2.1.1 GFED3.1

The total dry matter consumed by biomass burning in GFED3.1 (van der Werf et al., 2010) is estimated by the multiplication of the MODIS burned area product at 500 m spatial resolution (Giglio et al., 2010, for the MODIS era) and fuel consumption per unit burned area, the latter being the product of the fuel loads per unit area and combustion completeness. This estimation is conducted using the Carnegie–Ames–Stanford approach (CASA) biogeochemical modeling framework that provides estimates of biomass in various carbon “pools” including leaves, grasses, stems, coarse woody debris, and litter. Fuel loads in CASA are estimated according to carbon input information on vegetation productivity, and carbon outputs through heterotrophic respiration, herbivory, fires, and tree mortality (Giglio et al., 2010; van der Werf et al., 2010). Then, the biomass burning emission of a given species is calculated by multiplying the dry matter with an emission factor of that species (EF, in grams of species per kilogram of dry matter burned). The EF used in GFED3.1 (and most of the other datasets) is mainly chosen from Andreae and Merlet (2001) and/or Akagi et al. (2011), but may also be obtained from various other sources. The GFED3.1 dataset can be accessed through the following link: https://daac.ornl.gov/VEGETATION/guides/global_fire_emissions_v3.1.html (last access: 17 January 2020).

2.1.2 GFED4s

Compared to GFED3.1, the latest GFED version, GFED4s, has a few significant upgrades as described in detail by van der Werf et al. (2017), including (1) additional burned area associated with small fires which were previously omitted by the burned area product but now are compensated for by including the active fires to augment the burned area product MCD64A1 (Giglio et al., 2013; Randerson et al., 2012); (2) a revised fuel consumption parameterization optimized using field observations (e.g., van Leeuwen et al., 2014); and (3) further dividing forest into temperate and boreal forest ecosystems and applying different sets of emission factors. Among the existing BB emission datasets, GFED4s has hitherto been the most widely used by modeling communities, such as the Coupled Model Intercomparison Project Phase 6 (CMIP6; Van Marle et al., 2017) and AeroCom phase III experiments (https://wiki.met.no/aerocom/phase3-experiments). The link to the GFED4s dataset is http://www.globalfiredata.org (last access: 17 January 2020).

2.1.3 FINN1.5

The FINN1.5 biomass burning emission dataset is developed from its previous version FINN1 (Wiedinmyer et al., 2011) with several updates. It uses satellite observation of active fire (with a confidence level greater than 20 %) and land cover from the MODIS instruments on board the NASA Terra and Aqua polar-orbiting satellites, together with the estimated fuel consumption to derive biomass burning emissions. The burned area in each active fire pixel is assumed as 1 km², except for grasslands and savannas where it is assigned a value of 0.75 km². The fuel consumption at each fire pixel is estimated according to its generic land use–land cover type (LULC) which is assigned using values updated from Table 2 of Hoelzemann et al. (2004) in the various world regions based on Global Wildland Fire Emission Model (GWEM). With the estimated burned area, fuel consumption, and EF of individual species, the daily global open biomass burning emissions of each species are then calculated at a 1 km spatial resolution. The FINN1.5 emissions dataset is archived at http://bai.acom.ucar.edu/Data/fire/ (last access: 17 January 2020).

2.1.4 GFAS1.2

The GFAS (Kaiser et al., 2012) estimates dry matter combustion rate by multiplying FRP and biome-specific conversion factors (units: kilograms of dry matter per megajoule). The global distribution of FRP observations is obtained from the MODIS instruments on board the Terra and Aqua satellites and are then assimilated into the GFAS system. The gaps in FRP observations, which are mostly due to cloud cover and spurious FRP observations of volcanoes, gas flares, and other industrial activity, are corrected or filtered in the GFAS system. Eight biome-specific conversion factors are calculated by linear regressions between the GFAS FRP and the dry matter combustion rate of GFED3.1 in each biome (see Table 2 and Fig. 3 in Kaiser et al., 2012). The biomass burning emission of a given species is calculated by multiplying the dry matter with an emission factor of that species. More information on the latest GFAS product can be found at https://apps.ecmwf.int/datasets/data/cams-gfas/ (last access: 17 January 2020).

2.1.5 FEER1.0

The FEER1.0 (Ichoku and Ellison, 2014) multiplies its emission coefficients C_e with MODIS FRP data that have been preprocessed and gridded in the GFAS1.2 analysis system (Kaiser et al., 2012) to derive biomass burning aerosol emission rates. The C_e in FEER1.0 for smoke aerosol total particulate matter (TPM) was derived through zero-intercept regression of the emission rate of smoke aerosol (i.e., R_sa) against the corresponding FRP (Ichoku and Kaufman, 2005; Ichoku and Ellison, 2014) at pixel level within each grid. C_e corresponds to the slope of the linear regression fitting. In the FEER methodology, R_sa is estimated through a spatiotemporal analysis of MODIS AOD data along with wind fields from the NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) reanalysis dataset (Rienecker et al., 2011). The smoke aerosol C_e in FEER1.0 is available at $\mathrm{1}{}^{\circ }\times \mathrm{1}{}^{\circ }$ spatial resolution global grid and covers most of the land areas where fires have been detected by MODIS at least 30 times during the period 2003–2010 (Ichoku and Ellison, 2014) to ensure statistical representativeness. In the current version of the FEER1.0 emission dataset, C_e values for other smoke constituents, say OC, at each grid cell are obtained by scaling the C_e of smoke aerosol according to the ratio of their emission factors, such as EF_oc to EF_sa (i.e., ratio of emission factor for OC to that for total smoke aerosol). The FEER1.0 dataset is available at http://feer.gsfc.nasa.gov/data/emissions/ (last access: 17 January 2020).

