Introduction
Open biomass burning is known to be an important source of black carbon (BC),
which is the major absorbing component of carbonaceous aerosol and one of the
main atmospheric species contributing to climate forcing (Bond et al., 2013;
IPCC, 2013). On the global scale, the radiative forcing of BC, including the
effects of BC on ice and snow surfaces, has been estimated to be as high as
+1.1 Wm-2 . In a more recent,
observationally constrained analysis, the BC radiative forcing after
subtracting the preindustrial background was estimated to be
+0.53 Wm-2 (with the uncertainty bounds of +0.14 to
+1.19 Wm-2) , which still suggests that it is
quite significant in comparison to the radiative forcing of 1.82±0.18 Wm-2 associated with carbon dioxide
(which is the main climate forcer). Open biomass burning (BB) is likely to
contribute about 40 % to the total BC emissions .
As a significant BC source, BB plays an especially important role in climate
processes in the Arctic, where the increase of the annual surface temperature in
the period since 1875 has been almost twice as large as that in the rest of the
Northern Hemisphere . Several studies
e.g., have indicated that a
significant part (up to about 50 %) of this temperature increase could have
been induced by BC. There is an abundant amount of evidence that BB provides a significant
contribution to BC in the Arctic atmosphere in the spring and summer
e.g.,.
It has also been estimated that Siberian fires alone
contributed almost half (46 %) of the total BC amount deposited in the Arctic
over a period of 12 years (2002–2013). Radiative effects associated with BC
residing in the Arctic atmosphere include both direct radiation budget
changes causing strong warming of the Arctic surface and significant changes
in atmospheric stability and cloud cover .
Significant increases in surface temperature in the Arctic as a result of
perturbations of the meridional transport can even be caused by BC residing
in the midlatitude atmosphere . This indicates that to
correctly evaluate the effects of BC on the Arctic climate it is critical to
know its concentration not only in the Arctic but also in the atmosphere over
adjacent regions such as Siberia. Additionally, the deposition of BC on ice and
snow has been found to strongly contribute to Arctic warming by decreasing
surface albedo and promoting ice/snow melting which, in turn, may result in
further surface darkening and provide a positive feedback on the increase of
the surface temperature in the Arctic
.
The effects of biomass burning on atmospheric composition and climate are
commonly evaluated using chemistry transport and climate models relying on
data from BB emission inventories, such as the Global Fire
Emissions Database (GFED) , the Global Fire Assimilation
System (GFAS) emission dataset , the Emissions for
Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP)
inventory , the Fire Inventory from NCAR (FINN)
, and the Quick Fire Emissions Dataset (QFED)
, which are widely used in atmospheric and climate
studies. However, emission inventory data are likely to be affected by
considerable uncertainties due to a limited knowledge of spatiotemporal
characteristics and temperature regimes of fires, as well as due to the lack
of reliable estimates of emission factors for a variety of ecosystems and
environmental conditions. These uncertainties lead to large discrepancies
between emission estimates provided by different inventories. For example,
according to the FEI-NE inventory recently developed by , the
annual BC emissions from fires in northern Eurasia in the period from 2002 to 2015
are, on average, a factor of 3.2 larger than those given by the GFED4
inventory. Using the FEI-NE inventory in the FLEXPART
Lagrangian particle dispersion model, found the model
results to be in a reasonable agreement with surface BC concentrations
observed at several Arctic stations in the period from 2002 to 2013. Conversely,
a Bayesian inverse modeling analysis based on carbon isotope
characterization of BC measurements at Tiksi (East Siberian Arctic) from
April 2012 to April 2014 revealed that the best fit of the FLEXPART data to
the observations was achieved by reducing the GFED4 fire emissions by 53 %
; however, this estimate may reflect uncertainties in the
spatial distribution of the GFED4 emissions, as the sensitivity footprints in
this particular study only cover a part of Siberia . In
view of these rather controversial findings and the important role that BC
emissions from BB are likely to play in climate processes in the Arctic, it
is critical to obtain stronger observational constraints on BC emissions from
fires in northern Eurasia and its major BB BC source regions such as Siberia.
Note that the general term “black carbon”, which is used throughout this
paper, is rather generic and can be broken down into more specific terms,
including refractory black carbon (rBC), elemental carbon (EC), and
equivalent black carbon (eBC); these terms refer to three major measurement
approaches that are used to characterize carbonaceous matter, such as
laser-induced incandescence, thermal or thermal–optical methods
distinguishing between more and less volatile fractions of carbonaceous
aerosol material, and optical methods based on measurements of
light absorption coefficients . Accordingly, BC
emission data reported by a given emission inventory may be based on one or
more specific methods that were employed to evaluate the emission factors
used in the inventory. However, a concrete “measure” of BC is usually not
specified in BB emission inventories.
The main goal of this study is to investigate the feasibility of constraining
the BB BC emissions using retrievals of the aerosol absorption optical depth
(AAOD) from satellite measurements performed by the Ozone Monitoring
Instrument (OMI) in the near-UV region . To
achieve this goal, we address the case of the severe fires that occurred in
Siberia in 2012 (see, e.g., ). The other
goals of this study are to obtain “top-down” estimates of the monthly BC
emissions from fires in Siberia in the period from May to September 2012 and to evaluate the
corresponding data of the GFED4 and FEI-NE inventories for this period.
Previous applications of the OMI AAOD retrievals include, in particular,
the evaluation of BC emissions employed in global aerosol models
and the identification of the atmospheric variability of
AAOD at various scales .
used AAOD retrieved from OMI observations in an inverse
modeling analysis involving the GEOS-Chem (Goddard Earth Observing
System-Chemistry) global model to constrain BC emissions over southeastern
Asia (where the BC emissions are predominantly anthropogenic) for April and
October 2006; they found overwhelming enhancements (up to 500 %) in
anthropogenic BC emissions in April relative to a priori emission estimates.
In this study, the OMI AAOD measurements are analyzed utilizing simulations
performed for the northern Eurasian region (including Siberia) using the
CHIMERE chemistry transport model .
One of main difficulties with using the OMI AAOD retrievals for constraining BC
emissions stems from the well-established fact (see, e.g.,
) that the absorption of UV
and shortwave visible radiation by BB aerosol is strongly affected by brown
carbon (that is, by the light-absorbing fraction of organic carbon). In view
of this fact, explicit modeling of AAOD in the case of BB aerosol as a
function of its composition would inevitably involve major uncertainties
associated with the assumptions regarding the magnitude of the imaginary part
of the refractive index for organic carbon (OC) and the mixing state of
aerosol particles; these characteristics are likely strongly variable,
depending on the sources and atmospheric processing of BB aerosol
. To overcome this difficulty,
we follow an empirical approach that involves the
parameterization of AAOD as a function of the EC/OC (elemental carbon to
organic carbon) ratio and the aerosol extinction optical depth (AOD). This
parameterization is based on the analysis of experimental relationships
between the single scattering albedo (SSA) of BB aerosol particles and the
EC/OC ratio and is fitted to the retrievals of aerosol optical properties
from the multi-wavelength measurements made at the AErosol RObotic NETwork
(AERONET) sites in Siberia in summer 2012. The relationships between SSA and
the EC/OC ratio were reported by as a result of the
fourth Fire Lab at Missoula Experiment (FLAME-4).
Along with the OMI AAOD retrievals, we use AOD retrievals from satellite
measurements made by the Moderate Resolution Imaging Spectroradiometer
(MODIS). Numerous studies found the MODIS AOD retrievals to provide useful
observational information for the evaluation and estimation of BB emissions of
aerosol and co-emitted species
e.g.,.
In this study, the MODIS AOD data were used to constrain OC emissions and to
optimize the calculated AOD values. Note that since BC is usually a minor
component of BB aerosol, AOD is mostly determined by the organic (scattering)
fraction of BB aerosol ; thus, estimation of BC emissions
using only AOD measurements would require making some assumptions regarding
the quantitative aerosol composition. The AAOD measurements are much more
sensitive to the BC fraction of aerosol than the AOD measurements, even
though the OMI AAOD retrievals are also sensitive to OC due to its strong
absorption at shorter wavelengths.
