation (DETER) systems, developed by the Instituto Nacional de Pesquisas Espaciais (INPE) (F. G. Assis et al., 2019; Prodes, 2013). The Brazilian government has implemented the Action Plan for Prevention and Control of Deforestation in Amazonia (PPCDAm) in order to protect the Amazon rainforest from illegal deforestation and aim for sustainable development since 2005 (Aragão et al., 2018). After three phases of PPCDAm, the annual deforestation rate was reduced to one-third of the average deforestation rate between 1988 and 2004 during 2010–2015 (Fig. 1a), which was stable in magnitudes (∼7000km2) and favorable to the study on drought-induced air quality impacts.

The standardised precipitation-evapotranspiration index (SPEI) is a widely used multiscalar (30, 90, 180, 360, 720 d) drought index based on water balance combining precipitation and potential evapotranspiration derived from climate data (Beguería et al., 2014; Vicente-Serrano et al., 2010). The positive (negative) values of SPEI indicate wet (dry) condition. In this study, we used the 3 month SPEI index to monitor the agricultural drought conditions in the Amazon rainforest. Figure 1b shows the time series of the 20th percentile of regional SPEI index. The indices less than or equal to -1.5 are labeled with red color bars, indicating the occurrence of a severe drought, while yellow color bars refer to non-drought conditions with the index greater than -1.5. In this study, the severe dry months were considered as experiencing a drought if they were extended to more than three months (more than 3 consecutive red color bars in the plot) (Fig. 1b). Therefore, according to the SPEI index, drought events occurred in 2010, 2012, 2014 and 2015, with the longest duration lasting for eight months in 2010 and the strongest SPEI signal exceeding -2.5 in 2015, indicating extreme drought conditions during these years, which aligns with other literatures (Espinoza et al., 2024; Papastefanou et al., 2022).

Here we employed GCHP version 13.2.1, with stretch grid factor 4 targeted in 3° S, 60° W and 72 vertical levels in cubed-sphere grid C48, which allowed us to resolve the spatial resolution in the Amazon rainforest to ∼0.5° (∼50km). We adopted data from The Modern-Era Retrospective analysis for Research and Application, Version 2 (MERRA-2) (Gelaro et al., 2017) as the meteorological input (http://wiki.seas.harvard.edu/geos-chem/index.php/List_of_MERRA-2_met_fields, last access: 19 August 2026). The emissions in GCHP are handled by The Harvard-NASA Emission Component (HEMCO version 3.1.1) (Lin et al., 2021), which integrates various data inventories and scale factors. We implemented The Community Emissions Data System, Version 2 (CEDSv2) (McDuffie et al., 2020) for global base anthropogenic emissions and The Model of Emissions of Gases and Aerosols from Nature version (MEGAN) (Guenther et al., 2012) for biogenic emissions in all simulations.

Two sets of simulations ran from January 2009–December 2015, with 2009 used as model spin-up and 2010–2015 retained for analysis. This study period was chosen because it encompasses multiple intense fire seasons, including the major drought years 2010 and 2015, which are among the highest fire-emission years in the past decade. A comparison of regional DMB from the full GFEDv4s record (1997–2019) indicates that 2010 and 2015 are among the highest-emission years in the time series, while 2011–2014 spans lower to intermediate fire activities, suggesting that the 2010–2015 window samples both typical and extreme dry-season fire variability in the region (Fig. S1 in the Supplement). Additionally, deforestation rate remains low and stable during the experimental period (Fig. 1a), providing a quasi-constant anthropogenic land-use background that allows us to more cleanly isolate the influence of drought-driven variability on fire emissions, air quality, and public health impacts by comparing drought years (2010, 2012, 2014 and 2015) and non-drought years (2011 and 2013). Within this window, we could capture substantial interannual variability in drought conditions while avoiding the confounding effects of large swings in deforestation extent, which would complicate attribution of observed changes in air pollutant exposure and health burden. This design enables us to quantify the individual effects of droughts on fire-related air quality and public health under a relatively stable deforestation regime.

