To represent fire activity we used the active fire location data set MCD14ML produced by the University of Maryland and provided by NASA Fire Information for Resource Management System . The data product is based on MODIS mid-range and thermal infrared observations. MODIS sensors are flown on board the sun-synchronous polar-orbiting Terra and Aqua satellites which respectively pass the equator at 10:30 and 13.30 local time during the daytime hours and at 22:30 and 01:30 at night. The instruments have a wide swath of approximately 2330 km, each providing near-global coverage daily. For high latitudes the coverage is better due to increasing overlap between consecutive overpasses. Each detection in the data set represents an active fire in a 1 km2 pixel at the time of satellite overpass, and contains information on the retrieved fire radiative power.

Large and long-lived fire events, likely strong emission sources, were identified and selected for the analysis. Burning episodes larger than 100 km2 are not numerous, but account for more than 80 % of total burned area in boreal North America , and are a dominant mode of burning in parts of temperate regions as well . In order to identify such events, individual MODIS active fire detections were agglomerated into large wildfire events by performing two-step spatial–temporal clustering. First, any MODIS fire detections located closer than 10 km in space and 24 h in time were grouped together. Single detections not assigned to any of the formed clusters were removed from further analysis. The clusters were then filtered by selecting events with (i) a spatial bounding box containing all fire detections belonging to the cluster larger than 100 km2 and (ii) a duration longer than 7 days. The duration was determined by the time span between the first and the last MODIS active fire detections belonging to the cluster. The burning was considered uninterrupted if the largest temporal interval between subsequent MODIS fire observations was less than 24 h. During the second step of clustering, any of the selected events active at the same time and located closer than 150 km were grouped into large burning episodes and assigned a unique source label. These events were classified into boreal and temperate fires using the dominant emission source given in the GFEDv4 inventory for areas and periods in which the events were active.

Smoke transport for the selected fire events was simulated with the HYSPLIT model . Plume dispersion from a source location was represented by the motion of a large number of discrete particles moved by the wind field with mean and random components. Global Data Assimilation System (GDAS) meteorological archive data were employed to drive the model.

For each day of burning, particles were continuously released into the model domain from the locations of the individual active fire detections within the fire event. In order to represent fire diurnal cycle, different MODIS active fire observations were used to release particles for two 12 h intervals representing day and night emissions from 09:00 to 21:00 and from 21:00 to 09:00 local time respectively. Emission source number and locations for daytime periods were determined from the highest number of fire detections observed during a single Terra or Aqua daytime overpass with 10.30 and 13.30 equatorial crossing time. Similarly, emitted particle source numbers for the night periods were determined by the largest burning extent observed during one of the night-time overpasses with 22.30 and 1.30 equatorial crossing times. Notably, the Terra overpass at 22.30 at high latitudes makes observations of regions where local time is earlier than 21:00. In this study, however, all fires detected during this overpass were classed as night-time observations. If no valid observations were available for some of the time intervals, the count and fire pixel locations were set to a minimum non-zero value estimated for the burning episode from all daytime or night-time observations. This was done to avoid total temporary shut-down of the emissions, which is an unlikely scenario for a long burning episode. Every hour, 20 particles were released for each fire pixel. As a result, emitted particle number for a burning episode was determined by the number of active fire pixels observed during a given time period.

Particles were uniformly distributed between the surface and the top altitude of the planetary boundary layer as given in GDAS archive. Satellite-based plume height estimates indicate that in up to 80 % of the events analysed, injection heights were limited to the planetary boundary layer. While confinement of the emissions to the mixing layer underestimates injection height for the most energetic burning episodes, such configuration should nonetheless represent the majority of burning episodes.

Throughout the simulations, modelled particle positions, their age and source burning event identifier were recorded each day at local solar noon. The generated point clouds were later used to compare against Terra and Aqua aerosol optical thickness (AOT) observations.

