Understanding the role atmospheric aerosols play in the Earth–atmosphere system is limited by uncertainties in the knowledge of their distribution, composition and sources. In this paper, we use the GEOS-Chem based inverse modelling framework for retrieving desert dust (DD), black carbon (BC) and organic carbon (OC) aerosol emissions simultaneously. Aerosol optical depth (AOD) and aerosol absorption optical depth (AAOD) retrieved from the multi-angular and polarimetric POLDER/PARASOL measurements generated by the GRASP algorithm (hereafter PARASOL/GRASP) have been assimilated. First, the inversion framework is validated in a series of numerical tests conducted with synthetic PARASOL-like data. These tests show that the framework allows for retrieval of the distribution and strength of aerosol emissions. The uncertainty of retrieved daily emissions in error free conditions is below 25.8 % for DD, 5.9 % for BC and 26.9 % for OC. In addition, the BC emission retrieval is sensitive to BC refractive index, which could produce an additional factor of 1.8 differences for total BC emissions. The approach is then applied to 1 year (December 2007 to November 2008) of data over the African and Arabian Peninsula region using PARASOL/GRASP spectral AOD and AAOD at six wavelengths (443, 490, 565, 670, 865 and 1020 nm). Analysis of the resulting retrieved emissions indicates 1.8 times overestimation of the prior DD online mobilization and entrainment model. For total BC and OC, the retrieved emissions show a significant increase of 209.9 %–271.8 % in comparison to the prior carbonaceous aerosol emissions. The model posterior simulation with retrieved emissions shows good agreement with both the AOD and AAOD PARASOL/GRASP products used in the inversion. The fidelity of the results is evaluated by comparison of posterior simulations with measurements from AERONET that are completely independent measurements and more temporally frequent than PARASOL observations. To further test the robustness of our posterior emissions constrained using PARASOL/GRASP, the posterior emissions are implemented in the GEOS-5/GOCART model and the consistency of simulated AOD and AAOD with other independent measurements (MODIS and OMI) demonstrates promise in applying this database for modelling studies.
Atmospheric aerosols have a variety of sources and complex chemical
compositions. Desert dust (DD) aerosol is one of the most abundant types of
aerosol by mass, while the range of global dust emission estimates spans a
factor of about 5 (Huneeus
et al., 2011). Primary carbonaceous aerosol, which consists of black carbon (BC)
and organic carbon (OC) from combustion of fossil fuels, biofuels and
biomass, has strong light absorption that can affect the energy balance of
the Earth–atmosphere system (Bond et al., 2013). High
uncertainty in carbonaceous aerosol emissions (e.g.,
Bond et al., 2004) translates into a significantly high uncertainty in
evaluating their climate effects (Textor et al.,
2006). The Intergovernmental Panel on Climate Change (IPCC) estimates the
global mean direct shortwave radiative forcing due to primary carbonaceous
aerosol to be
Space-borne remote-sensing instruments offer an integrated atmospheric column measurement of the amount of light scattering by aerosols through modification of diffuse and direct solar radiation. Numerous satellite observations of the spatial and temporal distribution of aerosols have been conducted in the last 2 decades (King et al., 1999; Kaufman et al., 2002; Lenoble et al., 2013). The satellite retrievals of aerosol optical depth (AOD) and aerosol absorption optical depth (AAOD) are directly related to light extinction and absorption due to the presence of aerosol particles. AOD is a basic optical property derived from many Earth-observation satellite sensors, such as AVHRR (Advanced Very High Resolution Radiometer), MODIS (Moderate Resolution Imaging Spectroradiometer), MISR (Multi-angle Imaging SpectroRadiometer) and POLDER (Polarization and Directionality of the Earth's Reflectances) (Goloub et al., 1999; Geogdzhayev et al., 2002; Kahn et al., 2009; Tanré et al., 2011; Levy et al., 2013). AAOD is another valuable product to quantify the solar absorption potential of aerosols; however, only a few satellite aerosol products can provide retrievals of AAOD, and only with limited accuracy, for example OMI (Ozone Monitoring Instrument) on the Aura satellite making measurements in the UV range that have sensitivity to aerosol absorption (Torres et al., 2007; Veihelmann et al., 2007).