The first set of simulations (“fire-on”) utilized the GFEDv4s (Randerson et al., 2018) monthly DMB data in 0.25° to represent global fire emissions (Fig. 1c), while the second set of simulations (“fire-off”) were the control with all global biomass burning emissions turned off. We use monthly mean fire emissions rather than daily fields because our primary focus is on seasonal and interannual differences between drought and non-drought years, and running the full 2009–2015 period with higher-frequency emissions would substantially increase computational cost due to the large input/output burden for emission fields. Within the 2010–2015 study period, the DMB recorded by GFEDv4s peaks in every August or September in each year, and the peak magnitudes in identified drought years (2010, 2012, 2014 and 2015) are generally substantially higher than that in the two non-drought years (2011 and 2013), with 2010 showing the largest peak among all years. By utilising the GFEDv4s data in GCHP, we obtained the fire-induced chemical emissions by multiplying DMB with the species-specific and fire-type-specific emission factors. We then analyzed a temporal analysis of shortlisted chemicals in monthly level from January 2010–December 2015 within the Brazilian Amazon area (5° N–15° S, 47–73° W) and spatial analysis of these chemicals in monthly level from January 2009–December 2015 in South America (15° N–25° S, 40–90° W).

We compared the simulated surface PM2.5 and ozone concentrations from the two sets of simulations (“fire-on” and “fire-off”) together with satellite-observed PM2.5 concentrations to quantify the fire-induced health impacts in terms of premature deaths. The PM2.5 concentrations in GCHP were calculated as the sum of ammonium, inorganic nitrates, sulfate, black carbon, organic carbon, dust aerosol (first two size bins), fine sea salt aerosol (first size bins) and secondary organic aerosol (SOA). The full definition and detail equations used in GCHP can be found on the GEOS-Chem wiki (http://wiki.seas.harvard.edu/geos-chem/index.php/Particulate_matter_in_GEOS-Chem, last access: 19 August 2026).

To obtain realistic absolute PM2.5 levels while retaining the model-derived fire signals, we regridded the satellite-derived monthly PM2.5 concentrations (V5.GL.04), provided by the Atmospheric Composition Analysis Group of the Washington University in St. Louis (Shen et al., 2024) and adopted it as the baseline exposure field for health impact analysis (Fig. S2 in the Supplement). We then coupled it with a monthly grid-specific correction factor derived from the ratio between fire-on and fire-off simulated surface PM2.5 to construct a corresponding “no fire PM2.5” exposure field for the baseline exposure field, the field is given by: 1No firePM2.5=GCHP fire offPM2.5GCHP fire onPM2.5×satellitePM2.5, The GEMM NCD+LRI model estimates PM2.5-induced premature deaths using age-specific and sex-specific hazard ratios, baseline mortality and population data (Burnett et al., 2018). Applying the two exposure fields, the hazard ratio (HR) is given by: 2HR(z)=exp⁡θlog⁡zα+11+exp⁡-z-uv, where 3z=max⁡(0,PM2.5-2.4µgm-3), and θ, α, u, v are disease-specific and age-specific parameters describing the shape of hazard ratio function with 2.4 µgm-3 as the counterfactual PM2.5 concentration described by Burnett et al. (2018), corresponding to the lowest PM2.5 observed across the 41 cohorts, which is prescribed by GEMM. The number of PM2.5-induced premature death (M) is given by: 4M=HR-1HR×B×P×A, where B is the baseline mortality rate, P is the age-specific and sex-specific population, A is the grid area in each grid cell. Baseline mortality rates are a disease-specific paramters for 12 age groups (25–29, 30–34, 35–39, 40–44, 45–49, 50–54, 55–59, 60–64, 65–69, 70–74, 75–79, 80 or above) and two sex groups (male and female), obtained from Global Burden of Disease (GBD) (https://vizhub.healthdata.org/gbd-results/, last access: 19 August 2026). Detailed information can be found in supplementary material (Table S2 in the Supplement). Population data for each age groups and sex groups were derived from the 2010 population density estimated by Gridded Population of the World, version 4, revision 11 (GPWv4.11) (Center for International Earth Science Information Network, 2018). The aggregated population density is detailed in supplementary material (Fig. S3 in the Supplement). This dataset was produced as global rasters at 30 arc-second horizontal resolution and aggregated to 2.5 arc-minute, about 4.6 km at the equator. We regridded this population dataset to match our simulation output from GCHP for futher analysis. Finally, the difference in estimated premature deaths between two cases was interpreted as the mortality attributable to fire-emitted PM2.5.