MODIS AOT collection 5.1 data products M*D04_L2 were used in this study. The dark target algorithm retrieves AOT at 550 nm and 10km×10 km spatial resolution at nadir. MODIS pixel size increases with view angle, and pixels at the edge of the swath are approximately 9 times larger. For this study, all M*D04_L2 AOT retrievals with quality assurance confidence > 0 were selected. To maximize coverage, no cloud fraction filtering was applied. The AOT product global validation against ground-based AERONET AOT observations suggest a 1σ error which increases linearly with aerosol loading ± (0.05 + 0.20 %) for overland cases. A regional MODIS M*D04_L2 AOT product validation indicates that performance varies greatly within North America. The study found that for 0.2<AOT<1.4 conditions, root mean square error varies from -0.01+0.51×AOT in arid western America where retrieval is hindered by bright surfaces, to 0.01+0.31×AOT in boreal forest and 0.3+0.12×AOT in the eastern USA. The study reported positive bias in MODIS AOT for some locations, in particular for retrievals at extremely high aerosol loadings. The AOT retrieval values have an upper limit of 5.0, and in addition, opaque smoke is often rejected as bright surface or cloud by the algorithm , preventing retrievals over extremely optically dense plumes. Consequently, AOT near the emission source is often not retrieved and the algorithm performs better when plumes are dispersed into regional haze.

Elevated MODIS AOT observations were attributed to a specific fire event and emission period by comparing above background MODIS AOT retrievals to plume extent modelled by HYSPLIT (Fig. ). Attribution required three pieces of information: (i) the event-specific background AOT value, (ii) the modelled plume extent at local solar noon for each day of burning and (iii) the coinciding MODIS AOT observations. First of all, the background AOT value was estimated for each of the selected burning events. It was determined by the median value of the AOT retrievals within 150 km radius from the fire event centroid observed 2 days prior to ignition. For each day of fire activity, a modelled plume extent (Fig. d–f) was determined from the locations of all HYSPLIT particle endpoints at solar noon, and AOT observations (Fig. a–c) with the highest spacial coverage for the day and plume area were selected from either the Terra or Aqua platform.

After the required information was obtained, the following steps were performed for each day of burning in an attempt to estimate fire-emitted AOT. First, plume regions bounding the particles released during the previous three daytime and night-time emission periods were identified. An estimation of the emission was attempted individually for each of the regions which represented plume areas that emitted during a specific time interval. This allowed the estimation of emitted AOT for up to 3 previous days from a single day of MODIS imagery. Importantly, such an approach allows the estimation of some emission periods even if a full MODIS plume overview is not available. Emitted AOT attribution was performed for the plume regions and satisfied two conditions: (i) the region had at least 80 % of MODIS AOT areal coverage, assuming that a single AOT pixel represents 100 km2 area, and (ii) within-region AOT median value was higher than the estimated background value for the fire event.

MODIS AOTs for the selected plume regions were interpolated to a 25 km-resolution equal-area grid (Fig. g–i) by employing radial basis function interpolation with a linear kernel. Fire-emitted AOTs were estimated by subtracting the background value from the within-plume AOT. The estimated fire-emitted AOT in every within-plume grid cell was apportioned to different emission periods and different sources based on information on release time and source of the HYSPLIT particles contained within the cell. If all particles found within a grid cell were released during the same emission period and originated from a single source, the cell's AOT was simply attributed to that emission period and source. If a mixture of particles were found within a cell, indicating that multiple fires and multiple emission periods contributed towards the grid cell AOT, the attribution was performed by apportioning a grid cell's fire-emitted AOT in proportion to the numbers of modelled particles released during the emission periods and with origin found within the grid cell. For example, if a grid cell had an AOT value of 1, and 100 HYSPLIT particles were located within the cell during the satellite overpass, 80 of which were emitted two diurnal cycles ago and 20 during the previous diurnal cycle, the grid cell AOT was split accordingly between the emission periods. Panels (k) and (l) in Fig.  illustrate the partitioning of total plume AOT into two different emission periods. Similarly, if there were any particles emitted from different fire events, grid cell AOT was divided both between different emission periods and different fire events.

The AErosol RObotic NETwork (AERONET) level 2 retrievals of aerosol microphysical and optical properties were used to characterize particles in plumes under investigation. AERONET consists of ground-based globally distributed sun–sky scanning photometers with a narrow field of view. The instruments are continuously monitored and calibrated, and the retrieved properties have estimated accuracy ranges. The direct sunbeam extinction measurements provide spectral AOT at several wavelengths ranging from 0.34 to 1.02 µm with uncertainties of 0.01–0.02 . Measured AOT and angular distribution of sky radiances are used to retrieve column-integrated aerosol volume size distribution at 22 size bins from 0.05 to 15 µm and spectral refractive index at 0.44, 0.67, 0.87 and 1.02 µm. Size retrieval is expected to be accurate within 25 % for particles with radii between 0.1 and 7 µm and within 25–100 % for size bins outside this range. Scans at high aerosol loadings (AOT 0.44 µm ≥0.4) allow the retrieval of refractive index with estimated uncertainties of 0.04 and 30 % for real and imaginary parts respectively .