2.1.6 QFED2.4

In QFED (Darmenov and da Silva, 2015) biomass burning aerosol emissions are estimated using gridded MODIS Terra and MODIS Aqua FRP and emission coefficients C_e, which are the product of a constant value C₀ (1.37 kg MJ⁻¹, reported by Kaiser et al., 2009), satellite factor, and biome-specific scaling factor. The scaling factors used in QFED2.4, the version applied in this study, were obtained by comparing AOD from the Goddard Earth Observing System model (GEOS) and MODIS aerosol product in multiple sub-regions (Fig. 4 in Darmenov and da Silva, 2015). These scaling factors were further reduced to four values representative of fires in savanna, grassland, tropical forests, and extratropical forests – 1.8, 1.8, 2.5, and 4.5, respectively. The QFED2.4 used a sequential model with temporally damped emissions to estimate the emissions in cloudy areas. QFED is the standard fire emissions dataset in the near-real-time GEOS data assimilation system and the MERRA-2 reanalysis (Randles et al., 2017). QFED2.4 emissions are available from https://portal.nccs.nasa.gov/datashare/iesa/aerosol/emissions/QFED/v2.4r6/ (last access: 17 January 2020).

2.2.1 Description of the NASA GEOS model

The GEOS model consists of an atmospheric general circulation model, a catchment-based land surface model, and an ocean model, all coupled together using the Earth System Modeling Framework (ESMF; Rienecker et al., 2011; Molod et al., 2015). Within the GEOS model architecture, several interactively coupled atmospheric constituent modules have been incorporated, including an aerosol and carbon monoxide (CO) module based on the Goddard Chemistry Aerosol Radiation and Transport model (GOCART; Chin et al., 2000, 2002, 2009, 2014; Colarco et al., 2010; Bian et al., 2010) and a radiation module from the Goddard radiative transfer model (Chou and Suarez, 1999; Chou et al., 2001). The GOCART module used in this study includes representations of dust, sea salt, sulfate, nitrate, and black and organic carbon aerosol species. A conversion factor of 1.4 is used to scale organic carbon mass to organic aerosol (OA), which is on the low end of current estimates (Simon and Bhave, 2012). More discussion on this conversion factor can be referred to in Sect. 4.3.

In this study the GEOS model (Heracles-5.2 version) was run globally on a cubed-sphere horizontal grid (c180, ∼ 50 km resolution) and with 72 vertical hybrid-sigma levels extending from the surface to ∼85 km for the year 2008. The reason we chose 2008 is because it is the year assigned as a benchmark year by the AeroCom community with which this study is associated; it is also because the AeroCom multi-model study of biomass burning led by Petrenko (mentioned in the introduction) also chose 2008 as a focus year. As such, the results from these two studies can be intercompared to draw some synthesized conclusions. In addition, 2008 was chosen because it is a neutral El Niño–Southern Oscillation (ENSO) year, which represents normal burning conditions. The model was run in a “replay” mode, where the winds, pressure, moisture, and temperature are constrained by the MERRA-2 reanalysis meteorological data (Gelaro et al., 2017), a configuration that allows a similar simulation of real events as in a traditional offline chemistry transport model (CTM) but exercises the full model physics for radiation, for example, and moist physics processes. We used the HTAP2 anthropogenic emissions (Janssens-Maenhout et al., 2015) that provide high-spatial-resolution monthly emissions. The BB emissions are uniformly distributed within the boundary layer without considering the specific injection height of each plume. All six BB emissions are daily emissions with the diurnal cycle prescribed in the model: the maximum is around local noon, which is more prominent in the tropics, and is gradually weakened in the extra-tropics (Randles et al., 2017). The natural aerosols are either generated by the model itself (i.e., wind-blown dust and sea salt) or come from prescribed emission files (i.e., volcanic and biogenic aerosols).

2.2.2 Experiment design

In order to investigate the sensitivity of the modeled AOD to different BB emission datasets, seven experiments were conducted with the GEOS model, differing only in the source of biomass burning emissions. The first six runs are GFED3.1, GFED4s, FINN1.5, GFAS1.2, FEER1.0, and QFED2.4, using the corresponding biomass burning datasets described above in Sect. 2.1. A seventh run is called “NOBB,” where the model is run without including biomass burning emissions.

2.3.1 MISR retrievals

We evaluated the simulated monthly AOD at 550 nm with the monthly level 3 total AOD data at the 558 nm wavelength from the Multiangle Imaging SpectroRadiometer sensor on board the EOS Terra satellite (Kalashnikova and Kahn, 2006; Kahn et al., 2010). We used MISR version 23 data products (MISR v23, with filename tagged as F15_0032) in half-degrees, which can be downloaded from the website https://eosweb.larc.nasa.gov/project/misr/mil3mae_table (last access: 17 January 2020).