To estimate premature deaths induced by ozone exposure, we utilize established exposure-response relation from previous studies (Anenberg et al., 2010; Malley et al., 2017; Seltzer et al., 2018). 5HR=exp⁡βΔY,6AF=1-exp⁡-βΔX,7M=AF×B×P×A, where 8ΔX=max⁡(0,[O3]-counterfactual concentration), and HR is the hazard ratio reported by previous epidemiological study, ΔY is 10 ppb from epidemiological studies conducted by Jerrett et al. (2009) and Turner et al. (2016), AF is the attributable fraction of ozone exposure, ΔX is the simulated ozone exposure with counterfactual concentration, β is the exposure-response factor, M is the number of ozone-induced premature deaths, B is the baseline mortality rate and P is the age-specific and sex-specific population, and A is the grid area in each grid cell. The baseline mortality and population data used here are consistent with those employed for estimating premature deaths caused by PM2.5 exposure.

In this study, we used the averaging metric and cause-specific risks recorded by Jerrett et al. (2009) and Turner et al. (2016). Jerrett et al. (2009) used the April-September average of daily 1 h maximum ozone concentration (6mMDA1) with counterfactual concentration 33.3 ppb, while Turner et al. (2016) focused on the annual average daily maximum 8 h ozone concentration (AMDA8) with counterfactual concentration 26.7 ppb. These ozone counterfactual concentrations were taken directly from the underlying epidemiological studies and were used to define the shape of the concentration–response function, rather than to represent a modeled “no-fire” ozone level. We applied these prescribed counterfactuals and exposure–response functions consistently to both the fire-on and fire-off simulations, and attributed the difference in resulting premature deaths to the contribution of fire emissions. For respiratory diseases, the hazard ratios of 1.040 (95 % CI: 1.013, 1.067) from Jerrett et al. (2009) and 1.12 (95 % CI: 1.08, 1.16) from Turner et al. (2016). In addition, we applied hazard ratio of 1.03 (95 % CI: 1.01, 1.05) for cardiovascular mortality, also reported by Turner et al. (2016). Because the epidemiological studies underlying our exposure–response functions control for co-pollutants, we calculated the premature deaths attributable to PM2.5 and ozone independently and summed them to represent the total fire-induced health burden.

We evaluated the ozone concentration against ground measurements over the Amazon region. The ground measurements of ozone were obtained from the Tropospheric Ozone Assessment Report (TOAR) (Gaudel et al., 2018). In the first phase of TOAR, there are five ground stations in the Brazilian Amazon monitoring surface ozone concentration. “Br-am01”, “br-am04” and “br-am05” are located near Manaus. “Br-am02” is located near Santarém. “Br-am03” is near Porto Velho. These station locations are available in the supplementary material (Table S3 in the Supplement). These measured ozone concentrations are compared with the GCHP-simulated results in the bottom layer of the nearest grid cell (Fig. S4 in the Supplement). Among the five stations, the best estimation is station “br-am02”, in which the modeled ozone level from September 2014–September 2015 is highly similar to the measured concentration. Additionally, GCHP is able to reproduce the similar ozone level in station “br-am01” during late 2010 but slightly overestimates the ozone level during 2014. For the other three stations, the model successfully reproduces the temporal pattern but overestimates ozone levels. As these five TOAR stations are at different altitudes, ranging from 25–261 m, the altitude difference between the measurement sites and GCHP simulation grid cells may cause such overestimation. Moreover, performance discrepancies may also be caused by the plume height. In reality, especially during drought years, fire plumes from tropical forest and savanna burning in this region tend to remain relatively low, with characteristic heights decreasing from about 1100–800 m (Gonzalez-Alonso et al., 2019). The standard GCHP configuration used here likely injects a larger fraction of emissions at higher altitudes through its default plume-rise and vertical mixing treatment (Zhu et al., 2018). This vertical mismatch can affect the comparison with measurements because plume height controls dilution, photolysis rates, and the distribution of precursors and radicals, thereby influencing in-plume ozone production and downwind ozone levels (e.g., Alvarado et al., 2009; Val Martin et al., 2010). At present, we lack a quantitative estimate of how large this effect is in our specific simulations and therefore treat it as an important but unquantified source of uncertainty.