Available observations within areas identified by the dispersion analysis as biomass burning plumes were attributed to a specific emission event and land cover type. Only retrievals containing refractive index (AOT 0.44 µm ≥0.4) were selected. In order to minimize the presence of dust and urban aerosol dominated retrievals, cases with volume concentration of fine mode (particle diameter < 1 µm) fraction less than 0.8, sphericity parameter lower than 0.98 and absorption Ångström exponent lower than 1 were filtered out. To make the samples more representative of plumes for which particulate matter was estimated, we selected AERONET observations within plume areas dominated by particles that had aged over 1 to 3 days.

The available AERONET spectral refractive indices were used to infer smoke aerosol water uptake. We employed the Maxwell Garnett effective medium approximation which provides a method to derive volume fractions of the components in the mixture if their refractive indices are known. The approach for retrieving black carbon concentrations from AERONET climatologies is described in detail and demonstrated by . It was further developed to infer brown carbon content , aerosol water uptake , and to simultaneously retrieve fractions of carbonaceous absorbers and dust .

To infer water content we employed a three-component mixture of black carbon and organic–inorganic matter included in a water host (Table ). For black carbon we assumed the refractive index and density suggested in . The second inclusion was used to represent a broad range of chemical species observed in biomass burning plumes , including organic carbon, ammonium sulfate and ammonium nitrate. These species were represented by a single component because they have n values close to 1.53. This value is characteristic of dry ammonium sulfate , was measured for organic carbon and lies within the range of values measured for dry organic compounds . Volume fractions of the inclusions and water host were retrieved in two steps. First, we deduced the amount of black carbon utilizing the spectral imaginary refractive index of the component. The Maxwell Garnett mixing rule was applied to a range of different fractions of black carbon in a water with negligible imaginary index. Volume fraction of the inclusion was estimated, determining the configuration which provided minimum χ2: χ2=∑i=1Nkiret-kimg2kiret2, where kiret is the AERONET-retrieved imaginary index, kimg is the value calculated by the Maxwell Garnett mixing rule, i is the summation over the selected AERONET wavelengths. We used AERONET k at 0.87 and 1.02 µm to retrieve black carbon fraction, assuming that it is the only absorber in this part of the spectrum. k at shorter wavelengths can be enhanced by absorption by organic carbon , which is retrieved as a part of the second inclusion. After volume fraction of black carbon was established, we kept it fixed and varied the fraction of the second inclusion in the mixture, minimizing the Eq. () for the real part of the refractive index at all four AERONET wavelengths.

Particle mass within the atmospheric column can be inferred from smoke AOT observations if mass extinction efficiency (Bext) is known: Mplume=τplumeBext, where Mplume is mass of plume aerosols, and τplume is a product of mean fire-emitted AOT and plume area. Bext represents extinction in area units per unit of aerosol mass, usually expressed as [m2 g-1]. It can be measured or calculated invoking Mie theory. In situ measurements of fresh North American smoke suggest Bext values ranging from 3.9 to 4.6 m2 g-1 . Equivalent measurements for aged plumes are not available for the region, but smoke samples collected in other forest ecosystems indicate slightly larger Bext values ranging from 4.0 to 5.3 m2 g-1 for older emissions. Similar Bext at 550 nm ranging from 4.5 to 5.2 m2 g-1 were inferred by from AERONET retrievals of dominant particle size distributions and index of refraction for North American boreal regions. applied a uniform 4.6 m2 g-1 value in deriving FEER TPM emission coefficients. Notably, plumes in their analysis were relatively young, up to a few hours old at most. In contrast, smoke discussed in this study is aged for a few days.