2.3.2 AERONET sites

We also evaluated the modeled 3-hourly and monthly AOD at 550 nm and Ångström exponent (AE, 440–870 nm) with corresponding measurements from the ground-based AErosol RObotic NETwork (AERONET; Holben et al., 1998) sites situated in biomass burning source regions. AERONET Version 3 Level 2.0 data, which are cloud-screened and quality-assured aerosol products with a 0.01 uncertainty (Giles et al., 2019), were used in this study. The data can be downloaded from the website: https://aeronet.gsfc.nasa.gov/new_web/download_all_v3_aod.html (last access: 17 January 2020). The AERONET AOD at 550 nm is interpolated from the measurements at 440 and 675 nm. AE is calculated with AOD at 440 and 870 nm. We compared model simulations with AERONET data at 14 selected sites, representing the aerosol spatiotemporal characteristics at the different biomass burning regions shown in Fig. 1. The 14 regions were defined previously in GFED-related series of studies (e.g., van der Werf et al., 2006, 2010, 2017). Some regions, such as Northern Hemisphere South America (NHSA) and equatorial Asia (EQAS), have no AERONET sites with data measured in 2008; thus we also used the average of multiple years or climatology of AERONET AOD at each site for reference. Locations of these 14 selected AERONET sites are represented by the numbered magenta dots in Fig. 1.

3.1.1 Annual total

Figure 2 shows the spatial distributions of annual total biomass burning OC emissions in 2008 from the six BB emission datasets. The regions with high emission of OC in Africa, boreal Asia, and South America are pronounced in all six BB emission datasets, albeit to different degrees. The regional differences of the annual total biomass burning OC emissions in different BB emission datasets can be appreciated more quantitatively in Fig. 3. Relevant statistics for the six BB emission datasets in the 14 regions are also listed at the top of the panel in Fig. 3, with the mean of the six BB emission datasets in the first row (mean). We also used three different measures to quantify the spread of the annual total from the six BB emission datasets: (1) standard deviation (SD), (2) ratio of maximum to minimum ($max/min$), and (3) the coefficient of variation (cv, defined as the ratio of the SD to the mean). The cv values for the 14 regions are also ranked in Fig. 3 for easy reference (e.g., a ranking of 1 means that this region shows the least spread among the six BB emission datasets, while a ranking of 14 indicates that this region has the largest spread). The best agreements among the six emission datasets occurred in Northern Hemisphere Africa (NHAF), equatorial Asia (EQAS), Southern Hemisphere Africa (SHAF), and Southern Hemisphere South America (SHSA), which have the top cv ranks (1–4) and a relatively low $max/min$ ratio (a factor of 3–4). The worst agreements occurred in the Middle East (MIDE), temperate North America (TENA), boreal North America (BONA), and Europe (EURO), which have the lowest cv ranks (14–11) and a large $max/min$ ratio (a factor of 66–10). This diversity was mostly driven by the QFED2.4 emission dataset, which estimated the largest emission amount for almost all regions (except EQAS), especially in MIDE where the BB emission from QFED2.4 is more than 50 times higher than that from the two GFED versions. Globally, the QFED2.4 dataset showed the highest OC emission of 51.93 Tg C in 2008, which was nearly 4 times that of GFED4s at 13.76 Tg C (the lowest among the six BB datasets).

Figure 2The spatial distribution of annual total organic carbon (OC) biomass burning emissions for 2008 estimated by six biomass burning emission datasets (units: g m⁻² yr⁻¹). The global annual total amount for each dataset in 2008 is indicated in the parentheses.

Figure 3The regional annual total organic carbon (OC) biomass burning emissions for 2008 in six biomass burning emission datasets in 14 regions (units: Tg yr⁻¹). The global annual total amount is listed after the name of each dataset (GLOB_TOT). Relevant statistics for the six BB emission datasets in each region are also listed at the top of the panel in blue under the short name of each region, with the mean of the six BB emission datasets in the first row. Three different methods to measure the spread of the six BB emission datasets are shown as well: one absolute method, i.e., the standard deviation (SD) in the second row, and two relative methods, i.e., the ratio of max to min (i.e., maximum ∕ minimum) shown in the third row, as well as the coefficient of variation (cv), defined as the ratio of the SD to the mean, in the fourth row. The rankings of the regions reflecting the spread of the BB emissions datasets according to cv are shown in the fifth row (i.e., a ranking of 1 means that this region shows the least spread among the six BB emissions datasets, while a ranking of 14 indicates that this region has the largest spread among the 14 regions).