We also validated the simulated surface PM2.5 concentrations by comparing them with satellite-derived surface PM2.5, which is from a global monthly mean dataset with 1° spatial resolution (named as V5.GL.04), provided by the Atmospheric Composition Analysis Group of the Washington University in St. Louis (Shen et al., 2024). This dataset combines aerosol optical depth from multiple satellite instruments with GEOS-Chem simulations and ground-based AERONET sun-photometer observations to estimate surface PM2.5. Because GEOS-Chem contributes to the construction of this product, the PM2.5 evaluation is not fully independent in a strict sense; however, the inclusion of AERONET and multiple satellite retrievals provides an observationally constrained, regionally representative benchmark that is more informative than the very sparse in situ PM2.5 monitoring network currently available in the Amazon. We compared the performance between dry (July–November) and wet (January–May) season for every year, including both drought and non-drought years (Fig. S5 in the Supplement). Despite the overall underestimation of PM2.5 concentrations, there are significant seasonal variations because the contribution from wildfire emissions to regional PM2.5 concentration is much higher during the dry season. GCHP with GFEDv4s fire emission inventory has a better performance in the dry season and overall underestimate surface PM2.5 when compared to satellite data, especially in non-drought years (2011 and 2013). Among all drought years, the model performs better in years with severe drought, with R2=0.55 in 2010 and R2=0.42 in 2015. The high concentrations are likely dominated by fire emissions during the dry season with severe droughts, and therefore reduce the discrepancies from other emissions. Figure S6 in the Supplement illustrates the time series comparison between the simulated surface PM2.5 from GCHP with GFEDv4s and the satellite-derived surface PM2.5 obtained from V5.GL.04 between January 2010–December 2015. Although model underestimation is noticeable in the temporal patterns, they show a strong correlation, mirroring the trends as in the satellite-derived data. This indicates that, despite the general underestimation, the model captures the seasonal dynamics of PM2.5 variation and aligns well with satellite observations.

To investigate the underlying sources of biases between simulated PM2.5 concentrations and satellite observations, we apply a machine learning algorithm, specifically the Light-GBM model. By treating these biases as the predicted variable, we can systematically analyze how various factors contribute to the discrepancies observed. In this model, we utilized several key predictors, including ERA5 meteorological variables, such as surface temperature, wind speed, and humidity, together with detailed emission inventories, such as GFED and CEDS, that provide information on pollution sources (Fig. S7 in the Supplement). The Light-GBM model revealed that surface wind speed is the strongest predictor, indicating its significant influence on the observed bias. Relative humidity and surface temperature are the other two important variables, suggesting meteorological condition is the main cause of the model-observation bias. Therefore, we applied satellite-derived monthly PM2.5 concentrations (V5.GL.04) and a monthly grid-specific correction factor generated from the simulations to get better estimation for the health impact analysis.