To avoid making assumptions about smoke optical properties, Bext was inferred utilizing available AERONET-retrieved refractive indices and particle size distributions. We used Mie code to calculate Bext assuming spherical internally mixed particles: Bext=∫rminrmaxσext(n,k,λ,r)dN(r)dln⁡rdln⁡rVdryρdry34π∫rminrmaxr3dN(r)dln⁡rdln⁡r, where σext is the extinction cross section of a single particle which depends on refractive indices (n,k), wavelength and particle radius (r). Vdry is particle dry volume fraction, ρdry is particle dry fraction density, both determined from aerosol water uptake analysis (Sect. ). σext was calculated at 0.55 µm using Mie code for every radius in the AERONET size distribution and averaged n and k retrievals at 0.44 and 0.67 µm. The numerator in the Eq. () is the single particle extinction cross sections integrated over number distribution, while denominator is aerosol dry fraction mass within the column given by the product of particle density and integrated particle volume.

Uncertainties in AERONET smoke aerosol properties, particle density and daily fire-emitted AOT attribution were propagated using a Monte Carlo method, retrieving water volume fraction, mass extinction efficiency and deriving total TPM estimates for the biomes. Throughout the study we report median values and interquartile range for the distributions, unless otherwise stated.

An important advantage of the AOT attribution method presented in this study is that it allows us to gauge combined errors originating from uncertainties in plume injection height, dispersion modelling, MODIS AOT retrievals and applied interpolation. Critically, any systematic changes in fire-emitted smoke optical thickness in evolving plumes can be inferred as well. This was facilitated by a number of cases in which two or more AOT attributions based on imagery and taken on consecutive days were performed for the same emission period. Figure  shows daily AOT estimates for days of emission for which the attribution was achieved from imagery taken on three consecutive days, for both night-time and daytime emission periods.

Overall, determined smoke AOT based on retrievals at later stages of plume development tend to have a positive bias compared to estimates for the same period of emission obtained on previous days. Notably, the largest increase in estimated AOT is observed when comparing estimates for the previous night-time emission cycle (smoke aged for 3 to 15 h) to AOT attributed to the same period determined from the following day's imagery, after the plume has aged for an additional 24 h. Inferred changes in daytime fire-emitted AOT over the first 2 days of ageing are smaller. Optical thickness for temperate smoke increases by approximately 30 % from the first observation of daytime emissions which are already aged for 15 to 27 h, compared to estimates for the same emission period determined from the imagery collected the following day. Changes in estimated daytime fire-emitted AOT for boreal plumes appear to be negligible. Notably, consecutive 24 h of ageing does not change estimated plume AOT significantly for both biomes and both daytime and night-time emissions. A slight decrease in optical thickness is observed for boreal smoke, but this should be treated with caution given the level of uncertainties involved. For the limited number of emission cycles presented in Fig. , contributions of day and night emissions appear to differ between the biomes. Night-time emissions constitute 30–40 % of total fire-emitted AOT for temperate events. Boreal plumes are dominated by daytime emissions with night-time emissions comprising under 20 % of total daily AOT. The difference is influenced by a larger number of night-time active fire pixels observed for temperate fires (Fig. f) and, consequently, more particles released during the night-time emission period in the dispersion simulations.

The effect of increasing AOT over time could be in part explained by uncertainty in plume dispersion modelling. However, the modelling error is expected to increase with time and hence should be manifested by progressively larger disagreement and biases for older estimates. In contrast, the results suggest that the agreement between the estimates for the same emission period is reasonably static across the plume age categories (Fig. b). The bias, on the other hand, is clearly largest for the first and the second plume observations within the first two diurnal cycles. It is possible that the model-emitted night-time particles get mixed with subsequent daytime emissions during the transport, effectively scavenging part of AOT from the other emission periods during the attribution. However, the observed daytime AOT tends to increase as well. Additionally, there are significant differences in inferred AOT changes between boreal and temperate plumes, indicating that some physical processes might be driving the change.

Particulate matter estimation and comparison with other methods are based on fire-emitted AOT during emission cycles starting and ending at 00.00 UTC. For 159 and 125 emission periods for boreal and temperate events respectively, AOT was determined from imagery taken on consecutive days allowing us to estimate the attribution error. These estimates do not include the problematic previous night emissions. Figure a shows the differences in fire-emitted AOT estimates for these cases. Given that the differences are approximately normally distributed, we propagated 50 % one sigma uncertainty in attributed daily fire-emitted AOT to derive confidence intervals for TPM emission estimates.

