ACPAtmospheric Chemistry and PhysicsACPAtmos. Chem. Phys.1680-7324Copernicus PublicationsGöttingen, Germany10.5194/acp-16-9655-2016Technical note: Intercomparison of three AATSR Level 2 (L2) AOD products over
ChinaCheYahuiXueYongyx9@hotmail.comhttps://orcid.org/0000-0003-3091-6637MeiLinluGuangJiehttps://orcid.org/0000-0003-4630-1645SheLuhttps://orcid.org/0000-0002-5413-9150GuoJianpinghttps://orcid.org/0000-0001-8530-8976HuYincuiXuHuiHeXingweiDiAojieFanChengState Key Laboratory of Remote Sensing Science, jointly sponsored by
the Institute of Remote Sensing and Digital Earth of the Chinese Academy of
Sciences and Beijing Normal University, Institute of Remote Sensing and
Digital Earth, Chinese Academy of Sciences, 100101 Beijing, ChinaDepartment of Computing and Mathematics, College of Engineering and
Technology, University of Derby, Kedleston Road, Derby, DE22 1GB, UKKey Laboratory of Digital Earth Science, Institute of Remote Sensing
and Digital Earth, Chinese Academy of Sciences, 100094 Beijing, ChinaCentre for Atmosphere Watch and Services, Chinese Academy of
Meteorological Sciences, 46 Zhongguancun South Avenue, Haidian District,
100081 Beijing, ChinaHebei Key Laboratory of Environmental Change and Ecological
Construction, College of Resources and Environment Science, Hebei Normal
University, Shijiazhuang, Hebei Province, ChinaUniversity of Chinese Academy of Sciences, 100049 Beijing, ChinaYong Xue (yx9@hotmail.com)2August20161615965596745March201616March201629June20163July2016This work is licensed under a Creative Commons Attribution 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/This article is available from https://acp.copernicus.org/articles/16/9655/2016/acp-16-9655-2016.htmlThe full text article is available as a PDF file from https://acp.copernicus.org/articles/16/9655/2016/acp-16-9655-2016.pdf
One of four main focus areas of the PEEX
initiative is to establish and sustain long-term, continuous, and
comprehensive ground-based, airborne, and seaborne observation infrastructure
together with satellite data. The Advanced Along-Track Scanning Radiometer
(AATSR) aboard ENVISAT is used to observe the Earth in dual view. The AATSR
data can be used to retrieve aerosol optical depth (AOD) over both land and
ocean, which is an important parameter in the characterization of aerosol
properties. In recent years, aerosol retrieval algorithms have been developed
both over land and ocean, taking advantage of the features of dual view,
which can help eliminate the contribution of Earth's surface to
top-of-atmosphere (TOA) reflectance. The Aerosol_cci project, as a part of
the Climate Change Initiative (CCI), provides users with three AOD retrieval
algorithms for AATSR data, including the Swansea algorithm (SU), the
ATSR-2ATSR dual-view aerosol retrieval algorithm (ADV), and the Oxford-RAL
Retrieval of Aerosol and Cloud algorithm (ORAC). The validation team of the
Aerosol-CCI project has validated AOD (both Level 2 and Level 3 products) and
AE (Ångström Exponent) (Level 2 product only) against the AERONET data in a round-robin
evaluation using the validation tool of the AeroCOM (Aerosol Comparison
between Observations and Models) project. For the purpose of evaluating
different performances of these three algorithms in calculating AODs over
mainland China, we introduce ground-based data from CARSNET (China Aerosol
Remote Sensing Network), which was designed for aerosol observations in
China. Because China is vast in territory and has great differences in terms
of land surfaces, the combination of the AERONET and CARSNET data can
validate the L2 AOD products more comprehensively. The validation results
show different performances of these products in 2007, 2008, and 2010. The SU
algorithm performs very well over sites with different surface conditions in
mainland China from March to October, but it slightly underestimates AOD over
barren or sparsely vegetated surfaces in western China, with mean bias error
(MBE) ranging from 0.05 to 0.10. The ADV product has the same precision with
a low root mean square error (RMSE) smaller than 0.2 over most sites and the
same error distribution as the SU product. The main limits of the ADV
algorithm are underestimation and applicability; underestimation is
particularly obvious over the sites of Datong, Lanzhou, and Urumchi, where
the dominant land cover is grassland, with an MBE larger than 0.2, and the
main aerosol sources are coal combustion and dust. The ORAC algorithm has the
ability to retrieve AOD at different ranges, including high AOD (larger than
1.0); however, the stability deceases significantly with increasing AOD,
especially when AOD > 1.0. In addition, the ORAC product is
consistent with the CARSNET product in winter (December, January, and
February), whereas other validation results lack matches during winter.