CO is harmful because of its ability to displace oxygen in human blood cells and thus reduce blood oxygen levels (Byard, 2019). Figure 2a shows its emissions in the Amazon (5° N–15° S, 47–73° W), which are highly dependent on the incomplete combustion from deforestation fires. Therefore, CO emissions in the fire-off simulations remain at low levels (9×10-12kgm-2s-1), while the emission pattern in fire-on simulation was highly similar to the pattern of DMB shown in Fig. 1. The annual CO emission peaks were either in August or September because of the high fire occurrences during the dry season. The annual emission peaks in drought years (5.4-8.8×10-10kgm-2s-1) are generally two to three times higher than those in non-drought years (2.8-3.0×10-10kgm-2s-1), while the emission in 2010 is also an outlier, reaching 2.0×10-9kgm-2s-1 as the annual peak. Figure 2b shows the fire-on and fire-off monthly mean CO concentration from 2010–2015. The concentration in the fire-off simulation remains low and stable at level of 0.095 ppm while the concentration in fire-on simulation fluctuates and also peaks in every August and September, identical to the emission pattern. Due to the higher emission rate in dry season with drought events, the annual peak of the Amazon monthly mean CO concentration in 2010, 2012, 2014 and 2015 is consequently higher than that in 2011 and 2013. The average magnitude of annual peak during drought years is 0.29 ppm and that during non-drought years is just 0.18 ppm. Therefore, by comparing it to the concentration in fire-off simulation (0.095 ppm), fire emission contributes to about 50 % of the annual peak CO concentration in non-drought years and such proportion would further increase to about 67 % in drought years. Figure 3a shows the dry season mean CO concentration for each year from the fire-on simulation. These spatial concentration plots show that the high CO concentration pattern is mostly localized, suggesting that the concentration is associated with fire occurrence, which agrees with our finding from the time series plots.

Figure 2c shows the simulated Amazonian monthly mean PM2.5 concentration in fire-on and fire-off simulations from 2010–2015. PM2.5 concentration is similar to the CO concentration pattern because both of them are the primary pollutants generated by incomplete combustion during deforestation and savanna fires. The PM2.5 concentration in the fire-off simulation is stable and stays between 6–10 µgm-3 throughout the study period while the concentrations in the fire-on simulation are higher during dry seasons in both non-drought and drought years. The annual peak of PM2.5 concentration in non-drought years is around 12 µgm-3, while the average annual peak in drought years is 26 µgm-3 with an extreme high record reaching 46.4 µgm-3 in 2010. The fire-induced PM2.5 accounts for about 50 % of the regional concentration during the dry season in non-drought years, and enhances the dry season concentration in drought years to a level of two to four times higher than that in non-drought years. To further understand the relationship between droughts and PM2.5, we compared between the dry season mean PM2.5 concentration and SPEI drought index (Fig. 2d). As the negative value of SPEI indicates drought occurrence, the negative slope (-4.83) in linear regression suggests that PM2.5 concentration increases with the intensity of drought. The t-value and p-value are -6.23 and 9.81×10-5 respectively with an outlier in 2010, suggesting we can reject the null hypothesis of no difference under drought and non-drought conditions. Therefore, fire emissions lead to regional PM2.5 sources and droughts worsen the pollution by influencing fires in the Amazon forest. Figure 3b shows the dry season mean PM2.5 concentration from 2010–2015 from the fire-on simulation. Similar to the CO spatial pattern, the PM2.5 concentration pattern is mostly localized and rapidly diluted, suggesting PM2.5 concentration is associated with fire emission with a short lifetime in the surface layer.