Large and persistent fire events discussed in this study exhibit distinctiveness in FRP values and diurnal burning cycle. Median MODIS FRP retrieved for the boreal fires is 103 (94–117) MW, while median FRP for temperate events is 90 (78–103) MW. This suggests higher burning intensity and combustion rates for boreal fires. A more striking difference, however, emerges when comparing ratios of maximum active fire pixel counts detected during individual daytime and night-time satellite overpasses. The proportion of active fires at night are typically much higher for temperate fires. The average daytime to night-time pixel count ratio is 1.4 (1.1–1.9) for the fires in this biome compared to median value of 3.6 (1.8–4.8) for boreal fires. Such a pattern indicates a higher contribution of night burning for temperate events.

The identified AERONET observations of boreal and temperate smoke suggest distinctiveness in retrieved size distributions and refractive index (Fig. a–c). The selected observations indicate that boreal emissions tend to have larger particles with median volume median radius value of 0.19 (0.17–0.21) compared to 0.17 (0.16–0.19) µm obtained for temperate smoke. These differences may be influenced by differences in combustion phase between the biomes. Very intense and predominantly flaming fires emit larger particles than events with more important smouldering combustion . Substantial differences exist when comparing the indexes of refraction for boreal and temperate smoke. Boreal plumes exhibit higher median n value of 1.49 (1.47–1.52) in contrast to 1.43 (1.37–1.45) observed for plumes attributed to temperate forest fires. Although boreal smoke generally is more absorbing with median k value 0.008 (0.007–0.01)i compared to the 0.005 (0.004–0.008)i value obtained for temperate emissions, plumes from both biomes are only weakly absorbing and characteristic k values have a negligible influence on calculated Bext. Variability in the real part of the refractive index between the plume categories, on the other hand, is larger and indicates differences in particle chemistry.

Maxwell Garnett medium approximation calculations using the discussed optical constants result in substantially different inferred water content for the two sources (Fig. ). The variability is mainly driven by the real part of the refractive index. Inferred median black carbon fractions are less than 1 % for both classes and thus have minimal impact on water content retrieval. Median water volume fraction for boreal fires is 0.15 (0.1–0.31), whereas temperate plumes have median value of 0.47 (0.29–0.67). The derived values agree with water volume fractions inferred by using a similar approach, although dust was not included as one of the components in our retrieval. Converting the inferred median water volume fractions to geometric hygroscopic growth factors results in values of 1.05 and 1.24 for boreal and temperate plumes respectively. These estimates compare favourably to measured factors for biomass burning smoke , indicating near-hydrophobic particles for boreal plumes, while temperate smoke could be classed as less hygroscopic. Notably, measured geometric hygroscopic growth factors are reported at 90 % relative humidity. In contrast, water volume fractions inferred in this study are representative of ambient humidity levels, and as a result direct comparison is not very meaningful.

The main limitations of the presented method are (i) the assumption that aerosols with n ≥ 1.53 are dry and (ii) large uncertainties in the chosen n values and different components used in the retrieval. In addition to increasing water content, formation of organic compounds may alter aerosol optical properties. Measured n for dry ambient organic aerosol are typically lower than the 1.53 value used in this study, ranging from 1.47 to 1.53 and appear to change with age . Although the uncertainties in AERONET properties and particle density were propagated in the retrieval, water fractions inferred in this study critically depend on n of the dry major component being close to 1.53. Any departures from this value result in inaccurate water uptake retrieval.

The differences in plume particle properties, primarily n and particle size, coupled with distinctiveness in inferred volume water fractions, drive differences in simulated Bext for the dry volume content of the plumes. Boreal plumes have larger particles, higher values of refractive index, but smaller water fractions and hence a lower median Bext value of 5.7 (5.1–6.5), while emissions originating from temperate forests have a median Bext value of 6.7 (5.4–9.2) m2 g-1 due to inferred greater water content. The identified AERONET observations are for ambient plumes which are aged for at least 1 to 3 days, and consequently, computed Bext values for dry volume fractions are larger than the 4.7 ± 0.7 m2 g-1 value suggested for dry aged boreal and temperate emissions . Somewhat higher values ranging from 4.7 to 5.5 m2 g-1 were calculated for a set of AERONET retrievals from North American boreal forest . The main difference between that aerosol climatology and the retrievals used in this study are in the real part of the refractive index. climatology for boreal smoke generally represents drier plumes with an average n value of 1.5 compared to 1.49 and 1.43 median n values attributed to boreal and temperate emission in this study.