Introduction
The Pan-Eurasian Experiment (PEEX) is a multidisciplinary, multiscale, and
multicomponent research, research infrastructure, and capacity-building
program (Kulmala et al., 2015). One of the strategically most important tasks of PEEX is to fill in the observational gap in atmospheric in situ data in
the Siberian and Far East region and start the process towards standardized
and harmonized data procedures (Kulmala et al., 2011). Aerosols play a major
role in Earth's climate system, including intervening in the radiation budget
and cloud processes and affecting air quality and human health (Remer et
al., 2005; Samet et al., 2000; Tzanis and Varotsos, 2008; Kokhanovsky and de
Leeuw, 2009). The particles suspended in the troposphere scatter solar
radiation back to cool the atmosphere or absorb solar radiation, which warms
the atmosphere, causing changes in the net effect of aerosols. These
particles could also affect the formation and microphysical properties of
clouds as cloud condensation nuclei (Andreae and Rosenfeld, 2008). The source
of aerosols could be anthropogenic or natural (Varotsos et al., 2012).
Particles from different sources are mixed into aerosol masses to influence
aerosol optical depth (AOD), reduce visibility (Kinne et al., 2003; Varotsos 2005; Remer et al.,
2005), and cause spatial and temporal variability of AOD; therefore, the
largest uncertainties in the estimation of radiative forcing are introduced
by aerosols (IPCC, 2013).
Over the past 35 years, different types of satellites have been used to
obtain atmospheric information, especially aerosol properties (Griggs, 1979;
Kokhanovsky and de Leeuw, 2009). Remote sensing provides a means to obtain
global and long-term observations of aerosols, especially in the widest
oceans and remote regions where ground-based stations cannot be constructed. In
addition, polar-orbiting satellites and geostationary satellites obtain daily
global images, which help to capture changes in aerosol patterns and
properties (Prins et al., 1998; Torres et al., 2002). There are, however,
many difficulties in observing aerosols by satellites because depending on
the surface properties, the contribution to the signal received by the
satellite can vary drastically; aerosol components and concentrations are
constantly varying, and their sources cannot be precisely determined (Levy et
al., 2007).
The Advanced Along-Track Scanning Radiometer (AATSR) aboard ENVISAT is used
to observe the Earth in dual view, of which one component is nadir direction and the
other is forward direction with a viewing angle of 55∘ from nadir
view. The AATSR was designed to have seven spectral channels at wavelengths of
0.55, 0.67, 0.87, 1.63, 10.7, and 12 µm. The nadir spatial resolution
is 1 km × 1 km with a swath width of 512 pixels. Furthermore, the
AATSR instrument equipped two calibration targets: a black-body calibration
target for thermal channels and an opal visible calibration target for visible
and near-Infrared channels, aiming to implement self-calibration. The data
from AATSR can be used to retrieve AOD both over land and ocean, which is
important for the characterization of aerosol properties (Adhikary et al.,
2008). In recent years, some aerosol retrieval algorithms have been
established both over land and ocean, taking advantage of the features of
dual view, which can help eliminate the contribution of surface to top-of-atmosphere (TOA) reflectance. Aerosol_CCI, as part of the Climate Change
Initiative (CCI) (http://www.esa-aerosol-cci.org/), provides users with
three algorithms for AATSR data, including the Swansea algorithm (SU) (Bevan
et al., 2012), the ATSR-2/AATSR dual-view aerosol retrieval algorithm (ADV)
(Kolmonen et al., 2015), and the Oxford-RAL Retrieval of Aerosol and Cloud
algorithm (ORAC) (Thomas et al., 2009). The aim of this work is to evaluate
different performances of these algorithms in calculating AOD over different
regions of China in 2007, 2008, and 2010.
A ground-based sun photometer has been used to take sun and sky measurements
directly (Holben et al., 1998). The Aerosol Robotic NETwork (AERONET) has
constructed hundreds of sites all over the world as of 2015. These stations,
operated by the American National Aeronautics and Space Administration
(NASA), are operational worldwide, providing multispectral channel
validation data for satellite-retrieved data to complete synthetic
measurements on a global scale.
The China Aerosol Remote Sensing Network (CARSNET) is a ground-based aerosol
monitoring system that uses CE-318 sun photometers, similar to AERONET, and
has constructed 37 sites throughout China (Che et al., 2009). It has been
established that CARSNET AOD measurements are approximately 0.03, 0.01, 0.01, and 0.01 larger than measurements by AERONET in the 1020, 870, 670, and
440 nm channels, respectively (Che et al., 2009). In this paper, we combine
two aerosol observation datasets from AERONET and CARSNET as reference data
to validate these three AATSR AOD products over China more comprehensively.
The basic method for assessment is to compare the retrieval results with data
(AOD mainly) obtained by AERONET/CARSNET. However, this direct comparison of
retrieval results with AERONET data is limited due to different cloud
screening processes (de Leeuw et al., 2013), and such a limitation could
influence the validation reliability to some extent. To make the validation
more reliable, comparison of the retrieval results with high-quality data
from the Moderate Resolution Imaging Spectroradiometer (MODIS) or Multiangle Imaging Spectroradiometer (MISR) is also an effective method for validation (Kahn et al.,
2009). However, AERONET or other ground-based networks provide accurate
measurements without the influence of land surface reflection (Holben et al.,
1998), which means that comparison of retrieved AOD with ground-based
measurements is the basic method. The AATSR L2 products provided by
Aerosol_CCI have been validated by the validation team via a round-robin
(RR) test (de Leeuw et al., 2013). On this basis, we focused on assessing the
performance of AATSR aerosol L2 products in mainland China by comparing the
retrieval results with AERONET and CARSNET data.