The optimized emissions are validated using independent ground-based and
aircraft aerosol measurements performed in different parts of Siberia.
Therefore, through the use of a chemistry transport model, this study
integrates data from satellite and ground-based remote sensing and in situ
and aircraft measurements to not only obtain independent observation-based
estimates of BC emissions from Siberian fires but also to ensure their
reliability. Moreover, the use of a chemistry transport model allows
for the transport path of BB BC emissions in the atmosphere to be followed, and
for the part exported directly to the Arctic to be determined.
Schematic representation of the study design. Green is used to
depict the observational data used in the analysis. Red and dotted red
lines illustrate the iterative procedure aimed at the optimization of the BB BC
and OC emissions.
A general overview of the study design is presented in Fig. . The
methodology of the study is described in detail in Sect. . The
results of our analysis are presented in Sect. . Finally,
Sect. summarizes our findings followed by the concluding
remarks.
Results
Optimal estimates of the correction factors for BB emissions of BC and OC
The optimal estimates of the correction factors, FBC
and FOC, and their uncertainties for each of the 5 months
considered are reported in Table ; note that the subscript “m”
is omitted here and in the following for brevity. The estimates range from about 2.3 (in
May) to 3.7 (in September) for BC and from 1.5 (in May) to 2.7 (in August)
for OC. Table also lists our estimates of the FBC/FOC
ratios, which range from about 0.7–0.8 (in June–August) to 1.5–1.7 (in May
and September). The estimates of both FBC and FOC as well as of
their ratios are reasonably well constrained by the observations: the
respective uncertainties are less than or about 35 % for FBC and less
than 30 % for FOC in the summer months. The uncertainties are largest
in the estimates of FBC and FOC for September (45 % and 47 %,
respectively). This is not surprising considering that the fires were
relatively small in this month (see Fig. ), and so the number of
available observation data points in the optimization dataset is many
times lower for September (58 data points) than, e.g., for July (3017 data points).
The monthly variations in both the correction factors and their ratios
exhibit a rather pronounced seasonal pattern. Specifically, the values of
FOC are smaller in May and September than in the summer months. In
contrast, the values of FBC and the FBC/FOC ratio are much
smaller in the summer months than in May and September. Although the
differences between the correction factors for different months are mostly
not statistically significant, a more than twofold decrease in the
FBC/FOC ratio in June and July is statistically significant with
respect to both May and September.
While the monthly variations of FBC and FOC may, in principle,
account for changes in both the emission factors and in the conversion factor
α (see Eq. ), the variations in the FBC/FOC
ratio may be explained by changes in only the ratio of the BC and OC emission
factors. It seems possible that the variations of the FBC/FOC
ratio are partly associated with high variability of the contributions of
different fire types (featuring different emission factors) to the observed
FRP (see Fig. ): specifically, the contributions of agricultural
and grass fires to the integral FRP were 41 % and 21 % in May and September,
respectively, but only 14 %, 10 %, and 9 % in June, July and August. To examine
this possibility, we performed an additional estimation (see the
Supplement, Sect. S3) using the AAOD and AOD observations only
over a selected subregion (50–57∘ N, 60–115∘ E, see
Fig. ) where the relative contribution of agricultural and grass
fires to FRP was much larger and more uniformly distributed across the
different months than in the whole region (see Fig. S4). The estimates of the
FBC/FOC ratio obtained for this subregion (see Table S1 in the Supplement) show
much smaller (and statistically insignificant) variations between different
months, as compared to the corresponding estimates for the whole region (in
particular, the FBC/FOC ratio decreased in May but increased
in the summer months), whereas monthly variations of the FBC and
FOC factors themselves (see also Table S1) even increased.
Therefore, this additional analysis supports the possibility that the monthly
variations of the FBC/FOC ratio are associated with different
fire types; therefore, this would infer that the emission factors (for BC
or OC or the both species) specified in the GFED4 inventory and in our
simulation are biased in case of at least one fire type. However, it should
be noted that the variability of the FBC/FOC ratio for the whole
study region can also be explained by other reasons, such as spatial
variability of the emission factors across different ecosystems in the region
considered, as well as by the emission factor monthly variability which is
not represented by the constant emission factor values specified in the GFED4
inventory (see Table ). Based on the limited amount of
available data, we can not exclude these alternative explanations.
Evaluation of the optimized simulations of AAOD and AOD
In this section we examine whether the simulations that were employed in
the inverse modeling analysis are sufficiently reasonable and representative
of the observations that have not been used for the optimization of the BB
emission parameters. To this end, we compare our simulations, in which the BB
emissions have been computed using the correction factors presented above,
with a validation subset of the satellite data (see Sect. ).
A comparison of our simulations with in situ measurements is presented in the
subsequent two sections.
Spatial distributions of the mean values of AAOD at 388 nm (a, c, e)
and AOD at 550 nm (b, d, f) in the period from 1 May to
30 September 2012 according to (a, b) the OMI and MODIS observations,
respectively, and simulations performed with the optimized BB emissions (c, d) and without BB emissions (e, f).
Figure presents the spatial distributions of the temporally
averaged AAOD and AOD values over the study region according to the satellite
observations and our simulations performed both with the optimized BB
emission and with zero BB emissions. Note that blank pixels indicate that
either the satellite observations are available for less than 2 days in
these grid cells, or that the observed and/or simulated data have not been
included in the validation subset according to the criteria formulated in
Sect. . Evidently, both the observed and simulated (with the
BB emissions) data show rather similar spatial patterns, indicating the
presence of heavy smoke plumes over many areas in both western Siberia (in
particular, between Omsk and Krasnoyarsk) and eastern Siberia (southeast of
Yakutsk). Importantly, the effects of the same fires can be readily seen in
both the AAOD and AOD data. The differences between the satellite data and
simulations are also considerable (the root mean square errors normalized to
the mean values equal 0.49 and 0.46 in the cases of the AAOD and AOD
distributions, respectively). In particular, both AAOD and AOD tend to be
overestimated by the model in central Siberia and the “Far East” but
underestimated in western Siberia. These differences may be due to a variety
of reasons, including errors in the spatial allocation and the magnitude of
fire emissions, uncertainties in the satellite retrievals, as well as the
model's inability to take spatial and temporal variations in the
optical properties of the actual BB aerosol into account.
Time series of daily AAOD (a) and AOD (b) values averaged over the
study region according to the OMI AAOD and MODIS AOD observations and
simulations (“base-opt” and “base-ini”) performed with the optimized and
initial-guess BB BC and OC emissions, as well as (“bgr”) without fire
emissions. Note that whenever a sufficient number of the observational data
points was available (see Sect. ), the simulations were
averaged over the same grid cells as the observations; otherwise, the
simulated AAOD and AOD values were averaged over the whole study region.
The numbers given in the figure legends report the mean AAOD and AOD values
obtained by averaging over the observational data points and their simulated
matchups shown in the figure as well as the values of the correlation
coefficient.
Figure presents the temporal (daily) variations in the
spatially-averaged AAOD and AOD data. Considering that the number of
spatially resolved data points averaged over a given day strongly varies
from day to day and that the agreement between the daily mean data from the
simulations and observations is likely to degrade on days with a small amount
of available data, we required that each observational data point (and its
simulated matchup) shown in Fig. was composed of at least 10
values corresponding to different grid cells. Otherwise, an observational
data point was considered to be an outlier. These outliers were not included in
Fig. and were disregarded in the comparison statistics (reported in
the legend of Fig. ); the corresponding simulated values (shown in
Fig. ) were averaged over the whole study region. The results
presented in Fig. indicate that when the model used the optimized
emissions, it reproduced the daily variations both in AAOD and AOD reasonably
well, with correlation coefficient values of about 0.8 and very small biases
that do not exceed 5 %. The agreement of the simulations is evidently better
with the AOD observations than with the AAOD retrievals. This is an expected
result, given the fact that both the OMI-derived AAOD data and the
corresponding simulations are likely to have larger uncertainties than the
observations and simulations of AOD. The correlation coefficient values were
considerably smaller and biases were much larger when the model used the
“initial-guess” BB emissions calculated with the correction factors equal
unity. These findings indicate that the inversion of the AAOD and AOD
observations results in major improvements in the model performance.