NOx is a group of highly-reactive gases including nitric oxide (NO) and nitrogen dioxide (NO2), emitted from both anthropogenic sources (such as industrial combustion and transportation) and natural sources (such as lightning and wildfires). NOx, together with volatile organic compounds (VOCs) and solar radiation, is the precursor of tropospheric ozone. Figure 4a shows the Amazon (5° N–15° S, 47–73° W) monthly mean NOx emission in both simulations. Similar to CO and PM2.5, NOx emission in the fire-off simulation remains low and stable with 5.0×10-12kgm-2s-1 throughout the simulation period, while NOx emission in the fire-on simulation increases with similar pattern as enhanced DMB, indicating that NOx emissions in Brazilian Amazon are also dominated by wildfires. Additionally, drought events amplify the NOx emission to two to four times (2.0-3.4×10-11kgm-2s-1) higher than that during non-drought years (1.2-1.3×10-11kgm-2s-1), in which 2010 is the outlier with the annual peak emission rate up to 6.6×10-11kgm-2s-1. Figure 4b shows the simulated isoprene, one of the major biogenic VOC (BVOC), emissions in Brazilian Amazon, which is simulated by MEGAN in GCHP. In our configuration, MEGAN biogenic emissions are independent of wildfire impacts, so isoprene emissions are identical in the fire-on and fire-off simulations and peak every September with magnitudes between 5.2-8.1×10-10kgm-2s-1. In reality, wildfire smoke and aerosols may influence vegetation and thus biogenic emissions, but such feedbacks are not included in the current GEOS-Chem model. Despite the identical emissions, the annual minimum of monthly mean isoprene concentration in the fire-on simulation is found between August and September and is 3.2–4.7 ppb, which is about 1–2 ppb lower than in the fire-off simulation over the same period (Fig. 4c). We also estimate the ratio of NOx emission over isoprene emission in Fig. 4d. The ratio remains in a low proportion (below 0.16) (e.g. Ren et al., 2022), suggesting that NOx is the limiting factor of surface ozone formation in the Amazon and fire emissions are the primary factor to generate NOx and ozone to worsen the regional air quality.

Figure 5a shows the simulated Amazon monthly mean ozone concentration in fire-on and fire-off simulations from 2010–2015. There is an annual concentration peak in every August in both fire-on and fire-off simulations. However, the annual ozone concentration peak is about 14 ppb in the fire-off simulation, while it ranges from 22.1–24.2 ppb during non-drought years and 24.9–35.1 ppb during drought years in fire-on simulation. The fire-induced ozone concentration contributes to about one-third of the total concentration in non-drought years while the contribution ranged from one-third to more than 50 % in drought years. To better demonstrate the impacts of droughts on ozone, we compare dry season mean ozone concentration and SPEI drought index (Fig. 5c). As the negative value of the SPEI indicates drought occurrence, the negative slope (-3.39) in linear regression suggests that ozone concentration increases with the intensity of drought. The t-value and p-value are -17.89 and 6.38×10-9 respectively, suggesting we can reject the null hypothesis of no difference between drought and non-drought conditions.

Despite the lower contribution fraction-wise of fires to ozone concentration than for CO and PM2.5, ozone pollution induced by fire emission is a regional air quality issue. Figure 5b shows the dry deposition velocity of ozone in the Amazon, which is mainly determined by meteorological and land use conditions. With stable dry deposition velocity throughout the simulation period, ozone dry deposition rates mainly depend on its concentration, leading to higher deposition rates in drought years than non-drought years and consequently causing more damage to the vegetation and ecosystems in drought years. Figure 6 shows the dry season mean ozone concentration from 2010–2015 from the fire-on simulation. These spatial concentration plots show that the fire- influenced ozone was advected toward the southern part of the Brazil, suggesting the affected area of ozone is larger than that of PM2.5 and causing regional air pollution. Moreover, the fire contribution to ozone concentration shows a larger impact area during dry season in drought years than during non-drought years. Therefore, fire emissions are suggested to be one of the dominant sources deteriorating regional ozone pollution, and drought events would further worsen it in both pollutant magnitude and impact area by amplifying fire emissions in Amazon forest.

Utilizing the GEMM NCD+LRI model with GCHP simulated grid-specific correction on satellite-derived monthly mean PM2.5 concentrations, population density, and cause-specific baseline mortality rates, we also estimated the premature deaths induced by PM2.5 exposure in Brazil from 2010–2015. The total PM2.5-induced premature deaths showed a fluctuating trend over the simulated period (Fig. 7). The PM2.5 induced death is estimated to be 22 595 in 2010 and 22 931 in 2011. There is a slight decrease to 21 621 in 2012, 22 081 in 2013 and 22 068 in 2014 but then shows substantial increase to 23 113 in 2015. The total ozone induced premature mortality is comparably lower than that of PM2.5, causing 4068 (95 % CI: 2012–6080) deaths in 2010, 3553 (95 % CI: 1757–5313) in 2011, 4499 (95 % CI: 2225–6727) in 2012, 3697 (95 % CI: 1827–5534) in 2013, 4844 (95 % CI: 2399–7229) in 2014 and 4757 (95 % CI: 2356–7102) in 2015 (Fig. 7).