Reference data and validation statistics
AOD is the most important parameter in terms of aerosol properties and is
different from other retrieved parameters in the Aerosol_CCI project.
The Aerosol_CCI project adopts three aerosol retrieval algorithms for the ATSR-2/AATSR instrument, including the Swansea algorithm (SU) (Bevan et al.,
2012), the ATSR-2/AATSR dual-view aerosol retrieval algorithm (ADV) (Kolmonen
et al., 2015), and the Oxford-RAL Retrieval of Aerosol and Cloud algorithm
(ORAC) (Thomas et al., 2009). All three algorithms are able to retrieve aerosol properties both over land and ocean. The ADV algorithm was
originally developed for retrieving AOD properties over land at wavelengths of
0.555, 0.659, and 1.61 µm (Veefkind et al., 1998). The main advantage
of ADV is the introduction of a k-ratio approach to eliminate the contribution of
reflection to TOA reflectance, which uses the ratio of the reflectance
measured in the forward and nadir views (Flowerdew and Haigh, 1995). The ORAC
algorithm is designed to retrieve AOD properties in each of four AATSR
short-wave channels both over land and ocean, including AOD, effective
radius, and surface reflectance. The build of the forward model used in the ORAC
algorithm is based on the radiative transfer code DISORT. A parameterized model
of surface reflectance distribution is used in retrieval and combines with
the AATSR dual view to make up for the lack of a priori reflectance (North et al., 1999). An iterative optimization method is
employed to determinate AOD, aerosol type, and surface reflectance.
The distribution of selected AERONET and CARSNET sites in mainland
China in 2007, 2008, and 2010. The blue and red points represent AERONET and
CARNET sites, respectively.
AATSR L2 data (see Table 1) are daily products with a spatial resolution of
10 × 10 km2, and contain a quality flag or a level of
confidence for each pixel (de Leeuw et al., 2013). Compared to the Level 3
(L3) product with a spatial resolution of 1∘× 1∘,
daily L2 data have a higher spatial resolution, which helps to capture
greater detail of aerosol properties and is further explored in our follow-up
study/research after this manuscript.
Details of AATSR AOD products.
AlgorithmVersionSensorMain parametersResolution coverageADV/ASV2.3AATSRAOD,ANG10 km, 1∘ globalSU4.21AATSRAOD, ANG10 km, 1∘ globalORAC03.04AATSRAOD, aerosol type10 km, 1∘ global
It has been demonstrated that the ground-based observation data from AERONET have the ability and precision to be used as reference data when
users validate AOD (Holben et al., 1998). There are 8 AERONET sites in
mainland China providing Level 2.0 (L2) data (cloud-screened and
quality-assured) for 2007, 11 sites for 2008, and 10 sites for 2010, from
which the AOD measurement data are available on the website. However, most of
these sites are distributed in the eastern China coastal area, as shown in
Fig. 1, which, however, does not meet the requirements of comprehensively
validating the aerosol properties over all of China. Substantial hazardous
aerosol pollution affects most regions of northern (Li, 2014) and eastern
China in winter, and heavy dust aerosols from the Taklimakan Desert in
western China can be transported long distances to eastern China, even to
Japan (Takahashi, 2011), resulting in regional differences.
The measurements from another network, CARSNET, equipped with calibrated
CE-318 instruments, have the same accuracy as AERONET. CARSNET has more
sites than AERONET in mainland China, and the spatial distribution of
the CARSNET sites is distributed more evenly. Therefore, for the purpose of
assessing different performances of these three AATSR L2 AOD products, we
selected ground-based measurements from both of these two networks as
reference data.
AERONET provides AOD data at three data quality levels: Level 1.0
(unscreened), Level 1.5 (cloud-screened), and Level 2.0 (cloud-screened
and quality-assured) (http://aeronet.gsfc.nasa.gov/new_web/index.html).
Here, we selected AERONET L2 data that are screened and quality-assured.
Because both the AERONET and CARSNET data are AATSR products without
band-effective wavelengths, we interpolated the ground-based data to the
550 nm wavelength. The AOD of the L2 datasets were compared with
AERONET and CARSNET observation data using scatter plots and linear regression
of the data. The comparisons were made for collocated satellite and
ground-based observations (Ichoku et al., 2002); i.e. AOD pixels were
selected from the ground-based
measurements within a spatial extent of ±25 km of ground-based stations and
a time range of ±30 min of the AATSR overpass. At least five AATSR AOD retrievals and two AERONET or CARSNET
observations are required in each collocation (Levy et al., 2010).