The relationship between AAOD and AOD values according to (a) the
satellite and AERONET data and (b) corresponding simulated data from the
“base-opt” model run. In the case of the satellite data, each data point
represents a value of AAOD or AOD for a given cell of the model grid. The
AERONET data are described in Sect. ; the corresponding
modeled data were extracted for grid cells and days matching the AERONET
observations. The figure legends report the equations of a linear regression
without an intercept, the mean values of the AAOD and AOD for the different
datasets, and the values of the correlation coefficient for each set.
Figure compares the relationships between AAOD and AOD according
to the satellite observations and our simulations. The relationships include
all of the gridded daily data points selected for the validation dataset. As
follows from Eq. (), the relationship between AAOD and AOD is
indicative of the EC/OC ratio in BB aerosol particles. Therefore, the
adequacy of the relationship between the modeled AAOD and AOD values is an
important prerequisite for accurate estimations of the EC/OC ratio in the BB
aerosol emissions. Figure 7 also shows the similar relationships between the
AAOD and AOD data derived from the AERONET measurements, which were used to
evaluate the parameters of Eq. (), and between their modeled
counterparts.
Evidently, although the model can not explain some strong variations in the
AAOD/AOD ratios derived from the observations (which may be enhanced due to
temporal and spatial inconsistencies between the OMI and MODIS measurements),
it reproduces the “observed” relationship quite well on average.
Specifically, both the observations and simulations indicate that the ratios
of the average values (indicated by angle brackets) of AAOD and AOD, as well
as the slopes of the regression lines fitted to the AAOD and AOD values, are
close to 0.1 (±11 %). According to Eq. (), this value
corresponds to an EC/OC ratio of about 0.045, which is rather similar to that
of 0.052 assumed in the GFED4 inventory for BB emissions in extratropical
forests. Similar values of the 〈AAOD〉/〈AOD〉 ratio are characteristic
for the AERONET data and their simulated matchups, although the latter is
slightly positively biased. The consistency between the 〈AAOD〉/〈AOD〉
ratios in the satellite observations and AERONET data can be considered as
evidence that the measurements of the optical properties of BB aerosol at the
AERONET sites are sufficiently representative of the typical optical
properties of BB aerosol in the whole study region. Note that the cluster of
green points above the regression line in Fig. b indicates a
distinct contribution of the agricultural and grass fires featuring much
larger ratios of the BC and OC emission factors (see Table ) than
the predominant forest fires.
Time series of the EC and OC mass concentrations (µgm-3)
measured at ZOTTO in comparison to their simulated matchups
from the “base-opt”, “base-ini”, and “bgr” simulations. The values of several
statistical characteristics are reported in the figure legends; the
confidence intervals for the biases are evaluated in terms of the 90th
percentile. The EC and OC concentrations from the “bgr” simulation are
plotted using the corresponding axes on the right-hand side of the panels.
Evaluation of the simulated BC and OC concentrations against observations at ZOTTO
Figure illustrates the evaluation of the simulated concentrations
of EC and OC against the corresponding in situ observations at the ZOTTO
site, which in summer 2012 was surrounded by numerous fires (see
Fig. ). The observational data points shown in Fig.
represent EC or OC concentrations detected in the individual aerosol filter
samples. The simulated data from the CHIMERE model, which was run with both
the optimized and initial-guess BB emissions as well as without BB emissions,
were averaged over each individual sampling period; the sampling periods were different
lengths for different samples (see Sect. ). The comparison
statistics, including the mean value, the difference between the mean values
of the simulated and observed data (the bias) along with the 90 % confidence
interval, and the correlation coefficient are reported for each simulation in
the legends of Fig. . Note that to the best of our knowledge,
aerosol simulations performed with a chemistry transport model have never
been previously evaluated against EC and OC measurements in Siberia.
Both the EC and OC concentrations predicted by the model for the “base-opt”
and “base-ini” cases correlate very well (r>0.9) with the corresponding
observations. The EC concentrations are somewhat overestimated on average in
the simulations with the optimized emissions: the agreement of the mean
concentrations would be perfect if the simulated concentrations were reduced
by 23 %. However, a predominant part of this difference between the mean
simulated and measured concentrations can be explained by random model
errors. A remaining smaller part of the difference may be explained by the
uncertainty in our estimates of the emission correction factors (and thus in
BC emissions specified in the model), which is about 35 % (see
Sect. ). Conversely, the EC concentrations in the simulation with the
initial-guess emissions are a factor of 1.32 too low on
average. The fact that the “base-ini” simulation demonstrates a slightly
better performance in terms of the correlation coefficient than the
“base-opt” simulation may be indicative of a smaller monthly variability of
the BC emission and/or conversion factors representative of the forest fires,
which predominate in the vicinity of the ZOTTO site, compared to the
variability of the same parameters representative of the fires across the
whole study region.
The OC concentrations simulated with the optimized emissions appear to be
slightly biased low, but the available bias estimate is not statistically
significant. In contrast, the OC concentrations are strongly (by more than a
factor of 2) underestimated in the “base-ini” simulation: this result is
consistent with a similar underestimation of AOD in the same simulation (see
Fig. ) and further supports our finding (see Table )
that the initial-guess OC emissions should be strongly increased. Note that
according to our “bgr” simulation, i.e., if BB emissions in Siberia were
completely absent, both EC and OC concentrations at the ZOTTO site would be
more than an order of magnitude lower than observed. This fact indicates
that possible uncertainties in anthropogenic EC emissions are not likely to
be responsible for any noticeable bias in the EC concentrations simulated
with BB emissions. Overall, the above comparison indicates that our top-down
BB EC and OC emission estimates, considered in combination with their confidence
intervals, are consistent with the EC and OC observations made in central
Siberia.
CO mixing ratios (ppb) derived from the measurements that
were made in the framework of the YAK-AEROSIB experiment on 31 July and 1 August 2012 (a) in
comparison with the corresponding simulated values from the
“base-opt” CHIMERE run (b). The mixing ratios are overlaid on the CHIMERE
grid; the mixing ratio values were calculated by averaging the original measurement
data and their simulated matchups over the region covered by each grid cell
that was intersected by the aircraft trajectory.
Comparison of the simulated data with aircraft measurements
Figure shows the tracks of the flights performed
in the framework of the YAK-AEROSIB campaign from July to August 2012. The
northern and southern sectors of the trajectory correspond to the flights
performed on 31 July and 1 August, respectively. The flight tracks are
overlaid onto the grid of our model and are shown along with the observed and
simulated values of the CO mixing ratio, which were averaged over the region
covered by each grid cell that had been intersected by the aircraft
trajectory. One can notice several grid cells (north and south of Krasnoyarsk
and around Yakutsk) where the CO mixing ratios (both in the measurements and
in the simulations) exceed 400 ppb. These “hot spots”, corresponding
to high percentiles of the CO mixing ratio, were not found in the respective
data from the “bgr” simulation (which are not shown in Fig. ) and
thus are likely due to BB emissions. Note that crossing BB smoke coinciding
with high CO plumes has been confirmed by direct visual/olfactory evidence as
well as by a clear increase in the K+ ion concentration in the forest fire
plumes (Antokhin et al., 2018). Taking these considerations into account, we
used high percentiles of the CO observations to pinpoint occurrences when the
aircraft traversed BB plumes.