The increase in attributed AOT in aged plumes determined in this study is consistent with well-documented smoke particle evolution. Aerosols grow considerably in size as plumes age. Particles undergo rapid changes during the first few hours after emission due to combined effects of condensation and coagulation , with reported growth rates in volume median radius as high as 0.04 µm per hour . On a timescale of days, plume particles continue to grow in dense plumes but at substantially lower rates, primarily due to coagulation and hygroscopic growth. Reported increases in volume median radius at these timescales are of the order of 0.02–0.03 µm . Condensation of organic and inorganic species and secondary particle production increase particle plume mass, while coagulation only transforms particle distribution. Both processes alter smoke optical thickness, mainly by enlarging scattering cross section and scattering efficiency, which is a strong function of particle size. Condensation has been reported to increase particle mass by up to 30–40 % in Amazonian plumes, but is thought to be important only during the first 24 h at most . The inferred increase in fire-emitted AOT over the first 2 days of ageing reported in this current study only partially overlaps with this period. The first few hours of plume development when condensation is thought to be the most active are not represented; therefore condensation is unlikely to contribute significantly towards the inferred AOT growth. A growth in volume median radius of 0.02 µm due to coagulation theoretically could increase scattering efficiency by up to 30 % without changes in plume mass, but this process cannot explain the differences in the magnitude of AOT change observed between the biomes.

An additional factor driving changes in AOT is water uptake by smoke particles. Absorption of water depends on air relative humidity and aerosol solubility which in turn tends to increase with atmospheric processing. It increases particle size further, enhancing scattering cross section. Hygroscopic growth factors measured and inferred by optical methods for biomass burning smoke at 80 % relative humidity range from 1.1 to more than 2 . suggested an average enhancement factor of 1.35 ± 0.2. Bext values derived for dry volume fraction in this study suggest median scattering cross-section enhancement factors of 1.2 and 2 for boreal and temperate plumes, assuming the 4.7 Bext value for dry smoke .

Notably, the magnitude of AOT increase over time, shown in Fig. , corresponds to inferred median water fractions for the two biomes. Temperate emissions exhibit generally hydrophilic particles with much greater water content, while boreal plumes seem to contain much less aerosol water. This distinctiveness could be due to different ratios of smouldering and flaming combustion. Field measurements indicate that prescribed burns, and in particular wildfires in temperate regions, have lower combustion efficiencies . Temperate fires discussed in this study have lower mean FRP values and a less pronounced diurnal burning cycle, and the emitted plumes have higher ratios of night-time emissions. Smouldering night-time smoke has been reported to contain more soluble organic compounds , which could explain the presence of more hydrophilic aerosols in temperate plumes. In addition, factors not accounted for in this study, such as significant differences in relative humidity and atmospheric processing between the biomes, may be partly responsible for the inferred variability in water uptake.

On an individual event basis the relationships between daily particulate emissions given by the global inventories and this study exhibit varying degrees of agreement. Figure  shows the TPM from this study and GFED for the events for which estimation was performed for at least seven diurnal cycles. Although some fires exhibit only fair or weak agreement, the result is nonetheless encouraging considering the error in AOT attribution and conversion to TPM method in this study, and large uncertainty associated with the date of burn in daily burned area product on which GFED depends. Robust linear fits between GFED TPM and daily estimated TPM, shown in Fig. , indicate considerable variability in slopes, even in comparison to the events with generally good agreement. This suggests distinctive combustion and emission characteristics for individual events. As well as variability on a per burning event basis, large differences exist when comparing relationships for fires in boreal and temperate forests. Notably, for every tonne of GFED TPM, this study shows TPM ranges from 0.46 to over 2 t for boreal burning events, while for temperate fires the conversion factors range from approximately 1 to more than 5. The relationships are similar in terms of agreement when comparing daily TPM estimates with other inventories (not shown), but scaling factors, which are needed to reconcile the estimates, differ.