We conducted collocations according to year (2007, 2008, and 2010) and dataset
(ADV, ORAC, and SU). In total, 20 ground-based observation sites, including 12
AERONET sites and 8 CARSNET sites, were in Chinese territory in 2007, of
which 6 AERONET and 8 CARSNET inland sites were selected. For 2008, we
selected 8 AERONET and 24 CARSNET inland sites, for a total of 32 sites,
ignoring the island sites and those near the shoreline. For 2010, only 6 CARSNET sites are available for us, and a total of 14 inland sites
was selected, with 8 AERONET inland sites (see Table 2).
Collocated pairs are analysed using statistical methods. Bias describes the
average difference between satellite retrievals and ground AOD. Then, to
determine how well the satellite data match the ground-based observation
data, the relationship between them is explored. Some basic statistics are
shown on the scatter plot, including the root mean square error (RMSE):
RMSE=1n∑i=1nτsat,i-τaero,i2,
where τaero,i represents the ground-based observation data
and τsat represents the satellite retrievals. Mean
satellite-retrieved AOD (MSA) and mean AERONET and CARSNET AOD (MAA)
represent the central tendency of the data. Relative mean bias (RMB) is used
to determine under- or overestimation of the AOD retrievals; it is the ratio
of MSA to MAA:
RMB = MSA/MAA.
Mean bias error (MBE) is the mean difference between the satellite
retrievals and AATSR AODs, and the mean absolute error (MAE) is the absolute
value of the mean bias error. Together with RMB, the MBE and MAE are used to
determine the magnitude of the difference between the two datasets.
κ statistics
In the scatter plot of the collocated pairs, the retrieved data and the
corresponding collocated ground-based observation data could be considered
as two arrays, and the main purpose of KAPPA is to explore how these two
arrays match each other. For the retrieval of aerosol properties, the
performances of most algorithms decrease in effectiveness with increasing
AOD, i.e. difficulties in retrieving AOD will be increased as AOD
increases. Obviously, when only using bias, the absolute
value of the difference between ground-based data and AATSR AOD data in each
collocation pair, as an assessment standard for different AODs, is
insufficient and lacks persuasion. Therefore, the combination of bias and bias/Ground, i.e. the ratio of bias to the value of the reference data in each collocation pair,
used in the KAPPA coefficient will account for this shortage and provide a
new statistic for assessing the agreement between two arrays, taking
advantage of the KAPPA coefficient.
The KAPPA coefficient was originally proposed as a descriptive statistic
indicating the degree of beyond-chance agreement between two ratings per
subject in a dichotomous form (Bloch and Kraaemer, 1989). KAPPA coefficients
with various forms could also be used to measure the accuracy of thematic
classifications (Rosenfield and Fitzpatrick-Lins, 1986). KAPPA is, in short,
a measure of “true” agreement (Cohen, 1960). The pairs collocated by
matching ground-based data with AATSR L2 AOD data could be regarded as two
different arrays so that we introduced the KAPPA coefficient to assess
agreement between these two arrays. Based on the concept of the KAPPA
coefficient proposed by Cohen (1960), an appropriate modification with a
two-category nominal scale is shown in Table 3.
To estimate the KAPPA coefficient, one needs to determine which pairs are
true or which pairs are “relevant”. However, if only given matched
collocation pairs, we cannot determine which pair is relevant. Therefore,
the design of criterion 1 and criterion 2 needs to be reasonable and fit for
the purpose of validation.
For criterion 1, if bias is greater than the mean of
bias, then it is marked as “far from truth”, and if not,
it is marked as “close to truth”. Here, the bias was assessed from the
first quartile to the third quartile for eliminating possible “outliers”.
The bias only indicates the absolute error of the retrieved
AOD, and it still needs another statistic for criterion 2, i.e. bias/Ground, which indicates the relative error of AOD retrieval.
For criterion 2, if bias/Ground| is greater than 0.2, then it is marked as “far from truth”, and if not, it
is marked as “close to truth”. For the conventional formula of calculating
the KAPPA coefficient, see
K=Po-Pc1-Pc,
where Po is the proportion of observed agreement and P is the proportion
of chance agreement.
Po=a+dnPc=F1×G1n+F2×G2nn
Algorithms for AATSR AOD retrieval used to underestimate AOD over different
regions in China include the ADV, ORAC, and SU algorithms. On this basis, the
agreement between ground-based observation data and satellite retrievals is
assessed based on the ADV and SU algorithms (Che et al., 2015). The main aim
of this new KAPPA coefficient is to evaluate the comprehensive performance of
these algorithms. Its function is to represent not only the degree of
underestimation but also the level of agreement between different datasets.