Despite their ability to provide global coverage in high spatial resolution, satellite measurements alone are not sufficient for addressing the question regarding the distributions, magnitudes and fates of aerosols in the atmosphere. These aspects can be studied using chemical transport models (CTMs), which rely on meteorological data from external databases with atmospheric physics, considering the physical and chemical processes in the atmosphere, and allow modelling of the detailed distribution of aerosol for any chosen time period (e.g. models by Balkanski et al., 1993; Chin et al., 2000; Takemura et al., 2000; Ginoux et al., 2001; Bessagnet et al., 2004; Grell et al., 2005; Spracklen et al., 2005; Mann et al., 2010). However, CTM simulations are limited by uncertainties in knowledge of aerosol emission characteristics, knowledge of atmospheric and aerosol processes, and the meteorological data used. As a result, even the most recent models are expected to capture only the principal global features of aerosol. For example, among different models, quantitative estimates of average regional aerosol properties often disagree by amounts exceeding the uncertainty of remote sensing of aerosol observations (Chin et al., 2002, 2014; Kinne et al., 2003, 2006; Textor et al., 2006). Therefore, there are diverse and continuing efforts to harmonize and improve aerosol modelling by refining the meteorology, atmospheric process representations, emissions and other components (e.g. aerosol aging scheme, particle mixing state) (Watson et al., 2002; Dabberdt et al., 2004; Generoso et al., 2007; Ghan and Schwartz, 2007; He et al., 2016; R. Wang et al., 2014, 2016).
One of the most promising approaches for reducing model uncertainty is to improve the aerosol emission fields (that is input for the models) by inverse modelling, i.e. fitting satellite observations and model estimates and by adjusting aerosol emissions (e.g. Bennett, 2002). For example, Dubovik et al. (2008) developed an algorithm for inverting CTMs and implemented the approach to retrieve distributions of aerosol emissions using MODIS data. The algorithm was used to implement the first formal retrieval of global spatial and temporal emission distributions of fine-mode aerosol from the MODIS fine-mode AOD data. Wang et al. (2012) and Xu et al. (2013) use MODIS radiances to constrain aerosol sources over China. Huneeus et al. (2012, 2013) optimize global aerosol emissions using MODIS AOD with a simplified aerosol model (Huneeus et al., 2009). However, as discussed in Dubovik et al. (2008) and Meland et al. (2013), MODIS AOD (as well as currently available aerosol satellite data) contains only limited information to evaluate aerosol types, properties or speciated emissions. Further, inconsistencies among representations of aerosol microphysics between the CTM and the aerosol retrieval algorithm can have significant influences on inverse modelling of aerosol sources (e.g. Drury et al., 2010; Wang et al., 2010). Therefore, the retrieval of aerosol emissions from satellite observations remains very challenging.
Recently a new dataset of spectral AOD and AAOD was generated using the GRASP
(General Retrieval of Atmosphere and Surface Properties) algorithm from
POLDER/PARASOL (Polarization & Anisotropy of Reflectances for Atmospheric
Sciences coupled with Observations from a Lidar) instrument
(Dubovik et al., 2011, 2014; data available
from ICARE data distribution portal:
Here we develop an inverse modelling approach to retrieve the spatial and temporal distributions of DD, BC and OC aerosol emissions simultaneously from PARASOL/GRASP spectral AOD and AAOD using the GEOS-Chem model (Bey et al., 2001) and its adjoint (Henze et al., 2007). Section 2 describes the model and data used in this study. The dust and carbonaceous aerosol model in the GEOS-Chem adjoint of Henze et al. (2007) is that of the GOCART (Goddard Chemistry Aerosol Radiation and Transport) model implemented in GEOS-Chem (Fairlie et al., 2007; Park et al., 2003), which is fully conceptually consistent with the aerosol model used in the inversion by Dubovik et al. (2008). The details of inverse modelling and performance evaluation of the inversion framework using numerical tests are presented in Sect. 3. In order to interpret the retrieval results and improve our understanding of aerosol emissions, we retrieve 1 year of daily DD, BC and OC emissions (see Sect. 4). Evaluation of these inversion results using independent AERONET, MODIS and OMI observations, as well as implementation of the posterior emissions in the GEOS-5/GOCART model, is presented in Sect. 5. Conclusions and discussion of the study's merits and limitations are considered in the Sect. 6.
Distribution of PARASOL/GRASP AOD retrievals per
0.1
Aerosol refractive index, size distribution and particle density for DD, BC, OC, SU, SS and host water employed in this study.