Download

Overall, two FRP-based BB emissions, QFED2.4 and FEER1.0, were a factor of 2–4 larger than the other BB datasets. This result is consistent with the findings of Zhang et al. (2014) over sub-Saharan Africa. It is worth noting that the BB emission amount of GFAS1.2 was close to that of GFED3.1, reflecting the fact that GFAS1.2 is tuned to GFED3.1 (described in Sect. 2.1.4). Globally, FINN1.5 yielded more OC emissions than the two GFED datasets and GFAS1.2 (e.g., 40 % larger than GFED4s). Regionally, FINN1.5 was generally comparable to the two GFED datasets in most of the regions, but was higher than them in the tropical regions, such as EQAS, Southeast Asia (SEAS), Central America (CEAM), and Northern Hemisphere South America (NHSA). Interestingly, FINN1.5 was even the largest among all six datasets over the EQAS region, which might be associated with its assumption of continuation of burning into the second day in that region (to be discussed in Sect. 4.1.2). The global OC emissions from GFED4s were lower than those from its GFED3.1 counterpart, although higher in several other regions, such as TENA, CEAM, NHSA, boreal Asia (BOAS), and central Asia (CEAS). Possible explanations for these differences among the six global BB emissions datasets are provided in Sect. 4.1.

3.1.2 Seasonal variation

Biomass burning is generally characterized by distinct seasonal variations in each of the 14 regions and globally, as shown in Fig. 4. Overall, there were four peak fire seasons across the regions: (1) during the boreal spring (March–April–May), fires peak in BOAS mainly because of forest fires (see the contribution of different fire categories in Table 3 of van der Werf et al., 2017); in CEAM, NHSA, and SEAS because of savanna and deforestation fires; and in central Asia (CEAS) mainly due to the agricultural waste burning to prepare the fields for spring crops. (2) During the boreal summer (June–July–August), fires peak in BONA and TENA, mostly due to wildfires that occur under the prevailing dry and hot weather, in EURO probably associated with the burning of agricultural waste. In addition, we found that fire peaked in MIDE in the three FRP-based datasets, i.e., QFED2.4, FEER1.0, and GFAS1.2. This might be associated with the failure to filter out the gas flares from the FRP fire product, especially in QFED2.4 (Darmenov and da Silva 2015). (3) During the austral spring (September–October–November), fires peak in the southern hemispheric regions of SHSA, SHAF, and AUST (Australia and New Zealand), associated with savanna burning (in addition to deforestation fires in SHSA). In SHSA, the two GFED versions peaked in August, 1 month earlier than the rest; (4) during the boreal winter (December and January), fires peak in NHAF, particularly along the sub-Sahel belt (Fig. 2), where savanna fires are associated with agricultural management and pastoral practices across that region (e.g., Ichoku et al., 2016b). Overall, all six BB emission datasets exhibited similar seasonal variations, although they differed in magnitude. In particular, it is noteworthy that in EQAS, the annual OC emissions from GFED4s were lower than those of GFED3.1 by 18 %, but higher by a factor of 2 in the month of August when peatland burning is predominant.

Figure 4Monthly variation in organic carbon (OC) biomass burning emissions for 2008 in six biomass burning emission datasets in 14 regions and the globally (i.e., GLOB, highlighted with a black box). The annual total emission is listed on the right side of each panel.

Download

For reference, biomass burning black carbon (BC) emissions are also shown, but in the Supplement (Figs. S1, S2 for annual total and Fig. S3 for seasonal variation), which exhibited features similar to those of OC. The amounts of biomass burning BC emission were almost proportional to their OC counterparts (about 1∕10 to 1∕15 of OC).

3.2.1 Global spatial distribution

Comparisons for September and April in 2018 are shown in Figs. 5 and 6, respectively, representing the peak biomass burning months in the Southern Hemisphere and many regions in the Northern Hemisphere, respectively. The MISR AOD is displayed in the top left panel and the model biases (model minus MISR) from the seven individual experiments are shown in the rest of the panels.

Figure 5(a) The spatial distribution of monthly mean AOD at 558 nm for September 2008 from MISR with white representing missing values. The global average value (ave) is shown in parentheses. Panels (b)–(h) are for GEOS model biases (i.e., model at 550 nm minus MISR at 558 nm) in seven model experiments, i.e., bias in (b) NOBB, (c) GFED3.1, (d) GFED4s, (e) FINN1.5, (f) GFAS1.2, (g) FEER1.0, and (h) QFED2.4.

Figure 6Same as Fig. 5 except for April 2008.

In September 2008, the high AOD observed from MISR (Fig. 5a) in the Southern Hemisphere was mostly attributable to biomass burning. A large fraction of Southern Hemisphere Africa (SHAF) featured high AOD (greater than 0.5). The area-averaged AOD over the entire SHAF was 0.331 (see Table S1 for the area-averaged MISR AOD in each region). The observed AOD peaked in central Africa (nearly 1.0) and gradually decreased westwards. A large negative model bias (−0.283) was found in the NOBB run over the region of SHAF (greenish shading in Fig. 5b; see Table S1 for the area-averaged model biases in each region). The negative bias was reduced most significantly in the QFED2.4 run (see Fig. 5h) to −0.044, followed by the FEER1.0 run (see Fig. 5g) to −0.079, but only to a limited extent in GFED4s and GFAS1.2 (see Fig. 5d and f, respectively), whose biases were still as large as −0.208.