Validation results and analysis
We collected different validation reference data of AERONET and CARSNET in
2007, 2008, and 2010. Only 14 ground-based observation sites are available in
2007, of which some are located close to each other. Most are located in
different provinces; however, the total number of sites is small and the
space distribution is not uniform. Therefore, the number of matches is
relatively small for all of the algorithms. More AERONET and CARSNET data are
available in 2008, with a total of 32 sites, made up of 8 AERONET sites and 24
CARSNET sites. There are 14 AERONET and CARSNET sites providing data for validation
in 2010. The focus of this paper is to determine the differences between the
ADV, ORAC, and SU L2 AOD products (see Table 4 and Fig. 1). Therefore, we
calculated statistics and analysed the validation results separately by year
(see Table 4 and Fig. 1).
For 2007, the RMSE is 0.095 and the RMB is 0.704, which reflects the
tendency of underestimation. This type of underestimation is more severe
with increasing AOD. Low dispersion and slight underestimation make the
KAPPA coefficient high (0.473), demonstrating that the ADV algorithm
performs well in calculating the AOD over China in 2007. The ADV algorithm
is appropriate for the retrieval of low AODs, especially for those less than
1.0; thus, the MSA for 2007 is 0.244. For 2008, the lower RMB (0.621)
suggests more severe underestimation, and higher relative mean bias (0.130) indicates lower
accuracy. Similar to 2007, the MSA of the ADV is 0.211. Therefore, the
KAPPA coefficient, which measures the overall performance, is 0.329, lower
than that of 2007. For 2010, ADV/AOD has the lowest RMSE (0.089) with the lowest
accidental error of the 3 years. However, the KAPPA coefficient is
0.180, also the lowest of the 3 years. The most obvious feature of the
ADV algorithm is underestimation with the highest MSA being 0.250 in 2007 and
the lowest being 0.173 in 2010. The ADV algorithm can retrieve low AOD values
with high accuracy. This “ability” is systematic for either high AODs or
low AODs. This also limits the range of application of the ADV algorithm,
especially in calculating AODs in high-value ranges.
Scatter plots of AATSR ADV, ORAC, and SU L2 AOD products with
ground-based data in China for the 3 years of 2007, 2008, and 2010. The black
solid line represents the 1–1 line. The magenta points are means for specific
ranges of AERONET and CARSNET AOD, and the magenta lines are the mean ±2σ of retrievals in a certain range. The areas and colours are
determined by the means of uncertainty (MU) dataset in AATSR L2 products and
the standard deviation of retrievals (Std_S) in a collocation frame of
50 km × 50 km, respectively.
The ORAC algorithm
The ORAC algorithm performed well for 2007, achieving a KAPPA coefficient of
0.474. However, the distribution of matches is dispersed, implying high RMSE
(0.206). In terms of the degree of fitness, its performance is not
effective. However, there is no obvious trend of underestimation or
overestimation, and accidental errors influence the accuracy of the ORAC
algorithm. The MSA of the ORAC is 0.324. ORAC has the most matches of the
three algorithms. Differently from 2008, no obvious underestimation occurs in
the results of 2007 and 2010. For 2008, the RMB is 0.829, suggesting a
slight underestimation trend. The applicability of ORAC is high, with MSA of
0.271. The collocated pairs are relatively dispersed, influencing the RMSE.
For 2010, the same dispersion of points in the scatter plot and low KAPPA
coefficient are observed. Overall, the ORAC algorithm tends to retrieve AODs
unstably for either high AODs or low AODs and with a slight underestimation in
2007. The results of 2008 and 2010 share common features, indicating that
accidental error is larger than systematic error.
The SU algorithm
The SU algorithm performed well for all 3 years, achieving KAPPA
coefficients of 0.409, 0.484, and 0.520, respectively. The RMBs are 0.816,
0.713, and 0.720 for 2007, 2008, and 2010, respectively, demonstrating the
underestimation of the SU product. The applicability of SU is high, with an MSA
of 0.293 for 2008. The most obvious feature of the SU algorithm is its
stability in retrieving AOD for different years or different regions (Fig.
4). The MSA ranges from 0.270 for 2010 to 0.330 for 2007, and the KAPPA
coefficient ranges from 0.520 to 0.409, which suggests that the SU algorithm
performed better at retrieving low AODs. The SU algorithm has the best
performance in terms of AOD retrieval, as it has the highest KAPPA
coefficient (0.520). Overall, the SU algorithm can be applied to retrieve
AOD in different ranges with high precision. Factors influencing the
performance of the SU algorithm include a small systematic error and even
a smaller accidental error.
Uncertainty analysis based on aerosol loading
In the previous section, we validated all three AOD products over mainland
China in 2007, 2008, and 2010, discovering that all three products tend to
exhibit underestimation to some extent. For the purpose of ascertaining the
causes of the underestimation, in this section, we focus on analysing the AOD
uncertainties leading to differences between retrieved AODs and ground-based
AODs in special conditions. Collocated pairs are divided into three groups
according to aerosol loading, including light loading (τ<0.15), heavy loading (τ>0,4), and moderate loading (Levy et
al., 2010). It is obvious that the AOD bias increases with increasing AOD for
all three products. These products have one feature in common, that is, the
AOD bias tends to be negative, which indicates that the underestimation
becomes more significant with increasing aerosol loading. The ADV and SU
algorithms perform well in estimating AOD, i.e. with little underestimation
(lower MBEs of -0.04 and -0.02, respectively, as shown in Table 5), when aerosol
loading is low (light loading) (Fig. 3).