The study area (30
GEOS-Chem is a global three-dimensional CTM driven by
assimilated meteorological data from the NASA Goddard Earth Observing System
Data Assimilation System (GEOS-DAS) (Bey et al.,
2001). We use the GEOS-Chem (v9-02) model for aerosol simulation with 47
vertical layers and 2
The GEOS-Chem model assumes external mixing for aerosol components with
lognormal size distributions. The modal (
An adjoint model can be used as a tool for calculating the gradient of a scalar model response function with respect to a large set of model parameters simultaneously (Fisher and Lary, 1995; Elbern et al., 1997, 2000, 2007; Henze et al., 2004; Sandu et al., 2005). The adjoint of the GEOS-Chem model was developed specifically for inverse modelling of aerosols or their precursors and gas emissions (Henze et al., 2007, 2009). The 4D-variational data assimilation technique is used to optimize aerosol emissions by combining observations and model simulations. The adjoint of GEOS-Chem has been widely used to constrain emissions. For example, Kopacz et al. (2009) utilized MOPITT measurements of carbon monoxide (CO) columns to optimize Asian CO sources. Zhu et al. (2013) constrain ammonia emissions over the US using TES (Tropospheric Emission Spectrometer) measurements. Zhang et al. (2015) use OMI AAOD to constrain anthropogenic BC emissions over East Asia. However, these studies have focused on a single aerosol or gas species and kept others constant during the inversion since the satellites or other available observations of aerosols generally did not provide enough accurate information to estimate contributions from different species. The recent development of the PARASOL/GRASP retrieval, which retrieves more detailed and accurate aerosol information (see in Sect. 2.3), thus presents a new opportunity for constraining emissions from different aerosol species simultaneously, which has only been considered in a few studies (e.g., Xu et al., 2013).
Validation of 1 year of PARASOL/GRASP spectral AOD and
AAOD re-scaled to 2.0
GRASP is a recently developed aerosol retrieval algorithm that processes
properties of aerosol and land surface reflectance. The algorithm is
developed for enhanced characterization of aerosol properties from spectral,
multi-angular polarimetric remote-sensing observations
(
In this study, we adopt 1-year (December 2007 to November 2008) PARASOL
products of spectral AOD and AAOD from GRASP to retrieve DD, BC and OC
emissions over the study area (in Sect. 4). In order to evaluate the
reliability of PARASOL aerosol products from GRASP, we compared
PARASOL/GRASP retrievals with AERONET measured AOD and AAOD at four sun
photometer channels (440, 670, 870 and 1020 nm) in Fig. 3. Here, we use
level 2 AERONET version 2 data, which are cloud screened and quality assured
(Smirnov et al., 2000). From all 1-year
measurements collected from 28 sites, we extract data between 13:00 and
14:00 local time. This provides a 60 min window centred at the
PARASOL over-passing time of
Our inverse modelling approach optimizes BC, OC and DD emissions at the
2
The algorithm iteratively seeks adjustments to emissions in order to
minimize the differences between observations and simulations as quantified
by the cost function,
The minimization of the quadratic form given by Eq. (3) can be obtained with
steepest descent iterations:
Diagram illustrating retrieval of aerosol emissions from satellite measurements.
In principle, the methodology assumes that the a priori information is available,
i.e. before the inversion, which here is the default model emissions.
Unfortunately, the covariance matrix
Diagram illustrating the inversion tests from synthetic measurements.
In addition, the GEOS-Chem adjoint model has been previously used for
calculation of the gradient of Eq. (9) with respect to a vector of emissions
scaling factors
In order to optimize the specification of a priori constraints and an initial guess, a
number of synthetic tests were carried out in Sect. 3.2. It should be noted that
using an a priori estimate of emission
In this section, a series of numerical tests were performed to verify and illustrate how the algorithm inverts the synthetic measurements and to tune the algorithm settings (e.g. initial guess, emission correction time resolution and BC refractive index). The retrieved results were compared with “true emissions”. Synthetic measurements are PARASOL-like spectral AOD and AAOD at six PARASOL wavelengths, simulated from 16 days of BC, OC and DD emissions, which, for simplicity, are specified to be constant over the 16 days, yet different from the prior model emissions in order to test the algorithm performance under the circumstance that a priori knowledge of the emission distribution is limited. Figure 5 shows the design of the inversion test from synthetic measurements.
Iterative comparison of spectral AOD and AAOD residual with two spectrum weight options.