In Southern Hemisphere South America (SHSA), where the area-averaged MISR AOD was 0.188, the maximum AOD was ∼0.7 in central Brazil (Fig. 5a). The negative bias averaged over SHSA was −0.132 in the NOBB run (Fig. 5b). The bias was most significantly reduced in the FEER1.0 run to −0.021 (Fig. 5g), but it appeared overcorrected in the QFED2.4 run to 0.020 (see reddish shading in Fig. 5h). The reduction of negative bias was again the least in the GFED4s run (Fig. 5d) and the GFAS1.2 run (Fig. 5f), whose biases were still as large as −0.081 and −0.080, respectively.

Being mixed with and often surpassed by other aerosol types, however, the contribution of biomass burning aerosols to the total AOD is hardly distinguishable from those of other sources in the peak biomass burning months in certain regions, such as April (Fig. 6) in the regions of Southeast Asia (SEAS), central Asia (CEAS), and boreal Asia (BOAS). Such complicated situations lead to the difficulties in evaluating the BB emission datasets with AOD observations, especially when the background AOD represented by the NOBB run was already overestimated, for example, in the region of CEAS, or when the MISR AOD was missing, for example, in the region of BOAS (where MODIS AOD was missing as well, not shown).

3.2.2 Seasonal variations in AOD at AERONET sites

In order to better quantify the sensitivity of the simulated AOD to the six different BB emission datasets, we further compared the simulated monthly AOD with the ground-based AOD observations from AERONET stations by choosing one representative station in each region (see Fig. 7: the panels representing the AERONET stations in Fig. 7 were arranged in a way that their placements correspond to those of their respective regions in Fig. 4 for easy reference). The exceptions are two regions NHSA and EQAS, where valid AERONET observations could not be found during 2008. Thus, we used the multi-year climatology of AOD at Medellín and Palangkaraya to represent NHSA and EQAS, respectively. We also included the climatology of AERONET AOD in the other 12 AERONET sites for reference. As shown in Fig. 7, the annual cycle of AOD in 2008 at available sites (thin brown bars) was similar to its respective climatology (thick light gray bars) to within 0.05. The MISR AOD was plotted for reference as a green diamond. In this section, the modeled monthly mean AOD was calculated by averaging over the modeled instantaneous AOD in each month, while the monthly AOD of AERONET and MISR was simply calculated by averaging over available observations in each month.

Figure 7Monthly variation in AOD (at 550 nm wavelength) for 2008 over 14 AERONET sites selected from the respective 14 regions (with the country of each site indicated in parentheses). The climatology of AERONET AOD (i.e., AERONET-clim) is represented by thick light gray bars and the monthly mean AERONET AOD for 2008 by thin brown bars, with their corresponding annual mean values shown in parentheses. MISR monthly mean AOD at 558 nm is represented by the green diamonds, and the seven GEOS experiments with different biomass burning emission options are represented by the lines in different colors. The corresponding annual ratios (ratio of model ∕ AERONET) listed on the right-hand side of each panel are estimated by averaging over monthly ratios.

Download

Contributions from non-BB emissions to the total AOD are represented by the NOBB experiment (black line in Fig. 7). Runs with different BB emission datasets showed almost identical AOD during non-biomass-burning seasons at each selected AERONET station in each region, thereby allowing their differences to be noticeable during the biomass burning peak seasons. At Alta Floresta in Brazil (Fig. 7(5)), Mongu in Zambia (Fig. 7(9)), and Chiang Mai Met Sta in Thailand (Fig. 7(12)), where the biomass burning emissions dominated the peak AOD, almost all experiments underestimated AOD during the respective peak biomass burning seasons. However, the fact is that the contribution of non-BB AOD was usually more than that of BB AOD during the burning seasons at most of the selected AERONET sites, except at the above three sites, Therefore, it is difficult to disentangle the effect of biomass burning on the total AOD in most situations, especially when the model has difficulty representing the non-BB AOD, leading, for example, to overestimation at three high-latitude (>55^∘ N) AERONET sites (the three panels in the top row of Fig. 7), i.e., Fort McMurray in Canada (Fig. 7(1)), Tõravere in Estonia (Fig. 7(6)), and Moscow_MSU_MO in Russia (Fig. 7(10)). However, it is apparent that the simulated AOD values based on QFED2.4 were overestimated during October and November at Fort McMurray in the USA, indicating that QFED2.4 overestimated BB organic carbon emission during these 2 months. In general, at most of the AERONET sites, the simulated AOD values based on QFED2.4 were the highest and closest to AERONET AOD during the corresponding peak of the biomass burning seasons, followed by FEER1.0 and FINN1.5 and then GFED3.1, GFEDv4, and GFAS1.2.

3.3.1 Alta Floresta in Brazil (Southern Hemisphere South America, SHSA)

The monthly mean AOD observed from AERONET at Alta Floresta is 0.47 during September 2008 (Fig. 8a). It shows that the simulated AOD from all six experiments captured the high aerosol episode observed in the AERONET dataset during 13 September (AERONET AOD is about 1.0–1.2). The simulation with QFED2.4 BB emission produced the closest agreement with the AERONET observed AOD with an average ratio of 1.00. In contrast, the simulated AOD with FEER1.0 (ratio =0.73), FINN1.5 (ratio =0.55), GFAS1.2 (ratio =0.42), GFED3.1 (ratio =0.40), and GFED4s (ratio =0.36) tended to be underestimated most of the time. All experiments showed relatively low skill at capturing the temporal variability of the observed AOD at Alta Floresta (corr =0.24–0.60). The Ångström exponent (AE: an indicator of particle size) from AERONET is 1.66 (not shown), indicating that small particles, most likely those from smoke, dominated the total aerosol loading at Alta Floresta (Eck et al., 2001). All experiments matched the observed AE (not shown).