Scatter plot of AERONET and CARSNET AODs with ADV AOD bias or
uncertainties in China in 2007, 2008, and 2010. The areas and colours of
bubbles represent MU and Std_S sampling area of 50 km × 50 km, respectively. Colours represent different groups: blue
denotes light loading, green denotes moderate loading, and red denotes heavy
loading. Each group has one box, the bottom and top borders of which
represent MBE +2σ and
MBE -2σ, respectively, containing 96 % of
scattered points of each group. The centre line of each box represents the
MBE of each group.
Under complex conditions, the ORAC overestimates AOD in regions of light
loading and moderate loading compared with AERONET and CARSNET, as shown
in Fig. 3. ADV tends to underestimate AOD more severely, with MBE =-0.11
in a moderate aerosol loading region. Similarly to ADV, the underestimation of SU
in moderate aerosol loading becomes more severe, with MBE =-0.07. ORAC
performs the best in retrieving in a moderate aerosol loading regions without
underestimation or overestimation, even though the bubbles are distributed
discretely with an SD of 0.18 (see in Table 5). The performances of all three
algorithms are at the same level, with close MBEs, SDs, and RMSEs in heavy
aerosol loading regions.
The top and bottom borders of the box we draw represent the interval of
-2σ,2σ, which contains most of the data
(approximately 95 %) for a given group. The data outside the box are
“possible outliers” based on the largest error contained in each group.
Those possible outliers have one feature in common in that the
corresponding points in the bias scatter plot are far away from other points.
Otherwise, the points below or above the box are different. If a point is
above the box, which indicates that the satellite-retrieved AOD is larger
than the ground-based observed AOD, this outlier tends to be caused by a
residual cloud. The ground-based network measures AOD from only one point;
however, the satellite-retrieved AODs in each collocated pair are an average
of 25 pixels. Any one of these 25 pixels with a cloud residual will lead to
an increased AOD in a collocated pair. Therefore, we conclude that the outliers above the box are possibly caused by cloud residual. From
this point of view, there are 6, 6, and 2 bubbles above each box for the ADV product for
light, moderate, and heavy aerosol loading, respectively. However, these
bubbles are not possible outlier due to the means of uncertainty (MUs) and Std_Ss being relatively small, as shown in Fig. 3. Similarly, the bubbles from the SU product above each box
are not possible outliers. For the ORAC product, most of the bubbles above
each box are possible outliers due to larger Std_S (> 0.2).
Most possible outliers are concentrated in light (13 bubbles) and
moderate (14 bubbles) aerosol loading regions as shown in Table 5, influencing
ORAC's performance in estimating AOD. The bubbles below the box are different
from those above the box. Most of them are only below the boxes of moderate
and heavy aerosol loading, indicating that all these algorithms have
limitations of underestimation in estimating AOD in moderate and heavy
aerosol loading regions, especially when the AOD loading increases.
We make these groups because aerosols exhibit different behaviours with
different loading conditions. In general, the bias or uncertainty of
satellite-retrieved AOD will increase with increasing AOD or aerosol
loading. As discussed above, all of these algorithms underestimate AOD at
different levels; similarly, it is worth noting that underestimation becomes
more severe with increasing AOD or aerosol loading.
Statistics of comparison between AOD bias and ground-based
measurements. Proportion is the ratio of the number of bubbles falling in each
box to the total number. RMSE1 and RMSE2 are RMSEs of AOD bias with ground-based
measurement and AOD uncertainty, respectively.
Additionally, we make a comparison of AOD bias, which is retrieval errors
observed, with AOD uncertainty in AOD retrieval for each pixel the from AATSR L2
dataset. AOD retrieval error observed (AOD bias) and AOD uncertainty in
retrieval are different as evaluating merits. The range of SU AOD uncertainty
is from 0.025 to 0.3, smaller than others, even in heavy aerosol loading
regions. Most bubbles of the ADV product in Fig. 3. are from 0 to 0.4 of AOD
uncertainty. The AOD bias and uncertainty are small in light aerosol loading
and moderate for ADV and SU products, as shown in Fig. 4. For the ORAC product,
there is no obvious regularity between AOD bias with AOD uncertainty in three
aerosol loading regions, especially those bubbles with high Std_S.
Uncertainty analysis of individual ground measurement sites
For the purpose of further evaluating the different performances of these
three algorithms in estimating AOD over mainland China, we validate these
products on a site-by-site basis. It is significant to explore the roles of
different factors in estimating AOD. There are several factors that may have
impacts on AOD calculation, including land cover, aerosol type, and elevation. Therefore, we analyse different validation results of each site to
study how these factors work (see Table 6).
Scatter plots of ADV, ORAC, and SU AOD uncertainty with AOD
bias over China for the 3 years of 2007, 2008, and 2010. The area and colours
of bubbles represent Std_S and AOD, respectively.