In our inversion framework, the observed aerosol parameters contain AOD and
AAOD at six PARASOL wavelengths. In principle, the weighting of observations
of AOD and AAOD at these different wavelengths should be defined as
The retrievals are conducted with Option A and Option B (the
inversion is conducted under Retrieval C scenario; see in the following
sections), with other settings held constant. Comparison of spectral
residuals after 20 iterations are shown in Fig. 6, which indicates that
Option B has a better fit for AAOD than Option A by increasing the weights
for AAOD, although spectral AOD can be fitted comparably well using either
option.
Inversion test for retrieving BC, OC and DD emissions from synthetic measurements with three different initial guess schemes: (A) prior model emissions – Retrieval A; (B) spatially uniform – Retrieval B; (C) prior emissions with spatially uniform background – Retrieval C.
In future studies, it is expected that more adequate information for
As mentioned in Sect. 3.1, the emission retrieval is an ill-posed problem
and utilization of a priori constraints and initial guesses are essential factors
for the retrieval. In our retrieval framework, the emissions are adjusted
using scaling factors for an initial guess of emissions,
Scatter plots among BC, OC and DD emissions retrieved from Retrievals A, B and C versus true values.
Thus, the following tendencies were observed in the conducted test.
In this method, the prior model emissions are directly used as the initial
guess; therefore the adjustments of emissions are limited to the grid boxes
with prior model emissions
For Retrieval A, the retrieval highly relies on the accurate distribution of
model prior emissions because the retrieval can only adjust the emissions
on the grid boxes in which the model prior emissions are non-zero, and thus the
retrieval could not create new sources. In our inversion test, the model
prior emissions are different from the truth both for distribution and
strength. Therefore, as shown in Fig. 8, Retrieval A produces
overestimations over the grid boxes for which
For Retrieval B, we investigate the use of spatially uniform initial guesses
for the emissions. With this initialization, we allow BC, OC and DD emissions
to be generated everywhere over land and ocean. In addition we are not using
a priori knowledge of aerosol emissions. From the third row “Retrieval B –
True” Fig. 7, the algorithm can determine the intensive aerosol emission grid
boxes, in which high aerosol loading is observed. However, the desert dust and
carbonaceous aerosol sources were not correctly reproduced since a uniform
emission is used everywhere. The scatter plots between retrieved emissions
from Retrieval B and true values are also shown in Fig. 8. In this case,
the retrieval could produce overestimations over some grid boxes in which
In Retrieval C, the retrieval was initiated using prior model emissions but
including a spatially uniform value over land grid boxes in which
Sensitivity test for retrieving DD, BC and OC emissions over 16 days with two scenarios of assumption of emission correction time resolution.
Aerosol sources are known to have high temporal and spatial variability.
However, because PARASOL observations have limited temporal coverage (e.g.
Figure 9 shows the retrieval maximum uncertainty
It should be noted that the regularization parameter defining the contribution of the a priori term in all tests was chosen to be very small (i.e. 0.0001) in order to make the retrieval rely mostly on the observations. Thus, the good convergence to the sought solution was obtained with minimum constraints. It is planned to investigate this aspect in future studies.
Test of BC particle refractive index influence on the
retrieval of BC emissions. The scatter plots are a grid-to-grid comparison of
retrieved 16-day averaged emissions (blue: Retrieval C; green: Retrieval E)
with the “true” BC emissions. The shaded grey area represents
Aerosol particles' light scattering and absorption efficiencies are
determined by their complex refractive indices, expressed as
The synthetic measurements of AOD and AAOD are simulated with a BC refractive
index
Overall, these sensitivity tests suggest that our inversion scheme is
capable of determining the strength and spatial distribution of BC, OC and
DD emissions simultaneously from the multispectral PARASOL/GRASP AOD and
AAOD products in the following manner.