Specifically, we selected PM2.5 and BC (eBC) measurements matching the
CO mixing ratios exceeding the 90th percentile (395 ppb) or
80th percentile (277 ppb) of the distribution of the CO mixing
ratios. The average CO mixing ratios in the selected subsets of the
measurement and simulated data were 602 and 374 ppb
for the 90th percentile and 465 and 311 ppb
for the 80th percentile, respectively. As the selection criterion was only applied to
the observational data that manifest strong subgrid variability, the fact
that the average CO mixing ratios are larger in the observations than in the
simulations does not necessarily mean that the model underestimates the CO
mixing ratios in the BB plumes. More importantly, the corresponding average
CO mixing ratios simulated without BB emissions (110 and 111 ppb for
the selection criteria based on 90th and 80th percentiles,
respectively) are much smaller than those simulated with BB emissions; this
fact confirms that the BB plumes observed during the YAK-AEROSIB campaign are
reasonably well matched in the simulations by large concentrations of CO
originating from vegetation fires.
Relationships between BB BC and PM2.5 concentrations obtained
from the YAK-AEROSIB observations and from the simulations performed with the
optimized BC and OC emissions. The relationships were obtained by
selecting PM2.5 and BC concentrations matching (in space and time) the
measured CO mixing ratios that exceed the 90th percentile (a) or
80th percentile (b) of the observed distribution of the CO mixing
ratios. Purple dots depicted in (b) represent the observational
data shown in (a). The figure legends give the equations for a
simple linear regression without an intercept. The shaded areas indicate the
90th percentile confidence intervals for the linear regression lines
fitted to the measurement data.
Figure shows the relationships between the PM2.5 and BC
mass concentrations selected as explained above. To evaluate these
relationships, they were fitted with linear regressions without intercepts;
the fit equations are reported in the legends of Fig. . Assuming
that the contribution of the background aerosol fraction to the selected BC
and PM2.5 measurements was negligible, we regard the value of the slope
of the best fit line as an estimate of the BC/PM2.5 ratio in BB
aerosol measured during the flights. Note that according to our simulations,
the background BC and PM2.5 concentrations corresponding to the selected
measurements were, on average, very small (only 0.02 and 14 µgm-3, respectively, for the selection criterion based on the
80th percentile) compared to the range of the values presented in
Fig. . The slopes of the fits to the observational data are about
0.021 for any of the two selection criteria considered. This value is in the
middle of the range of the eBC/PM2.5 ratio values (0.01–0.045) previously observed
in Siberian smoke plumes (Kozlov et al., 2008). For comparison, the
BC/PM2.5 ratio for fresh BB aerosol in extratropical forest is assumed
to be 0.033 in the GFED4 inventory (van der Werf et al., 2017); that is, a
factor of 1.5 larger than the value found in this study.
The large scatter of the experimental data points may reflect the actual
variability of the BC/PM2.5 ratios in BB aerosol particles sampled by
the aircraft instruments, although it may also be due to the measurement
uncertainty, including temporal mismatches between BC and PM2.5
measurements. The emissions from the flaming and smoldering phases of fires
have very different BC/PM2.5 ratios, and an aircraft flying through
plumes near the fires often passes through sub-plumes originating from the
different fire phases and thus having very different compositions. After some
transport, the smoke from the flaming and smoldering parts of fires becomes
well mixed in the plumes. This may explain why the scatter is smaller in the
relationships between EC and OC concentrations in the BB aerosol samples
collected at the ZOTTO site (see Mikhailov et al., 2017, and Fig. a
therein). In contrast, the scatter of the simulated data points is very
small. The variability of the BC/PM ratios may be strongly underestimated in
our simulations as a result of the simplistic model representation of the
complex patterns of the spatial and temporal variability of BB BC and PM2.5
emissions and also due to the probably inadequate representation of the BB
aerosol aging processes in CHIMERE.
The BC/PM2.5 ratio in our simulations is about 3 % and 10 % larger than
the corresponding estimate derived from the YAK-AEROSIB measurements with the
selection criteria based on the 90th and 80th percentiles of the CO
mixing ratio, respectively. As the eBC concentrations measured with an
Aethalometer are likely to be different from EC concentrations measured using
a thermo–optical method (see Sect. ), these differences are
not indicative of any biases in our estimates of BC emissions (which are
evaluated in this study as emissions of EC). Furthermore, any discrepancy
between the slopes of the best fits to the observational and simulation data
could easily be eliminated by decreasing the correction factors FBC
(and thus BC emissions) in the simulations within the uncertainty range of
the optimal estimates of FBC.
Unfortunately, due to the absence of frequent measurements of an independent
tracer of biomass burning in the YAK-AEROSIB observations, we could not use
them for an evaluation of model predictions of the absolute values of BC and
PM2.5 concentrations. Unlike the measurements at ZOTTO, which were
performed in an almost pristine environment, the aircraft trajectory during
the YAK-AEROSIB campaign passed over polluted areas near large cities, where
the contributions of anthropogenic sources to the BC and PM2.5
concentrations could be considerable or even predominant. Nonetheless, we
could compare our simulations with the campaign-average concentrations.
Accordingly, we found that the average respective BC and PM2.5 concentrations were
0.62 and 22 µgm-3 in the observations, while the
average concentrations of their simulated matchups were 0.44 and 28 µgm-3, respectively.
In view of the potentially large measurement uncertainties as
well as the limited representativeness of the aircraft measurements at the
scales resolved in our simulations, the differences between these average
concentrations can not be considered as clear evidence for biases in either BB
or anthropogenic emissions specified in our model. Overall, the comparison of
our simulations with the YAK-AEROSIB data shows a reasonable agreement,
although it also highlights the difficulties and uncertainties associated
with the validation of BC simulations against the optical measurements of
aerosols.
Gridded estimates of the BB BC emission totals (gm-2)
obtained in this study (a) and those calculated using GFED4.1s (b)
and FEI-NE (c) data for the period from 1 May to 30 September 2012.
BC and OC emission estimates
Figure shows the spatial distribution of the
average BB BC emissions calculated for the study period in accordance with
Eq. () using the MODIS FRP data and optimal estimates of the
correction factors (see Sect. ) constrained with the OMI AAOD
and MODIS AOD retrievals. Not surprisingly, the distribution of BC emissions
generally replicates the spatial patterns of FRP (see Fig. ) and
is also similar to the distributions of AOD and AAOD shown in
Fig. . Grid cells with strong BC emissions cover vast areas in
western and central Siberia, as well as in eastern Siberia (east of Yakutsk
and along Russia's border with China). For comparison, Fig. also
shows the corresponding spatial distributions based on the data from the
GFED4.1s and FEI-NE emission inventories. All the distributions look rather
similar, although there are also many differences between them. Most of the
differences appear to have a random character, but it is noticeable that the
emissions obtained in this study and based on the FEI-NE data tend to be
stronger in many “hot spots” than those based on the GFED data. Greater
FEI-NE BC emissions compared to those from GFED4 can be explained by an
almost factor of 2 difference in the BC emission factors assumed in
FEI-NE and GFED4, as well as by differences in the methodologies used to estimate
fuel loadings. Similar reasons (that is, biases in the emission factors
and/or in the fuel consumption estimates involved in the GFED4 inventory) may
be behind the differences between the GFED4 data and our estimates. It is
also noticeable that a much larger number of grid cells in the distributions
based on our estimates are associated with relatively weak emissions in the
range from 0.01 to 0.05 gm-2 than in the distributions
based on both the GFED and FEI-NE data. This difference indicates that
emissions from some small fires (especially in agricultural areas) may be
missing in the GFED and FEI-NE inventories (based on the burnt area data) but
are taken into account in our calculations based on the FRP measurements.
Optimal estimates of the BC and OC mass (in Gg) emitted from
fires in the study region for individual months of 2012 and for the whole
period considered (1 May–30 September). The numbers given in brackets are
confidence intervals reported in terms of the 90th percentile.
Species
May
June
July
August
September
All months
BC
96.2 (±32.0)
71.4 (±24.3)
139.0 (±48.4)
78.9 (±27.9)
19.8 (±9.1)
405.3 (±134.6)
OC
1032 (±331)
1774 (±430)
3895 (±983)
1889 (±538)
254 (±121)
8844 (±2197)
BC amounts (in Gg) emitted from fires in the study region in
the period from May to September 2012: the estimates constrained by satellite AAOD and AOD
satellite observations are presented in comparison to corresponding estimates
calculated using the GFED4.1s and FEI-NE data. The error bar in the positive
direction for the FEI-NE estimate is not shown to improve the readability of the
figure.