Figure 8Characteristics of the observed and the simulated aerosols at Alta Floresta during September 2008: (a) the 3-hourly time series of AOD at 550 nm. The AERONET AOD is represented by vertical gray bars, and the outputs from the six model experiments are represented by the color curves. The relevant statistics are listed: ave is the monthly average, ratio is the fraction of the simulated to the observed AOD at all observed hours, corr is correlation coefficient between the observed and the simulated AOD, and RMSE is root-mean-square error. (b) The 3-hourly time series of OC column mass density over the grid box where Alta Floresta is located (units: $\mathrm{1}\times {\mathrm{10}}^{-\mathrm{6}}$ kg m⁻² or mg m⁻²). (c) Same as panel (b) but for biomass burning OC emission rate (units: $\mathrm{1}\times {\mathrm{10}}^{-\mathrm{9}}$ kg m⁻² s⁻¹ or µg m⁻² s⁻¹). (d) MODIS Terra true-color image around Alta Floresta on 13 September 2008, overlaid with the active fire detections as red dots (image credit: https://aeronet.gsfc.nasa.gov/cgi-bin/bamgomas_interactive, last access: 29 December 2018, and https://worldview.earthdata.nasa.gov, last access: 29 December 2018).

The OC column mass loading (Fig. 8b) in each run resembled its corresponding AOD (Fig. 8a), implying that the day-to-day variation in OC column mass loading in this dry season dominates the simulated AOD in the model, rather than other factors such as relative humidity (RH). The OC column mass loading is determined by the regional scale of emission, transport, and removal processes of aerosols; the latter two processes being the same across the six experiments, given that the same model configurations were used. Therefore, the differences of OC column mass loading and thus AOD across the six experiments are attributed to the different choices of biomass burning emission datasets. Figure 8c shows the local biomass burning OC emissions (i.e., at the $\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$ grid box where this site is located) in the different biomass burning emission datasets. We found that there was a large contrast in the local biomass burning OC emission between 25 September (as high as 1–2 µg m⁻² s⁻¹) and the other days (close to zero) across all six experiments although to different degrees. Similar emission patterns are found when averaged over nine or 25 surrounding grid boxes (not shown). Such a sharp contrast was completely absent in the simulated OC column mass density (Fig. 8b) and AOD (Fig. 8a).

All of the foregoing evidence collectively suggests that the temporal variations in AOD (and aerosol mass loading) in Alta Floresta during the burning season do not directly respond to the local BB emission at the daily and sub-daily timescales but rather to the regional emission. The regional emission is further adjusted by the multiple processes determining the residence time of aerosols in air (typically a few days), such as the regional-scale transport and removal of aerosols. The MODIS Terra true-color image overlaid with active fire detections (red dots) on 13 September 2008 (Fig. 8d) confirms that there were no active fires (represented by red dots) detected at Alta Floresta (blue circle), and thus the dense smoke over the site during this peak aerosol episode must have been transported from the upwind areas rather than from localized BB emission sources. Therefore, accurate estimation of both the magnitude and spatial pattern of regional emissions is quite important.

3.3.2 Mongu in Zambia (Southern Hemisphere Africa, SHAF)

The case of Mongu is different from that of Alta Floresta. Figure 9d reveals that there were numerous active fire detections (represented by the red dots in this MODIS Aqua true-color image) at and close to Mongu (blue circle) on 12 September 2008, one of peak aerosol episodes as Fig. 9a shows. The simulated AOD from all six experiments captured two peak aerosol episodes observed from the AERONET dataset during 11–12 and 2–3 September (AOD about 1.0), albeit underestimated (Fig. 9a). All experiments also underestimated the sustained aerosol episode after 20 September. However, all model experiments almost reproduced the AERONET AE value of 1.80 throughout September at this site (not shown), confirming that the dominance of the fine-mode aerosol particles in smoke aerosols is captured by the model irrespective of the BB emission dataset used.

Figure 9Characteristics of the observed and the simulated aerosols at Mongu during September 2008: (a) the 3-hourly time series of AOD at 550 nm. The AERONET AOD is represented by vertical gray bars, and the outputs from the six model experiments are represented by the color curves. The relevant statistics are listed: ave is the monthly average, ratio is the fraction of the simulated to the observed AOD at all observed hours, corr is correlation coefficient between the observed and the simulated AOD, and RMSE is root-mean-square error. (b) The 3-hourly time series of OC column mass density over the grid box where Mongu is located (units: $\mathrm{1}\times {\mathrm{10}}^{-\mathrm{6}}$ kg m⁻² or mg m⁻²). (c) Same as panel (b) but for biomass burning OC emission rate (units: $\mathrm{1}\times {\mathrm{10}}^{-\mathrm{9}}$ kg m⁻² s⁻¹ or µg m⁻² s⁻¹). (d) MODIS Aqua true-color image around Mongu on 12 September 2008, overlaid with the active fire detections as red dots (image credit: https://aeronet.gsfc.nasa.gov/cgi-bin/bamgomas_interactive and https://worldview.earthdata.nasa.gov).