Six wavelengths (VIS–NIR) of AOD and AAOD from PARASOL/GRASP are needed to
retrieve BC, OC and DD emissions simultaneously. The weighting spectral factors for the PARASOL six wavelengths
( The BC, OC and DD emissions are allowed everywhere over land. The retrieval
is initialized by prior model emissions with a uniform background. The
retrieval with this initialization could detect new sources and perform
satisfactorily even when a priori knowledge of aerosol emissions is not fully
consistent with the assumed emissions. This scenario will be used in
Sects. 4 and 5. The emission corrections are assumed to be daily constant for DD and 4-day
constant for BC and OC. Owing to the limited observations available for
assimilation, this assumption helps to make the retrieval sufficiently
accurate and stable with a rather generic initial guess. The BC emission retrieval is sensitive to BC refractive index assumption,
which could produce a factor of
In this section, we discuss retrieval of DD, BC and OC emissions
simultaneously from the actual PARASOL/GRASP spectral AOD and AAOD data from
December 2007 to November 2008. The SU and SS aerosol simulations are kept
as the prior model. PARASOL/GRASP retrievals were aggregated to the same
horizontal resolution as the GEOS-Chem model (2
One of the important indicators of our inversion performance is the fitting
of PARASOL/GRASP spectral AOD and AAOD. We evaluate the GEOS-Chem-simulated
spectral AOD at 443, 490, 565, 670, 865 and 1020 nm using prior or posterior
emissions against the corresponding PARASOL/GRASP-retrieved AOD in Fig. 11. The posterior GEOS-Chem spectral AODs are simulated using retrieved DD,
BC and OC emissions, which will be presented in Sect. 4.2. Figure 11a
presents the annual average of the PARASOL spectral AOD from the GRASP
algorithm, whereas Fig. 11b and c show the same quantity from the
GEOS-Chem simulations with prior and posterior emissions, respectively. Here
we extract GEOS-Chem hourly AOD with the same PARASOL orbit partition at
13:00 LT, which is approximately the PARASOL overpass time of
13:30 LT. Figure 11d and e display the grid-to-grid comparison between
PARASOL/GRASP spectral AOD and prior and posterior GEOS-Chem simulation
during 1 year, colour-coded with the PARASOL Ångström exponent
Comparison of the annual spatial distribution of prior
One of the major discrepancies between the prior GEOS-Chem simulation and PARASOL/GRASP observation is that the model produces the highest annual average AOD values over the major dust source region of northern Africa; however, satellite data show the maxima AOD in central and the southern Africa, where carbonaceous aerosols usually dominate (although central Africa may also be influenced by dust events). Hence, compared to PARASOL/GRASP observations, the prior GEOS-Chem AOD is overestimated in northern Africa, while it is underestimated in the southern Africa biomass burning and Arabian Peninsula regions. Some recent studies by Ridley et al. (2012, 2016) and Zhang et al. (2015) also indicate that the GEOS-Chem model overestimates dust AOD in northern Africa. Meanwhile, Ridley et al. (2012) and Zhang et al. (2013) propose a new and realistic dust particle size distribution according to the measurements from Highwood et al. (2003), which can partially adjust the misrepresentation of dust near the source and over transport areas. This new particle size distribution has been adopted in our prior and posterior GEOS-Chem simulation. In addition, the underestimation of model-simulated AOD in biomass burning regions with the GFED emission database was also shown in other modelling studies (Chin et al., 2009; Johnson et al., 2016). The model-simulated spectral AODs with the posterior emissions agree with the PARASOL observations much better, in spite of slight systematic overestimations from 565 to 1020 nm (about 13 % on an annual average). This overestimation indicates some disagreement in modelling of AOD for these bands that needs to be investigated and addressed in future studies.
Same as Fig. 11, but for AAOD.
Comparison of monthly total DD, BC and OC emissions
(unit: Tg) over the study area between prior model (GFED3 and Bond
inventories for BC and OC, DEAD model for DD) and retrieved emissions; the
annual values (unit: Tg yr
Figure 11d and e show the statistics of prior and posterior GEOS-Chem-simulated AOD versus PARASOL/GRASP observed AOD at six wavelengths during the entire year. The number of matched pairs is 111 493. For the GEOS-Chem simulation with the posterior emissions, all the statistics parameters between model and observation are improved at all six wavelengths compared to the simulation with prior emissions. For example, the correlation coefficient has increased from 0.49–0.51 to 0.89–0.92 and the RMSE has decreased from 0.27–0.34 to 0.10–0.13. Such improvements are expected as the posterior emissions are retrieved based on the PARASOL/GRASP AOD data. We will show further evaluations with other datasets in Sect. 5.
Similar to the AOD analysis, here we evaluate the fitting of AAOD (Fig. 12). From the annually averaged spectral AAOD in Fig. 12, the prior
GEOS-Chem simulation (Fig. 12b) shows significant underestimations of AAOD
over the entire domain compared to PARASOL/GRASP observations (Fig. 12a).