Intercomparison of algorithms site by site
In this section, we select five representative AERONET and CARSNET sites with
more than 30 successful matches in 2007, 2008, and 2010 to guarantee an
appropriate statistical sample size. These selected sites are located in
different regions where the land cover and climatic pattern are different
and representative of mainland China. Two AERONET sites and three CARSNET
sites were selected, including SACOL and Xianghe from AERONET, and Lin'an,
Shangdianzi and Xilinhot from CARSNET. Most matches of the ADV and SU products
collocated with ground-based data occurred in March to October in 2007, 2008
and 2010, as shown in Figs. 5 to 9. The matches of the ORAC product were
distributed in each month over most sites.
Lin'an is located at 30.3∘ N, 119.73∘ E,
northwest of Zhejiang province. A total of 80 % of the 50 km × 50 km surrounding area is covered by green vegetation, and
the other 20 % is covered with urban land. The ADV and ORAC algorithm
underestimated AOD, with MBE = 0.13 and 0.12 in 2010, respectively (see
Fig. 5). The SU performed well in Lin'an, with slight underestimation. The
underestimation of the ADV algorithm is more severe than that of SU and
ORAC. Although the ORAC algorithm has the most matches in Lin'an, its
performance was unstable, which means that the level of underestimation was
different in different years.
Statistics of validation results of different products over
different sites.
Time series comparison of AATSR AOD with CARSNET AOD at
Lin'an in 2008 and 2010.
SACOL is situated along the southern bank of the Yellow River in Lanzhou
city, Gansu province. Lanzhou city has a temperate continental climate with
four clearly distinct seasons. The dominant land cover is grassland,
covering approximately 95 % of the spatial extent of the 50 km × 50 km area from the MODIS MCD12C1 land cover data. A total of
30% of the surface consists of arid and semi-arid areas, which can be a source of
dust aerosols. SU performs well in retrieving AOD over SACOL, with a low
RMSE (0.072). The accidental error in the retrievals using the ORAC
algorithm is obvious, leading to a high RMSE (0.170). However, as discussed
above, the ADV algorithm severely underestimated AOD in SACOL. The ADV
algorithm tended to severely underestimate the AOD of different ranges,
except for a small number of high-quality matches. The matches of the SU
product are of high quality for the 3 years. The ORAC has collocated
matches in January, February, November, and December (winter time), unlike
the ADV and SU products. However, the accuracy of ORAC in winter is highly
uncertain, as shown in Fig. 6.
Time series comparison of AATSR AOD with AERONET AOD at
SACOL in 2007, 2008, and 2010.
Shangdianzi is situated at 40.15∘ N,
94.68∘ E, with complex land cover of approximately 45 %
cropland, 30 % mixed forest, 18 % closed shrubland, 5 % grassland,
1 % water, and 1 % evergreen needle leaf forest. The SU algorithm has
high precision in AOD calculation over this site from March to October, when
most of the land cover is green. The ADV algorithm also performs well in
calculating AOD over these three sites, with a slight underestimation. The
performance of the ORAC algorithm in Shangdianzi is unstable, with strong
agreement with ground-based data from March to October and severe
underestimation in winter, as shown in Fig. 7.
Time series comparison of AATSR AOD with CARSNET AOD at
Shangdianzi in 2007, 2008, and 2010.
Time series comparison of AATSR AOD with AERONET AOD at
Xianghe in 2007, 2008, and 2010.
Xianghe is located to the southeast of Beijing and has the same climatic
conditions as Beijing. Approximately 98 % of the surface is covered with
urban land according to the MCD12C1 data of a 50 km × 50 km area. The performances of these three algorithms are at the same high-quality level (see Fig. 8). However, the ADV algorithm still underestimated
AOD at a level of MBE = 0.12 in 2007 and 0.10 in 2008.
Xilinhot is situated at 43.95∘ N,
116.07∘ E, at the centre of the Xilinguole grassland. The
main land cover is grassland (100 %) based on the MODIS MCD12C1 data, with
a spatial extent of 50 km × 50 km. The surface and
climate features of Xilinhot are similar to those of SACOL, and the
performance of the SU algorithm at these two sites is the same, i.e. both
with low RMSE (see Fig. 9). The ADV algorithm slightly underestimated AOD, with an MBE of 0.10–0.13. The ORAC AOD showed weak agreement
with the Xilinhot data, mainly because possible outliers exist in March
to June 2008 and March 2010.
Comparison of SU AOD with CARSNET AOD at Xilinhot in 2008
and 2010.
To guarantee statistical reliability, there must be more than 30 collocated
pairs at one site. The determination of the surface cover at each site is
based on the proportion (> 80 % for one land type) of each land
cover type from the MCD12C1 data at a spatial extent of 50 km × 50 km. If no land cover type accounts for a proportion larger
than 80 % at a given site, it will be identified as mixed; then, we select
two or more (sum > 80 %) land types with the largest
proportions as the main land cover. As the data volume is too low to infer
the year-to-year variability of performance at these sites, the analysis
gives some useful information, but it is important not to overinterpret
results from a small selection of data points.