The biomass burning OC emissions averaged over the grid box of Mongu exhibited distinct daily variations in each BB dataset (Fig. 9c). Similar emission patterns are found when averaged over nine or 25 surrounding grid boxes (not shown). At this site, the day-to-day variations in AOD still cannot be totally explained by the corresponding local emission at Mongu. For example, emission from FEER1.0 on 17 September is 6 times higher than that on 2 September (Fig. 9c), but the simulated AOD on 17 September is 2 times lower than that on 2 September (Fig. 9a). However, the magnitude of AOD at Mongu in each experiment corresponded to the magnitude of BB emission at the regional scale, since it is apparent that the overall higher regional BB emissions still resulted in higher column mass loading and thus AOD. For instance, FEER1.0 and QFED2.4, which have the largest monthly total biomass burning OC emission over the region of SHAF among the six BB emission datasets during September (2.27 and 2.92 Tg per month, respectively, as shown in Fig. 4), corresponded to the highest AOD (ratio =49 % and 46 %, respectively, as shown in Fig. 9a), while FINN1.5 and GFED4s, which represent the lowest monthly mean biomass burning OC emission over the region of SHAF (0.87 and 0.85 Tg per month, respectively, as shown in Fig.4), corresponded to very low AOD (15 % and 19 % of the observed AOD, respectively).

Although the temporal variation in the ambient RH may partially contribute to the day-to-day changes of the emission–AOD relationship, the close resemblance between the model-simulated AOD and column OC mass loading (Fig. 9b) excludes such possibility. This evidence therefore suggests that the temporal variations in AOD (and aerosol mass loading) in Mongu, where daily local BB emissions were present instead, do not directly respond to the local BB emission at the daily and sub-daily timescales during the burning season either as is the case in Alta Floresta. It further confirms the importance of accurate estimation of both the magnitude and spatial pattern of regional emissions as mentioned in the case of Alta Floresta. Therefore, over Southern Hemisphere Africa and Southern Hemisphere South America, an enhancement of regional BB aerosol emission amounts in all the BB emission datasets except for QFED2.4 in the latter region is suggested by this study in order to reproduce the observed AOD level although to different degrees.

4.1.1 Higher BB emissions estimated from QFED2.4 and FEER1.0

This study has shown that the QFED2.4 and FEER1.0 BB emission datasets are consistently higher than the others, with QFED2.4 being the highest overall. Some of the possible reasons responsible for this difference include the following.

• Constraining with MODIS AOD. The emission coefficients (C_e) used to derive biomass burning aerosol emissions in both QFED2.4 and FEER1.0 are constrained by the MODIS AOD, although in different ways (detailed in Sect. 2.1.6 and 2.1.5, respectively). This is not the case for the other BB emission datasets. Although GFAS1.2 uses the same FRP products as FEER1.0 in deriving dry mass combustion rate, its emission is tuned to that of GFED3.1. QFED2.4 applied four biome-dependent scaling factors to the initial constant value C₀ when deriving its C_e, by minimizing the discrepancy between the AOD simulated by the GEOS model and that from MODIS in corresponding biomes. The resulting QFED2.4 scaling factors are 1.8 for savanna and grassland fires, 2.5 for tropical forests, and 4.5 for extratropical forests (Darmenov and da Silva, 2015). This partially explains its very high OC biomass burning emission over the extratropical regions of TENA, BONA, and BOAS relative to the other emission datasets (Figs. 2–4). However, the high BB emission estimated by QFED2.4 is questionable during October and November of 2008 in the region of BONA (Fig. 4) according to the evaluation of its resulting AOD relative to the AERONET AOD at the Fort McMurray site (Fig. 7(1)). As for FEER1.0, the process of deriving C_e involved calculating the near-source smoke–aerosol column mass with the MODIS AOD (total minus the background) for individual plumes, thereby limiting influence from other emission sources (Ichoku and Ellison, 2014).

• Fuel consumption. In general, the FRP-based estimation approaches, such as GFAS1.2, QFED2.4, and FEER1.0, may enable more direct estimates of fuel consumption from energy released from fires, without being affected by the uncertainties associated with the estimates of fuel loads and combustion completeness (e.g., Kaufman et al., 1998; Wooster et al., 2003, 2005; Ichoku and Kaufman, 2005; Ichoku et al., 2008; Jordan et al., 2008). However, FRP from non-BB sources, such as the gas flare, could be mistakenly identified as BB sources. One example is over bare land on the eastern border of Algeria in MIDE (refer to the land type on the website http://maps.elie.ucl.ac.be/CCI/viewer/index.php, last access: 17 January 2020), where QFED2.4 shows high OC emission contrary to expectation (see Fig. 2f). Thus, additional screening of the FRP fire product is required.