Conversely, the posterior GEOS-Chem simulation (Fig. 12c) produces
much better agreement with the PARASOL/GRASP data for all wavelengths, with
a small overestimation of AAOD in the spectral range from 443 to 565 nm
(about 6 % on annual average) and a small underestimation at 865 and
1020 nm (about 9 % on annual average). Linked with the
Figure 12d and e show the comparisons of PARASOL/GRASP-observed AAOD at six wavelengths with the corresponding GEOS-Chem-simulated quantities using prior or posterior emissions. The very low linear regression slope between the model-simulated AAOD using prior emissions with observations (less than 0.11 over all six wavelengths) indicates that the prior simulations significantly underestimate the AAOD. In contrast, model simulations with the posterior emissions improve the slope to 1.01 at 443 nm and 0.70 at 1020 nm. Similar to the case of AOD, the agreements between the PARASOL/GRASP AAOD data and the model simulations are much better using the posterior emissions than using the prior emissions, with the correlation coefficients increased from 0.14–0.33 to 0.90–0.92 and the RMSE decreased from 0.022–0.044 to 0.008–0.023.
The retrieved and prior monthly total DD, BC and OC emission variations over the study area are shown in Fig. 13.
Figure 13 shows that the retrieved annual total DD emissions in the study
area is 701 Tg yr
Spatial distribution of seasonal desert dust aerosol
emissions:
As mentioned earlier, we considered two cases of BC aerosol refractive index
to perform the retrieval (Case 1:
Spatial distribution of seasonal BC emissions:
The spatial comparison of seasonal BC emissions is summarized in Fig. 15.
We plot model prior BC emissions from GFED3 and Bond anthropogenic
inventories in Fig. 15a, retrieved BC emissions from Case 1 in Fig. 15b,
and Case 2 retrieved BC emissions in Fig. 15c. Note that the colour bar
range in Fig. 15b is 2.5 times larger than that of Fig. 15a and Fig. 15c. Not surprisingly, the patterns of model prior emissions in Case 1 and
Case 2 retrievals are similar, with the highest BC emission source areas
located in biomass burning regions, such as central Africa during DJF and
southern Africa JJA. The large increases in the BC emissions in the
retrieval relative to the prior suggest that the current model-simulated
AAOD is much too low, which is consistent with the PARASOL/GRASP
observations in Sect. 4.2. Retrieval Case 2 shows a large increase over
the Arabian Peninsula, indicating there is an emission
The annual total OC emissions in Fig. 13 show that the retrieved annual
OC emissions are higher than the prior model by a factor of
Spatial distribution of seasonal OC emissions:
Comparison of retrieved DD, BC and OC aerosol emissions over the study area
with the GEOS-Chem prior model emission inventories showing basically
consistent spatial and temporal variation. However, the significant
differences are in the emission strength. The PARASOL/GRASP-based retrieval
reduces the GEOS-Chem annual DD emissions to 701 Tg yr
Density scatter plots of 1-year GEOS-Chem-simulated
AOD using the prior emissions
Same as Fig. 17, but for AAOD. The AAOD data were
aggregated into 50 bins for both the
In order to objectively evaluate our retrieved aerosol emissions based on PARASOL/GRASP spectral AOD and AAOD, we made a series of evaluations using independent datasets and models not used by our inversion. First, the posterior simulated 1-year AOD and AAOD (using Case 1 BC emissions) are compared with the sun-photometer-measured AOD and AAOD at 28 AERONET sites (shown in Fig. 1).
Figures 17 and 18 show the comparison of GEOS-Chem simulations using prior
and posterior emissions with AERONET measurements of AOD and AAOD,
respectively. The evaluation was conducted at four wavelengths (440, 675, 870
and 1020 nm) and GEOS-Chem hourly spectral AOD and AAOD are interpolated
based on the Ångström exponent. AERONET AOD and AAOD output averaged over time for
Time serial AOD
Comparison of the seasonal spatial distribution of prior
Figure 18 shows the density scatter plot comparisons for AAOD. However, unlike direct sun measurement of AOD, AERONET AAOD is inverted from almucantar measurements. To select sufficiently accurate retrievals, we applied standard quality screening criteria (e.g. Dubovik et al., 2002b and Holben et al., 2006). Therefore, there are fewer AERONET AAODs that matched with GEOS-Chem simulations than for AOD. The number of matched pairs is 3728. The low slope of the linear regression between prior model AAOD and AERONET (Fig. 18a) indicates that the prior model significantly underestimates AAOD. The posterior GEOS-Chem simulations using retrieved emissions (Fig. 18b) show the improvements validated with AERONET, with the correlation coefficients at 0.71, 0.64, 0.59 and 0.53. In addition, the RMSEs are also improved for posterior simulations.