Total TPM emission estimates obtained in this study for the wildfires examined are large in comparison to FEER and, to a lesser degree, to GFED and GFAS inventories, but are smaller than QFED estimates (Fig. ). QFED emissions are reported for PM2.5 aerosol fraction only, which typically constitutes 70 to 85 % of TPM for the biomes discussed . As a result, QFED TPM estimates should be approximately 20–40 % higher than indicated in Fig. .

Substantial differences exist between the estimates for boreal and temperate fires. For boreal forest events, total TPM emissions for this study are in close agreement with the bottom-up GFED and GFAS TPM estimates. The agreement indicates that application of the proposed 2.2 enhancement factor to GFAS TPM would overestimate boreal emissions for the events discussed. In fact, assuming an increase in aerosol mass and AOT in ageing plumes, boreal TPM emissions for this study are low compared the near-source GFED and GFAS estimates. Regional AOT-based QFED inventory suggests PM2.5 emissions are higher by 40 %, while near-source FEER TPM estimates are smaller by a factor of 2.8 when compared to TPM for this study.

For temperate forests, a striking contrast exists between GFED and GFAS inventories and methods based on regional AOTs. The largest estimates are given by the QFED inventory, which suggests PM2.5 emissions are higher by 50 % than the TPM estimates for this study. If bottom-up estimates of the boreal emissions agree well with this study's TPM for temperate events the discrepancies are much larger. Scaling factors of 3.2 and 4 are needed to reconcile GFED and GFAS emissions with the estimates obtained in this study. FEER emissions are closer to bottom-up approaches suggesting much lower emitted TPM compared to the other top-down methods. This appears to be characteristic to North America as has been reported in , indicating potential underestimation of the emissions in the region. For other continents, FEER generally predict higher TPM emissions than the bottom-up inventories and agree closely or even exceed QFED PM2.5 estimates.

The above emission budgets suggest particulate matter emission coefficients of 27 (23–30) and 31 (24–37) g per MJ-1 of time integrated GFASv1.0 FRP (Table ). They comprise approximately 70 % of coefficients derived for QFED PM2.5 emissions, and are 2.5 times larger than equivalent values derived using FEER emission coefficients. Notably, although differing in magnitude, all three top-down methods indicate slightly larger emission coefficients for temperate events. In contrast, more bottom-up approaches suggest 2.5 to 3 times larger emission coefficients for boreal forest. The TPM emission factors employed in GFAS and GFED inventories are identical for both forest types, but large differences exist in consumed biomass estimates. The GFAS inventory employs a 3-fold larger FRP to dry matter combustion rate conversion factor for boreal fires, attributable to high organic soil content in the biome. In contrast, emission coefficients for boreal and temperate forests derived in this study are statistically indistinguishable.

A number of factors may contribute towards the discrepancies between TPM estimates for this study and other methods. Relatively large estimates compared to near-source GFED, GFAS and FEER inventories may be influenced by unaccounted processes in ageing plumes. A several-fold growth in plume mass due to condensation and secondary particle production, however, seems implausible given that the reported magnitude of increase in particle mass driven by these processes is within 50 % . The difference may be partly due to large sizes of the events sampled in this study. Field measurements for large wildfires are scarce , and such events are underrepresented in compiled EFs. The agreement between top-down and bottom-up methods is better for boreal fires than it is for temperate events. Fire event sizes are similar for both biomes, at least for the fires sampled. Therefore, it seems that fire size considerations alone fail to explain the varying degree of agreement between this study's TPM, and GFED and GFAS estimates when comparing boreal and temperate cases. Comparably low FEER estimates, apart from ageing effects, might be partly determined by sampled event size. Infrequent and large fires prevailing in North American forests make it difficult to reliably derive combustion coefficients from near the source imagery .

Considering the above factors it seems that for the large fire events discussed, boreal emissions are underestimated by a factor close to 2 by the FEER inventory. Temperate TPM appears to be underestimated by factors of 2 to 4 by FEER, GFED and GFAS. QFED on the other hand, seems to overestimate particulate emissions by 40 to 50 %. The previously suggested GFAS TPM 2.2 enhancement factor seems to represent an average value for the region. It is not required for boreal fires, and is close to 4 for temperate plumes. While the underestimation by bottom-up GFED and GFAS may be driven by low-emission factors, the magnitude of the difference indicates that biomass-consumed estimates are the most likely source for discrepancies.