4.1.2 Features of FINN1.5

Globally, the FINN1.5 dataset is lower than QFED2.4 and FEER1.0 but larger than GFAS1.2, GFED3.1, and GFED4s (Fig. 3). Although FINN1.5 can capture the location of the large wildfires using the active fire products, the estimation of burned area is rather simple without the complicated spatial and temporal variability in the amount of burned area per active fire detection or variability in fuel consumption within biomes. For example, it estimates 1 km² of burned area per fire pixel for all biomass types except for savanna and grassland where 0.75 km² per fire pixel is estimated instead. That might partially explain why FINN1.5 is extremely low in AUST, as suggested by Wiedinmyer et al. (2011). Additionally, the FINN1.5 dataset is the least over boreal regions, such as in the regions of BOAS and BONA, where FINN1.5 is only one-third and three-fifths of GFED4s, respectively. Large forest fires dominate in BOAS and BONA, such that the direct mapping of burned area as calculated in GFED4s and GFED3.1 produces more biomass burning emissions (van der Werf et al., 2017). On the other hand, the BB emission in the FINN1.5 dataset is relatively large near the Equator. For instance, it is the largest among the six datasets over the region EQAS and the second largest over the regions of CEAM and SEAS (see Fig. 3). This might be attributed to the smoothing of the fire detections in these tropical regions to compensate for the limited daily coverage by the MODIS instruments due to gaps between adjacent swaths and higher chances of cloud coverage in tropical regions (Wiedinmyer et al., 2011). Thus, in FINN1.5, each fire detected in the equatorial region is only counted for a 2 d period by assuming that fire continues into the next day but at half of its original size.

4.1.3 Difference between GFED4s and GFED3.1

Globally and in some regions, biomass burning OC emission in GFED4s is lower than that in GFED3.1 (see Figs. 2–4), although the former has 11 % higher global carbon emissions and includes small fires (van der Werf et al., 2017). There are a few possible reasons, of which two major ones are as follows: (1) for aerosols, the implementation of lower EF for certain biomes in GFED4s than in GFED3.1 reduces the aerosol biomass burning emissions. As for the savanna and grassland, for instance, the GFED4s dataset mainly applies the EF value recommended by Akagi et al. (2011), which is 2.62 g OC kg⁻¹ dry matter burned, 18 % lower than the EF from Andreae and Merlet (2001) used in GFED3.1, which is 3.21 g OC kg⁻¹ dry matter burned (see Table 2). The new estimation of EF is 3.0±1.5 g OC kg⁻¹ dry matter burned as suggested by Andreae (2019). With it, the OC emissions in savanna and grassland can be slightly enhanced but would still be lower than those in GFED3.1. (2) The improvement in including small fires in GFED4s over GFED3.1 is probably offset by the occasional optimization of fuel consumption using field observations for overall carbon emissions. For instance, the turnover rates of herbaceous leaf vegetation (e.g., savanna) are increased in GFED4s, leading to the lower fuel loading and thus lower consumption for this land-cover type in GFED4s (van Leeuwen et al., 2014; van der Werf et al., 2017). Therefore, the biomass burning OC emissions are lower in GFED4s over SHAF, NHAF, and AUST (Figs. 3 and 4), where ∼88 % of the BB carbon emission is from savanna and grassland (van der Werf et al., 2017).

Table 2Comparison of emission factor (units: g species kg⁻¹ dry matter burned) used by GFED3.1¹ and GFED4s² (listed in the upper and lower parts of the cell, respectively, and bold if GFED4s is larger).

¹ Mainly from Andreae and Merlet (2001) with annual updates. ² Mainly from Akagi et al. (2011), supplemented by Andreae and Merlet (2001) and other sources. ³ GFED4s (van der Werf et al., 2017) further divides extratropical forest in GFED3 (van der Werf et al., 2010) into temperate forest and boreal forest. ⁴ Based on Christian et al. (2003) for CO₂ and CO.

Download Print Version | Download XLSX

On the other hand, there are regions in the Northern Hemisphere where GFED4s is higher than GFED3.1. For example, over CEAS and EURO, small fires associated with burning of agricultural residues contribute to 43.6 % and 58.6 % of the carbon emissions, respectively (van der Werf et al., 2017). In spite of the 30 % reduction of the EF in these two regions, the effect of including small fires in GFED4s is greater, resulting in the biomass burning OC emission from GFED4s being twice as high as that from GFED3.1. Another example is in BOAS where the biomass burning OC emissions are 10 % higher in GFED4s than in GFED3.1. This is likely attributable to the higher EF used in GFED4s for boreal forest fires than that in GFED3.1 (9.60 vs. 9.14 g OC kg⁻¹ dry matter; see Table 2), where 86.5 % of the carbon emission is from the Siberian forest (van der Werf et al., 2017).

It is interesting that the yearly total biomass burning OC emission from GFED4s is 20 % lower than that from GFED3.1 in EQAS (Fig. 4), even though the small fires are included and the EFs of peatland and tropical forest are higher in the former (Table 2). By examining the monthly variations over EQAS (Fig. 4), however, we found that GFED4s is actually higher than GFED3.1 in August by a factor of 2 when peatland burning is predominant but is equal to or lower than GFED3.1 in other months, particularly in May, leading to the overall lower annual total value in GFED4s.