Comparison between time series of AOD and AAOD at 440 nm from AERONET,
PARASOL/GRASP, and prior and posterior GEOS-Chem simulations from December
2007 to November 2008 are made at two AERONET sites (Mongu and Ilorin), and
the results are shown in Fig. 19. The geo-locations of these two sites are
already apparent in Fig. 1. Ilorin is located close to the active dust
sources in northern Africa, which are also influenced by seasonal
biomass burning events, especially from November to February. Mongu is
located close to the southern African seasonal biomass burning sources. The
posterior simulations better capture the time series variations in and
magnitude of AOD and AAOD from AERONET measurements. For example, in Mongu,
the prior simulation underestimates AOD and AAOD significantly. In
September, the underestimations are about 3 times (a bias of
Comparison of the seasonal spatial distribution of prior
All the evaluations considered thus far are based on simulations in the GEOS-Chem model. To evaluate how such results may be impacted by model biases owing to factors other than BC, OC and DD emissions, here we ask – can aerosol emissions retrieved from the GEOS-Chem-based inversion improve the aerosol simulation for another CTM? To investigate this, we implement our PARASOL/GRASP-based aerosol emission database in the GEOS-5/GOCART model (Chin et al., 2002, 2009, 2014; Colarco et al., 2010). The prior and posterior GEOS-5/GOCART model-simulated seasonal AODs are compared with MODIS observations in Fig. 20. GEOS-5/GOCART uses similar meteorological fields as GEOS-Chem, with the prior anthropogenic emissions from the Hemispheric Transport of Atmospheric Pollution (HTAP) Phase 2, biomass burning emissions from the Fire Energetics and Emission Research (FEER) database (Ichoku and Ellison, 2014), dust emissions calculated as a function of 10 m winds and surface characteristics (Ginoux et al., 2001), and volcanic emissions from OMI-based estimates (Carn et al., 2015). The PARASOL/GRASP-retrieved DD, BC and OC emissions over the study domain are used in the posterior simulations while other sources remain unchanged. On an annual average, the DD, BC and OC posterior–prior emission ratios in the study area are 0.53, 5.3 and 1.2 respectively.
Figure 20a shows the MODIS seasonal AOD at 550 nm. In order to have better
spatial coverage, we take MODIS collection 6 combined dark target and deep
blue AOD products at the spatial resolution of 1
Commonly, the ultraviolet, shortwave visible channels and polarimeter
measurements are considered to be main observation types sensitive to aerosol
absorption properties; therefore, AERONET, PARASOL/GRASP and OMI datasets
are often used as major long-term records of AAOD. We use the latest OMI
aerosol products (OMAERUV version 1.7.4) (Torres et
al., 2007, 2013) to evaluate the GEOS-5/GOCART model-simulated AAOD from
prior aerosol emission inventories and our retrieved aerosol emission
database. Meanwhile, collocated AERONET data over the study area are also
employed for the evaluation. Detailed assessments of OMI aerosol products are
described in other studies (Torres et al.,
2013; Ahn et al., 2014; Jethva et al., 2014). Figure 21 shows the validation
results. Figure 21a presents the OMI seasonal mean AAOD with original
OMAERUV version 1.7.4 spatial resolution 0.5
The major discrepancy between OMI seasonal AAOD (Fig. 21a) and the prior GEOS-5/GOCART-simulated AAOD is that the simulated AAOD is higher than OMI values in the northern African dust regions over all seasons, which can be attributed to the overestimation of dust particle absorption (Chin et al., 2009) and/or the total dust emissions. The posterior GEOS-5/GOCART-simulated AAOD shows a similar spatial distribution and magnitude to OMI values over dust regions with reduced differences, although the model is still overall higher than OMI, in particular over the southern African biomass regions in JJA.