Figure and Table report our top-down estimates of
the total monthly BC emissions from fires in the study region, as well as the
estimate of the integral BB BC mass emitted in the study region in the period
from May to September. The uncertainties of our estimates are reported in terms of
the 90th percentile confidence level. Our estimates are shown in
comparison with the corresponding values calculated using the GFED4 and
FEI-NE emission data. The uncertainty level in the GFED4 data has not been
reported, and therefore it is not indicated in Fig. . Note,
however, that previous studies in which AOD simulations based on the GFED
inventory have been evaluated against corresponding observations in different
regions of the world (see, e.g., Tosca et al., 2013; Reddington et al., 2016;
Petrenko et al., 2017) have indicated that the GFED data for BB aerosol emissions
may be very uncertain, such that they need to be corrected with adjustment
factors sometimes exceeding 10 on a regional scale. The uncertainty reported
for the FEI-NE data is 63 % (Hao et al., 2016). It has not been specified
whether this uncertainty characterizes gridded data or total regional
emission estimates; we assume here that the latter is true.
According to our estimates, the fires in the Siberian study region released
405 (±135) Gg of BC during the study period. This value is many
times larger than the total BC amount (25 Gg) that was emitted from
other sources in the study region and period, according to our calculations
based on the data of the ECLIPSE V5 emission inventory for 2010 (Klimont et
al., 2017). For comparison, our estimate of the total BB BC emissions is also
much larger than the total annual anthropogenic BC emissions in North America
(249 Ggyr-1) and less than a factor of 2 smaller than the
total annual anthropogenic BC emissions in Europe and Russia (660 Ggyr-1)
in 2010 (Klimont et al., 2017). About 40 % (139 Gg) of
the total amount of BC released from the fires during the whole study period
was emitted in July. The emissions were smallest in September (20 Gg)
and ranged from 71 to 96 Gg in May, June, and August. Note again that
BC emissions are evaluated in this study as emissions of EC.
Our estimates indicate that the total BC emissions from Siberian fires in the
period considered are strongly underestimated in the GFED4 inventory (by more
than a factor of 2), in which these emissions are estimated at about 198 Gg.
Taking into account that GFED is widely used as a “reference”
database for estimations of atmospheric and climatic effects of open biomass
burning, we believe that this is a significant finding. The relative
difference between our monthly BC emission estimates and the corresponding
GFED4 data is largest in September, exceeding a factor of 8; it is also large
(a factor of 3) in May. In contrast, the BC emissions in the FEI-NE inventory
(614 Gg) are larger than ours, although this difference is not
significant in view of the reported uncertainty in the FEI-NE data. Note
that, while no specific measure (EC, eBC, rBC) is identified for BC in the
GFED4 inventory, Hao et al. (2016) specified that the FEI-NE inventory
employed emission factors for refractory BC (rBC). However, we are not aware
of any procedure that could allow us to adjust for the differences between
rBC and EC. Overall, our top-down estimates provide a compromise between the
data of the GFED4 and FEI-NE inventories. Importantly, the evaluated
uncertainty in our estimates is much smaller than both the differences
between the estimates based on the two inventories considered and the
reported uncertainty of the FEI-NE data. Therefore, the satellite data
provide stronger constraints on BC emissions from Siberian fires, compared to
the state-of-the-art emission inventories.
(a) OC amounts (in Tg) emitted in the study region
in period from May to September 2012 according to this study and the GFED4.1s data along with
(b) the corresponding estimates of the BC/OC emission ratios (gg-1).
Although the estimation of OC emissions was not the focus of our study, our
top-down estimates of the OC emissions (see Fig. a and
Table ) are useful to consider here, as they allow us to further
evaluate the overall integrity of our method and results. We also report
BC/OC emission ratios (see Fig. b) calculated as the ratio of our
estimates for BC and OC emissions along with the emission ratios calculated
using the GFED4.1s data. Note that FEI-NE does not provide data on OC
emissions.
The results shown in Fig. a indicate that the pattern of monthly
variations of OC emissions is not very similar to that of BC emissions.
Specifically, the OC emissions in May are found to be much smaller than in
June, while the BC emissions were larger in May (see Fig. ).
However, similar to our BC emission estimates, our estimates of OC emissions are much
larger than the corresponding estimates based on the GFED4 data. Our estimate
for the integral emissions over the fire season considered is a factor of 2.2
larger than the corresponding estimate based on the GFED4 data. Based on
a comparison of satellite-derived and simulated AOD, several previous studies
have shown evidence that OC emissions provided by the GFED inventory may indeed
be underestimated in different regions of world, including Siberia (see,
e.g., Petrenko et al., 2012, 2017; Tosca et al., 2013; Konovalov et al.,
2014, 2015; Reddington et al., 2016), although it has also been argued (Konovalov
et al., 2015, 2017a) that models may underestimate AOD due to inadequate
representations of the BB aerosol aging processes. Therefore, it is possible that a
part of the differences between our optimal estimate of the OC emissions and
the corresponding GFED data may compensate for some missing processes (e.g.,
involving the formation of SOA due to oxidation and condensation of
semi-volatile organic compounds) in our model.
Spatial distributions of the relative contribution of both grassland
and agricultural (“grass”) fires to the BB BC emissions integrated over a
monthly period (a, c, d) along with the spatial distributions of the corresponding
BB BC emission values (b, d, f) for May (a, b), July (c, d), and September (e, f),
2012. The distributions were obtained using Eq. () and the optimal
estimates of the correction factors, FBC.
In spite of the very significant differences of our BC and OC emission
estimates with respect to the GFED4 data, the BC/OC emission ratio (0.046±0.014 gg-1) obtained in our analysis (see
Fig. b) is consistent, in the case of the integral emissions for
the study period, with that in GFED4 (0.054 gg-1).
Furthermore, the monthly variations of the BC/OC emission ratio according to
our estimates are qualitatively similar to those according to the GFED4 data.
Specifically, both the GFED4 inventory and our estimates indicate that the
BC/OC emission ratio was bigger in May and September than in the summer
months. However, our estimates also indicate that the BC/OC emission ratio
may be underestimated by GFED4 in May and overestimated in the summer months.
Monthly variations of the BC/OC emission ratio in the GFED inventory are a
result of changes in the presumed fire fuel: in particular, the monthly
variations shown in Fig. b indicate that, according to the GFED4
data, the contributions of agricultural and grass fires to the BB BC
emissions were slightly bigger in May and September than in the summer
months. The same factor can explain (at least partly) the monthly variations
in our estimates of the BC/OC emission ratio. To illustrate this point,
Fig. shows the spatial distributions of the relative
contribution of agricultural/grass fires to BB BC emissions integrated over a
month, along with the spatial distributions of the corresponding BB BC
emission values. The distributions were obtained for three different months
(May, July, and September) using Eq. () and the optimal estimates
of the correction factors, FBC. Evidently, the fires that occurred in
the study region in May mostly burned in agricultural lands and grasslands,
even though the BC emissions from intense forest fires were also quite
significant in several grid cells. In contrast, forest fires were clearly
predominant in July (as well as in the other summer months). Unlike the
situations in both May and July, BB BC emissions in September were not
clearly associated with any predominant fuel category: along with
agricultural/grass fires in the southwestern and southern parts of the study
region, there were relatively strong forest fires north of Tomsk and
Krasnoyarsk and west of Yakutsk. Taking these observations into account, it
can be speculated that the big difference between the BC/OC emission ratios
in July and September is, to some extent, a manifestation of the diversity of
fire regimes across the boreal region (Conny and Slater, 2002). Note that the
spatial distribution of our emission data is insufficient to enable us to
distinguish between agricultural and grassland fires. However, according
to the GFED4 inventory, agricultural burns strongly dominate over grass fires
both in May and September (by a factor of 5 at least).