Utilizing averaging metrics and exposure-response relationships from Jerrett et al. (2009) and Turner et al. (2016), we estimated the annual premature mortalities induced by ozone exposure in Brazil with population density and cause-specific baseline mortality rates between 2010 and 2015. Using the methods of Jerrett et al. (2009), the premature mortalities due to respiratory diseases caused by exposure to ozone in Brazil from 2010–2015 are 566 (95 % CI: 146–832), 427 (95 % CI: 110–630), 479 (95 % CI: 123–707), 389 (95 % CI: 100–574), 607 (95 % CI: 157–892) and 524 (95 % CI: 135–772) (Fig. 8). When using Turner et al. (2016)'s exposure-response relationship, the estimated premature mortalities due to respiratory diseases caused by ozone exposure are higher, which are 1724 (95 % CI: 1242–2254), 1506 (95 % CI: 1061–1970), 1909 (95 % CI: 1344–2497), 1573 (95 % CI: 1106–2061), 2044 (95 % CI: 1444–2664), and 2011 (95 % CI: 1420–2623) from 2010–2015 (Fig. 10). Apart from respiratory diseases, Turner et al. (2016) included cardiovascular diseases into premature mortalities estimations, which are 2343 (95 % CI: 798–3827) in 2010, 2046 (95 % CI: 696–3343) in 2011, 2589 (95 % CI: 881–4231) in 2012, 2124 (95 % CI: 722–3474) in 2013, 2800 (95 % CI: 955–4566) in 2014, and 2745 (95 % CI: 936–4479) in 2015 (Fig. 7).

In the following discussion, Turner's exposure-response relationship will be prioritized, as it includes both cardiovascular and respiratory diseases, offering a more comprehensive assessment of health impacts compared to Jerrett's model, which focuses solely on respiratory diseases. Additionally, Turner's model, established in 2016, provides more up-to-date and relevant data compared to Jerrett's 2009 study, making it a more robust choice for this analysis.

Although there is not an obvious trend in the total number of premature deaths, the difference in PM2.5- and ozone-induced premature deaths between fire-on and fire-off simulations align closely with the drought index (SPEI). Figure 9a shows that the fire-induced premature mortalities due to exposure to PM2.5 and ozone were significantly higher in drought years than in non-drought years. By comparing the estimated premature deaths with control simulations, we found that fire-induced PM2.5 and ozone cause 6.0 % and 18.6 % more deaths in non-drought years, and 8.9 % and 24.4 % more in drought years. Such differences in terms of premature mortalities are obviously larger in years with strong drought events (2010 and 2015). Figure 9b and c also demonstrates the correlation between the SPEI drought index and fire-induced premature mortalities due to exposure to PM2.5 and ozone. As the negative value of SPEI indicates drought occurrence, the negative slopes in linear regression from both graphs suggest that the number of deaths caused by PM2.5 and ozone both increases with drought intensity. For PM2.5, the t-value and p-value are -10.08 and 1.48×10-6. For ozone, the t-value and p-value are -9.20 and 3.4×10-6. Such statistics suggest that both null hypotheses can be rejected and that there are significant statistical differences of fire-induced premature mortalities between drought years and non-drought years.

Furthermore, despite premature deaths induced by exposure to ozone being lower than that to PM2.5, the health impact induced by ozone is more sensitive and closely correlated to drought conditions with an R2=0.56. The major reason behind is the difference in lifetime between these two pollutants. Ozone and its precursors could be advected to highly populated areas with longer lifetimes, while PM2.5-induced impacts were mostly localized within more scarcely populated areas.