As shown in Fig. 21a, the correlation coefficients of OMI seasonal AAOD
with AERONET vary from 0.42 in JJA to 0.83 in SON; meanwhile the RMSE is smallest in MAM
From the scatter plot of GEOS-5/GOCART-simulated AAOD versus OMI AAOD in
Fig. 21d–e, the significant increase in correlation coefficient from prior
to posterior simulations occurs in June–July–August (prior: 0.54; posterior:
0.76) as well as decreases in RMSE and MAE (prior: RMSE
In this study, we designed a method to retrieve BC, OC and DD aerosol
emissions simultaneously from satellite-observed spectral AOD and AAOD based
on the PARASOL/GRASP retrievals and the adjoint of the GEOS-Chem CTM. This method uses prior BC, OC and DD emissions as weak
constraints in the inversion by initializing the retrieval with prior
emissions added to uniform background values. A series of numerical tests
were performed, which show this assumption can provide a better fit to
observations, meanwhile this assumption allows the retrieval to produce rather good
results even if a priori knowledge of emissions is poor. Admittedly, the satellite
observations are sparse due to several factors, e.g., the clear-sky
conditions, global coverage orbit cycle. Nevertheless, the PARASOL 6
wavelengths AOD and AAOD from the GRASP algorithm are shown to be sufficient to
characterize the distribution and magnitude of BC, OC and DD aerosol
emissions simultaneously under the assumption of a DD emission correction
constant over 24 h and a 4-day correction constant for carbonaceous aerosol
emissions. The inversion test of synthetic PARASOL-like measurements results
in about 25.8 % uncertainty for daily total DD emissions, 5.9 % for daily
total BC emissions and 26.9 % for daily total OC emissions. In addition, it
was shown that using two different assumptions for BC refractive index (Case 1:
We evaluated the GRASP-retrieved 1-year PARASOL spectral AOD and AAOD with
AERONET ground-based observations and retrievals at 28 sites across the study
area (30
Our analysis of the retrieved aerosol emissions indicates that the prior GEOS-Chem model overestimates annual desert dust aerosol emissions by a factor of about 1.8 (with the DEAD scheme) over the study area, similar to other previous modelling studies (Huneeus et al., 2012; Johnson et al., 2012; Ridley et al., 2012, 2016). The retrieved annual BC and OC emissions show a consistent seasonal variation with emission inventories (GFED3 for biomass burning and Bond for anthropogenic fossil fuel and biofuel combustions). However, we find these BC and OC emissions to have broad underestimations throughout the study area. For example, emissions from the emission inventories for BC are significantly lower than our retrieved values by up to factor of 8 (Case 1) and 3 (Case 2), and for OC they are about a factor of 2 lower. These results are reflected in the model bias of AOD and AAOD from the prior GEOS-Chem simulation, e.g. significantly low bias over the biomass burning regions and high bias over the Sahara desert. Underestimation of BC and OC emissions in CTMs has been suggested previously (Sato et al., 2003; Zhang et al., 2015). However, we cannot rule out the possibility that differences between model and observations could also be attributed to the errors in removal processes and aerosol microphysical properties, in addition to the deficiencies in emissions (Bond et al., 2013). Nevertheless, the fidelity of our results is confirmed by comparison of posterior simulations with measurements from AERONET that are completely independent from and more temporally frequent than PARASOL observations. Specifically, to analyse the PARASOL/GRASP-based aerosol emission database further, we implemented these emissions in the GEOS-5/GOCART model and compared the resulting simulations of AOD and AAOD with independent MODIS and OMI observations. The comparisons show better agreement between model and observations with the posterior GEOS-5/GOCART results (lower biases and higher correlation coefficients) than prior simulations. In the future, we plan to apply our approach globally to longer records of observations to further investigate the inter-annual variability in aerosol emissions on global scales and to test our retrieved emission database in other models.
PARASOL/GRASP aerosol data are available from the ICARE data distribution
portal:
CC, OD, DKH and TL contributed to retrieval algorithm development and conducted the inversion test using synthetic measurements. FD, PL, HX and LL contributed to POLDER/PARASOL aerosol product generation by the GRASP algorithm. MC contributed to test-retrieved emissions in the GEOS-5/GOCART model as well as evaluation with independent measurements. CC and OD wrote the paper with input from all authors.
The authors declare that they have no conflict of interest.
This work is supported by the Laboratory of Excellence CaPPA – Chemical and Physical Properties of the Atmosphere – project, which is funded by the French National Research Agency (ANR). We would like to thank the GEOS-Chem and adjoint GEOS-Chem model developers; Daven K. Henze recognizes support from NASA ACMAP NNX17AF63G. We also thank the entire AERONET team, and especially the principal investigators and site managers of the 28 AERONET stations that we acquired data from. The authors are also grateful for the MODIS and OMI aerosol team (Omar Torres and Hiren Jethva) for providing the data used in this investigation and Huisheng Bian and Tom Kucsera for incorporating the PARASOL/GRASP emissions into the GEOS-5 model and provide the GEOS-5/GOCART simulation results. Edited by: Yves Balkanski Reviewed by: two anonymous referees