It is noteworthy that our estimates of the BC/OC emission ratios (in the
range from 0.036 to 0.042 gg-1) for the summer months are
only insignificantly – taking the confidence intervals into account –
different from the EC/OC ratio of 0.038 gg-1 that was
derived for BB aerosol by Mikhailov et al. (2017) from aerosol measurements
at ZOTTO in summer. Furthermore, our estimate for May (0.093±0.03 gg-1)
is in a good agreement with the EC/OC ratio (0.08±0.02 gg-1) found by Mikhailov et al. (2017) for BB
aerosol predominantly originating from agricultural fires in spring. As SOA
formation simulated with the “standard” aerosol module of CHIMERE contributes
very insignificantly to BB aerosol concentrations (Konovalov et al., 2015,
2017a), the ratios of the BC and OC emissions specified in our simulations
are quantitatively almost the same as the simulated BC/OC ratios in the
ambient aerosol particles, irrespective of their age. Therefore, our
estimates of the BC/OC emission ratios look reasonable in view of the
independent ambient observations in central Siberia. This finding confirms
the validity of our estimation method and the results obtained.
Estimates of the total BC emissions (in Tg) from fires in the
study region over the period from 1 May to 30 September 2012 for several test
cases of the estimation procedure.
Test case
Brief description
Total BC
Relative difference with respect
no.
emissions (Gg)
to the base case estimate (%)
1
BB aerosol photochemical age is larger than 11 h
416
+2.6
2
The SOA yield in the simulations is increased by a factor of 7
424
+4.6
3
BB aerosol particles are assumed to be hydrophilic and affected by in-cloud scavenging
431
+6.3
4
The background AOD is reduced by 50 %
477
+17.8
5
The background AOD is enhanced by 50 %
398
-1.6
6
The background AAOD is disregarded
443
+9.2
7
The OMI “final AAOD” data product is used instead of the retrieval data provided for different aerosol layer heights
317
-21.8
8
A weaker selection criterion (γ=0.5, see Eq. 8) is used
477
+17.8
9
A stricter selection criterion (γ=1.5, see Eq. 8) is used
388
-4.1
10
Any gridded data point considered includes at least 10 AAOD pixels
457
+13.0
Sensitivity tests
The confidence intervals for our optimal estimates of BC emissions
(Table and Fig. ) do not necessarily include possible
uncertainties and biases that may be associated with systematic model errors
and data selection criteria. Based on our understanding of likely reasons for
such uncertainties and biases, we specified 10 sensitivity tests listed in
Table 4. The sensitivity analysis was focused on the estimation of the total
BC emissions over the study period. The correction factors FBC and
FOC for each test case were obtained by applying Eqs. () and
() to the optimal (“base case”) estimates of the correction
factors (see Table ). One more iteration of the estimation
procedure was sufficient to obtain the test estimates of the total BC
emissions with a relative numerical error of 3 % or less. The total BC
emission estimates for each case and the relative differences with respect to
the base case estimate reported in Table are also listed in
Table .
Test case no. 1 addresses systematic differences between the photochemical
ages of BB aerosol observed at the AERONET sites that provided the data
considered in our analysis (see Sect. ) and those of BB
aerosol observed by satellites. Figure shows the histograms of
the photochemical ages estimated in accordance with Eq. ()
for the respective AERONET and satellite data from the datasets
selected for our analysis. Compared to BB aerosol observed from satellites
(which have a median photochemical age of 15.4 h), the BB aerosol at
the AERONET sites was typically more aged (with a median photochemical age of
25.8 h). If the relationship given by Eq. () is sensitive to
the photochemical age of the aerosol, these differences could result in some
bias in the modeled AAOD values. To get an idea about the significance of
such bias, we disregarded satellite data corresponding to photochemical ages
smaller than 11 h. The remaining satellite data have approximately the
same median photochemical age as the AERONET data. This restriction resulted
in a small change of the optimal BC emission estimate, which increased by
less than 3 %. This result does not necessarily mean that the BB aerosol
composition and its optical properties are not strongly affected by aging;
it may actually mean that changes of AOD and AAOD, as well as those of the
monthly BC emission estimates due to aerosol aging, tend to compensate each
other in the total BC emission estimate.
Histograms of the BB aerosol photochemical ages estimated in
accordance with Eq. () for the (a) satellite and (b) AERONET data
selected for this study. Note that very minor fractions of the data points,
which correspond to the age exceeding 60 and 56 h in the
cases of satellite and AERONET data, respectively, are not represented in the
histograms for the sake of better readability.
The impact of aerosol aging on our estimates is further addressed in test
case no. 2. Specifically, the goal of this test case is to assess a potential
bias in our BC emission estimates due to a probable underestimation of the
SOA contribution to aged BB aerosol. To this end, we performed a simulation
in which the yields of all SOA species from the oxidation of major volatile SOA
precursors (such as toluene, xylenes, isoprene, and terpenes) were enhanced by
a factor of 7 with respect to the “base-opt” simulation, while the reaction
list and the reaction rates (as well as all other simulation settings) were
kept unchanged. As a result of this model modification, a relative
enhancement of the averaged (over the whole period and region considered) POM
column amounts due to SOA formation increased from only 2.6 % (in the
“base-opt” case) to up to 27 % (in the test case). Note that the SOA enhancement
increased more strongly than the SOA yields, probably because some of the SOA
species are assumed to be semi-volatile in CHIMERE; thus, the condensed
fraction of these species increases with their total concentration. The increased SOA
contribution to POM in our test case simulation corresponds to the upper
margin of the wide range of the BB POM enhancements observed in aging BB
plumes in several field studies in North America (Cubison et al., 2011). In
spite of the considerable changes in the simulated POM fields, the optimal BC
emission estimate only showed insignificant change, increasing by 4.6 % (see
Table ). This result is in line with the above discussion regarding the
robustness of our BC emission estimates (see Sect. ) with
respect to the treatment of POM aging in our model. However, it is not surprising
that our estimates of the correction factors FOC and of the
OC emissions were found to be more sensitive to the changes in the POM
simulations; specifically, the top-down estimate of the total OC emissions
dropped by ∼15 % in the test case as a result of the increases in
aerosol abundances and of a slight decrease in the mass extinction
efficiency.
Test case no. 3 addresses the uncertainty associated with the representation
of wet deposition of BB aerosol particles in our simulations. As noted in
Sect. , we assumed that BB aerosol particles are hydrophobic
and therefore are not susceptible to in-cloud scavenging; accordingly, the
empirical uptake coefficient in the base case simulations with BB emissions
was set to be zero. This assumption can result in the overestimation of the
lifetime of BB aerosol particles, which tend to become more hydrophilic as
the aerosol ages (Paramonov et al., 2013); this, in turn, results in a negative
bias of our BC emission estimates. To get an idea of the magnitude of this
possible bias, we performed a test simulation (with the optimized BB
emissions) in which the empirical uptake coefficient was set to be unity;
this setting corresponds to the assumption that BB aerosol particles are
hydrophilic. The use of this simulation instead of the original simulation
with the optimized BB emissions in our estimation procedure only resulted in
a minor increase (∼6 %) in our optimal estimate of the total BB BC
emissions; the changes in the monthly estimates are found to be similarly
small. This is an expected result, as the major fires considered in our
analysis mostly occurred during dry periods with low precipitation.
Therefore, the test case no. 3 indicates that probable changes in the
hygroscopicity of ambient aerosol particles due to BB aerosol aging processes
will not significantly affect our BC emission estimates.
Test cases no. 4 and no. 5 are designed to evaluate the extent to which our
top-down BC emission estimate can be affected by a possible bias in the
background AOD values predicted by CHIMERE. To get an idea about such a bias,
we followed the approach suggested by Konovalov et al. (2014). Specifically,
we first selected the days and grid cells (irrespective of the availability
of AAOD data) in which the MODIS-retrieved AOD data are available and the
contribution of fires to the modeled AOD values (corresponding to the
selected the days and grid cells) does not exceed 10 % of the background AOD
values. We then evaluated the mean difference between the MODIS-retrieved
and modeled AOD for these selected data points. We found that the mean value
for the modeled AOD (∼0.17) is considerably higher than the mean value
(∼0.10) for the observed AOD. Taking into account that the bias in the
background AOD values in pixels affected by fires may be somewhat different
from that representative of background conditions, we considered larger
changes in the background AOD by increasing or decreasing it by 50 %. The
test results indicate that a probable positive bias in the background AOD
values is associated with some underestimation (by less than 20 %) of BC
emissions in our procedure; if the bias were absent, the difference between
the BC emission estimates inferred from the satellite observations and those
calculated with the GFED4 data would be even larger than in the base case.
The sensitivity of the optimal estimate is strongly asymmetric with respect
to the enhancement and reduction of the background AOD: this is probably due
to an impact of the changes in the background AOD on the selection of data
according to the criterion given by Eq. ().
Test case no. 6 addresses the uncertainties associated with the background
AAOD. On the one hand, the AAOD data have been retrieved from the OMI
measurements under the assumption that each observed pixel is characterized
by only one type of aerosol. Consequently, the absorption caused by other
types of aerosol has effectively been disregarded, although it might actually
affect the AAOD retrievals. Thus, to prevent overestimation of the BC
emissions, the background AAOD (predicted by CHIMERE) was subtracted from the
AAOD retrievals as suggested by Eq. (). On the other hand, BB
plumes typically reach much higher altitudes than anthropogenic
aerosol: this is taken into account in the OMAERUV retrieval algorithm by
assuming that the vertical distribution of urban/industrial aerosol is
largest at the surface, while the concentration of carbonaceous aerosol in
smoke layers at mid- and high-latitudes typically peaks at 6 km. The
AAOD values retrieved by assuming that the aerosol layer is residing near the
ground are much larger than those corresponding to the assumed heights of 6 km
or even 3 km. So, if the aerosol in a given pixel is
identified as carbonaceous BB aerosol, a part of AAOD corresponding to the
anthropogenic aerosol is likely to be underestimated in the retrievals.
Therefore, a simple subtraction of the background AAOD values from the AAOD
retrievals may result in an underestimation of the BC emissions in our
analysis. To get an idea about the maximum magnitude of this underestimation,
the background AAOD values were entirely disregarded in test case no. 6. The
test result indicates that the underestimation is probably rather small (less
than 10 %); however, it may actually be larger if the background
AAOD values in our simulations are biased high. Unfortunately, we can not
properly evaluate the possible overestimation of the BC emissions in the case
where the background AAOD is strongly underestimated. However, as noted
above, the simulated background AOD is overestimated, so it seems reasonable
to assume that the background AAOD is also overestimated. Accordingly, we
believe that the uncertainty of the best estimate of the BB BC emissions with
respect to the intrinsic uncertainty associated with the background part of
the AAOD retrievals is likely within the difference between the estimates
given by the “base-opt” case and test case no. 6.
As noted above (see Sect. ), the AAOD retrievals
corresponding to different assumed altitudes of the aerosol center of mass
were selected in our analysis using the smoke layer heights derived from
our simulations. Ideally, this approach ensures that the AAOD retrievals are
consistent with the observed variations in the location and intensity of the
fires. Nonetheless, in view of the possible uncertainties in the simulated
vertical distributions of the BB aerosol, it is also useful to consider the
BC emission estimates derived from the standard (OMI “final”) data product
based on rather rough (climatological) estimates of the smoke layer heights.
This is done in test case no. 7. We found that the standard data product
yields a 22 % lower BC emission estimate than the base case estimate.
The difference between the two estimates is considerable, but it is still
well within the uncertainty limits of the base case estimate.
Test cases no. 8 and no. 9 examine the sensitivity of our estimates to the
selection criterion defined by Eq. (). Specifically, we used a 50 %
higher and 50 % lower values of γ for test cases no. 8
and no. 9, respectively. The larger value of γ selects data points
with larger values of AOD and vice versa. The emission estimates obtained
with a smaller value γ are more prone to uncertainties associated with
the background AOD. Conversely, a stricter selection criterion results
in the loss of information about relatively small fires. Nevertheless, the test
results show that the sensitivity of our estimates to the big changes in
γ is relatively weak, suggesting that our base case estimates are
sufficiently robust with respect to the selection criterion considered.
Finally, test case no. 10 is designed to address a potential issue concerning
the representativeness of the OMI retrievals in view of the rather coarse
resolution of our simulations. It seems reasonable to expect that when, for
example, only one AAOD observation corresponding to BB aerosol is available
for a given grid cell, the mean observed AAOD value inferred in our procedure
for this grid cell is likely to be overestimated, as AAOD over the rest of
the grid cell's area may be much smaller. However, the overestimation can
hardly be very large for the very intense and widespread Siberian fires
considered, because, in this case, the smoke plumes are likely to cover a
large fraction of the grid cell area. To examine this issue, we disregarded
any gridded AAOD data points that comprised less than 10 different AAOD
observations (data pixels), while the maximum number of the pixels per grid
cell in the data considered equals 26. Contrary to our expectations, we found
that the estimate obtained in test case no. 10 is larger (by 17 %). This
increase is found to be mostly due to an increase in the optimal estimates of
FOC. Apparently, the above limitation resulted in the selection of MODIS
AOD data that are more representative of grid cells affected by major fires
and are matched by smaller AOD values simulated for the base case. Therefore,
the result of this test is not indicative of any representativeness issue for
the available OMI retrievals.
In general, the results presented in this
section demonstrate that our estimate of the total BC emissions from Siberian
fires is sufficiently robust with respect to possible uncertainties in the
input data and the choices made in the estimation procedure. In particular,
these results strongly support our findings that the GFED4 inventory
significantly underestimates the BC emissions from Siberian fires.
(a) The estimates of the BC mass (in Gg) transported from the
study region across the polar circle into the Arctic along with (b) the
corresponding estimates of the BC transport efficiency (see
Sect. ).
BC transport into the Arctic
As argued in the introduction, studying BB BC emissions in Siberia is
stimulated by the need to properly evaluate the role of BC in Arctic climate
change. Therefore, it is important to know not only the amount of BC emitted
from the fires but even more so the amount of BC transported into the Arctic.
Using the three-dimensional hourly fields of BC mass concentrations and of
the meridional component of wind speed from our optimized simulations, we
calculated the hourly BC fluxes from the study region across the polar circle
(66∘33′ N) and then integrated them over altitude (from the
surface up to the model domain top coinciding, approximately, with the
tropopause), longitude, and time on a monthly basis. The fluxes were
calculated separately for BC emitted from fires and from anthropogenic
sources. In this way, we evaluated the total masses of BB and anthropogenic
BC transported from the study region into the Arctic each month (see
Fig. ). Note that a part of the BC mass transported into the
Arctic may be transported back out of it to the study region (when the
corresponding transport times are shorter than the typical lifetime of BB
aerosol with respect to deposition); however, such backward transport of BC
is mostly not taken into account in our calculations, as the model
domain does not extend to the whole Arctic. We also evaluated the BC
transport efficiency, defined here as the ratio of the BC amounts transported
to the Arctic to the corresponding amounts of BC emitted from Siberian fires.
This definition is similar but not identical to that introduced in Evangeliou
et al. (2016), where the transport efficiency was defined as the ratio
between the mass of BC deposited in the Arctic and the mass of BC emitted
from a given region.
Our estimates indicate that vegetation fires contributed a predominant part
(as large as 95 %) of the integral BC mass transported into the Arctic from
the study region during the 5 months considered (see Fig. a).
This amount corresponds to an overall transport efficiency of about 27 % (see
Fig. b): that is, about a quarter of the total BC emitted from
Siberian fires was transported into the Arctic. This estimate of the
transport efficiency is comparable with that (about 30 %) obtained by
Evangeliou et al. (2016) for BC emitted from fires in Asia in the summer
periods of 2012 and 2013. Our results show that the transport efficiency was not
constant across the different months. In particular, it exceeded 60 % in
September and was less than 15 % in June. Interestingly, the total BB BC mass
transported into the Arctic in September is found to be slightly larger than
that in June (see Fig. a), in spite of the fact that the amount
of BB BC emissions is more than a factor of 3 larger in June than in
September. This fact emphasizes the potential climatic importance of the
fires that occur in Siberia in early fall. According to our results (see
Sect. ), the BB emissions from these fires (which were most
intense about 300–500 km west of Yakutsk, see
Fig. f) are very strongly (by a factor of 8) underestimated in
the GFED4 inventory (but note also that our BB BC emission estimate for
September is very uncertain). By using the model run for test case no. 3
(see Sect. ) instead of the “base-opt” run, we made sure that
disregarding the impact of BB aerosol aging on the hygroscopicity of aerosol
particles in our simulations did not have a significant effect on our
analysis of BC fluxes. In particular, the overall transport efficiency
evaluated under the assumption that BB aerosol particles are composed of
hydrophilic material turned out to be only slightly smaller (25.2 %) than the
corresponding base case estimate (27.6 %); among the individual months, the
transport efficiency decreased most in May (from 29 % in the base case to
24 % in the test case). As a caveat, it should be noted that due to
interannual meteorological variability, our monthly estimates of the
transport efficiency in 2012 may not be applicable to other years. To improve
the current understanding of the role of Siberian fires in Arctic
warming, the analysis suggested in this paper should be extended to a
multi-annual period.
Conclusions
We have investigated the feasibility of
constraining BC emissions from open biomass burning with AAOD retrievals from
OMI satellite measurements by considering the case of the severe fires that
occurred in Siberia in 2012. We developed an inverse modeling procedure
enabling the optimization of BB emissions based on MODIS FRP measurements by
combining OMI AAOD retrievals and MODIS AOD data with simulations performed
using the CHIMERE CTM. To limit possible errors in the simulated AAOD data due
to uncertainties in the absorption properties of the BB aerosol, we employed
an empirical parameterization predicting AAOD as a function of AOD and the
ratio of BC and OC column densities. The parameterization is based on the
experimental findings reported earlier (Pokhrel et al., 2016) and is fitted
to data from two AERONET sites in Siberia; it assumes that the SSA of BB
aerosol particles is a linear function of the elemental to total carbon
ratio. As a result of the application of our inverse modeling procedure to
the measurement and simulation data characterizing the BB aerosol in Siberia
during the period from 1 May to 30 September, we evaluated the monthly
correction factors for BB BC and OC emissions calculated using the FRP data
and obtained top-down estimates of the total BC and OC amounts emitted each
month in the period considered. Note that our estimation method implies that
the BC emissions are evaluated as emissions of elemental carbon (EC) measured
using a thermo–optical technique.
To validate the optimized BC and OC emissions, we used them to perform
simulations that were evaluated against independent observational data.
Specifically, we first compared our simulations with the OMI AAOD and MODIS
AOD data that had been withheld from the optimization procedure. A reasonable
agreement between the observations and simulations is found in the spatial
distributions and daily time series of the both AAOD and AOD data. In
particular, the correlation coefficients for the time series of spatially
averaged AAOD and AOD values were found to be 0.79 and 0.84, respectively.
Our simulations were further compared with in situ measurements of EC and OC
mass concentrations at the top of the 300 m tower at the ZOTTO site
(Mikhailov et al., 2017), situated at a remote location in central Siberia.
Although the simulated EC concentrations turned out to be about 23 % larger
than the observed concentrations, the bias was not found to be significant considering
the uncertainties of our emission estimates and random model errors. A minor
negative bias of about 7 % is found in the simulations of OC concentrations.
It should be noted that unlike the satellite data, which cover the whole
study region, the in situ measurements of BB aerosol may contain some local
features of fire regimes and fuels, which could not be reproduced in our
simulations. We also compared our simulation with optical measurements of BC
and PM2.5 mass concentrations onboard an aircraft in the framework of
the YAK-AEROSIB experiments. Due to a problem with distinguishing between the
significant (on average) contributions of anthropogenic and BB sources to the
measured aerosol concentrations, a direct comparison of the simulated and
measured BC concentrations would not be sufficiently informative of the
accuracy of our simulations of BB aerosol. Instead, we focused on a
comparison of the relationships between the BC and PM2.5 concentrations
in the simulations and observations using measurements of CO concentration
to select the observations most representative of BB aerosol. The slopes of
linear fits to the BC and PM2.5 data from the simulations and
observations are found to be in good agreement (within 10 %). This finding
further confirms that the BB aerosol composition was simulated adequately.
We found that Siberian fires emitted 405±135 Gg of BC in the
period from May to September 2012 (at the 90 % confidence level). The BB BC emissions were
largest in July, when 139±49 Tg of BC was emitted and smallest in
September (20±9 Gg). Our estimates were compared to the
corresponding estimates obtained from the GFED4 and FEI-NE databases. Our
estimate of the total BB BC emissions in the study region and period is found
to be a factor of 2 larger than the GFED4 estimate, but a factor of 1.5
smaller than the FEI-NE estimate. The differences of our monthly and
season total BC emission estimates with respect to both GFED4 and FEI-NE data
are statistically significant, although the differences with respect to the
FEI-NE estimates are smaller than the large uncertainty range reported for
the FEI-NE data.
The results of several sensitivity tests indicate that, although our
estimates can be influenced to some extent by a number of factors associated,
in particular, with data selection criteria and uncertainties in the
simulations of optical properties of aerosol in the absence of fires, the
possible bias in our estimate of the total BC emission is unlikely to exceed
the estimated uncertainty of about 35 %.
In spite of the significant differences between our BC emission estimates and
the GFED data, the ratio of the total BC and OC emission estimates derived
from the satellite data (0.046±0.014 gg-1) is found
to be consistent with the ratio of the corresponding BC and OC emission
totals according to the GFED data (0.054 gg-1). However,
there are considerable differences between the BC/OC emission ratios obtained
in this study and those calculated using the GFED4 data for the different
months. In particular, a larger value of the ratio of BC and OC emissions in
May is found in this study (0.093±0.030 gg-1)
compared to that suggested by GFED4 (0.06 gg-1): this
difference may be indicative of an underestimation of BC emissions from
agricultural burns and grass fires in the GFED4 inventory.
Finally, we estimated that about a quarter of the huge BC amount emitted from
Siberian fires in the period from May to September 2012 was transported across the polar circle
into the Arctic. Therefore, the results of this study have a direct
implication for reducing major uncertainties associated with the current
estimates of sources of BC in the atmosphere and snow/ice cover in the Arctic
and for improving the general understanding of the role of BC in the Arctic
climate system.
Overall, our analysis demonstrated that the OMI AAOD retrievals combined with
the MODIS AOD data can provide useful constraints to the BB BC emissions. It
is especially noteworthy that in the case considered in this study, the
entire uncertainty range for the BC emission estimates constrained with the
satellite measurements turned out to be a factor of 1.5 smaller than the
difference between the corresponding estimates provided by the two
state-of-the-art emission inventories, GFED4 and FEI-NE. A major factor
limiting the accuracy of the top-down estimates of BC emissions from Siberian
fires is the uncertainty of the AAOD simulations. To reduce this uncertainty,
more data from remote sensing and in situ measurements of aerosol optical
properties and composition (such as measurements of SSA and of the BC/OC and
OC/POM ratios) in northern Eurasia are needed. Another significant
uncertainty source in our estimates is associated with the estimation of the
altitude of the aerosol layer center of mass. Accordingly, future
developments of our approach should include an evaluation and optimization of
the simulated vertical distribution of BB aerosol using suitable satellite
observations, such as, e.g., Cloud-Aerosol Lidar and Infrared Pathfinder
Satellite Observations (CALIPSO) (Vaughan et al., 2004).