Analysis of algorithm performances in western China
Because sufficient ground-based data in western China are lacking for the
AERONET measurements, only data from CARSNET sites are used in 2008. We
selected six CARSNET sites located in western China in which there are more
than 25 matches.
Urumchi, situated at 43.78∘ N, 87.62∘ E,
serves as the provincial capital of Xinjiang Uyghur Autonomous Region and is
the most remote city in China in terms of distance to any sea. The dominant
land cover at the spatial extent of 50 km × 50 km is
grassland, which accounts for approximately 85 %. The ADV, ORAC, and SU
algorithms all severely underestimated AOD, with MBE = 0.22, 0.12, and 0.17, respectively. The MBE is lowest mainly because of the outlier in
April, which decreases the MBE (see Fig. 10).
Time series comparison of AATSR AOD with CARSNET AOD at
Urumchi in 2008.
Ejina is situated at 41.95∘ N, 101.07∘ E,
and its main land cover is barren ground (84 %). The performances of ORAC
and SU are at the same high-quality level, with MBEs of 0.02 and 0.09,
respectively. Another reason why we chose this site is that there are no
matches of ADV products successfully collocated with ground-based data.
Based on Fig. 11, the ORAC algorithm has strong applicability in Ejina and
high accuracy in retrieving AOD. The SU algorithm also performed well. This
demonstrates that another limitation of the ADV algorithm is its
applicability in calculating AOD in China. Dunhuang is situated at
40.15∘ N, 94.68∘ E and is surrounded by
barren ground (85 %). The same situation is true for Ejina, which causes a slight underestimation at each point but high R and low RMSE for the ORAC
algorithm (Fig. 12). The performance of the SU algorithm was not as good
as that of the ORAC because of its underestimation with MBE = 0.10. The
limits of underestimation and applicability of the ADV were more obvious at
this site, as it only had six matches and showed severe underestimation with
MBE = 0.17. Tazhong is situated at 39∘ N,
83.67∘ E and is surrounded by barren or sparsely vegetated surface.
Almost all land cover is barren ground according to the MODIS MCD12C1 data.
Similar to the former two sites, the ADV product did not have any successful
matches at this site (Fig. 13). Both the ORAC and SU algorithms exhibited
severe underestimation of retrievals, with MBE = 0.17 and 0.20,
respectively. The outliers of the ORAC product in February are much higher
than the observation data, causing the lower MBE.
Time series comparison of AATSR AOD with CARSNET AOD at
Ejina in 2008.
Time series comparison of AATSR AOD with CARSNET AOD at
Dunhuang in 2008.
Time series comparison of AATSR AOD with CARSNET AOD over
the site of Tazhong in 2008.
DR distribution of specific sites.
SiteDR < 11 < DR < 33 < DR < 55 < DRTotalUrumchi47402190Ejina51431196Tazhong63175388Dunhuang57311291
Seasonal distribution of validation results of three algorithms.
The prevailing climatic pattern in western China is a temperate continental
climate with four distinct seasons and less precipitation in winter and
spring. In conclusion, compared to eastern China, the applicability of the
ADV algorithm is not strong, and the underestimation is more severe. In the
four selected sites in western China, the performance of the ORAC algorithm
is best, even though severe underestimation occurs at some sites. The
accuracy of the SU algorithm is not as high as the ORAC product, with more
severe underestimation and lower applicability.
Intercomparison
In conclusion, the SU algorithm performs well in calculating AOD over
different land covers from March to October. Slight underestimation occurs
over barren ground or sparse vegetation at different times, and there are no
obvious features in terms of precision in the time series over grasslands.
For complex land surfaces where the dominant land cover is vegetation, the
SU algorithm is extremely effective in estimating AOD. In the last section,
we draw the conclusion that the SU algorithm underestimates AOD over mainland
China in 2008 probably because the dominant land cover in western China is
barren or sparse vegetation, over which the SU algorithm underestimates AOD
more severely.
Scatter plots of AATSR ADV, ORAC, and SU L2 AOD products with
ground-based data in China in the spring, summer, autumn, and winter time of
2007, 2008, and 2010.
Scatter plot of AATSR AOD and DT and DB AOD.
Comparisons of Ångström exponent of ORAC and SU
products. The area and colours of bubbles represent AOD uncertainty and AOD
values, respectively.
The ADV algorithm underestimates AOD at most of the selected sites. We
categorize these sites as four classes according to the MBEs of different
sites: Class 1 (MBE < 0.1), Class 2
(0.2 > MBE > 0.1), Class 3
(0.3 > MBE > 0.2), and Class 4
(MBE > 0.3). The ADV algorithm underestimates AOD over all
selected sites, leading to all selected MBEs being larger than 0. We make
such categories for the purpose of assessing the contribution of different
surfaces to AOD estimation. Only Xianghe of 2008 belongs to Class 1, and
Lin'an, Shangdianzi, and SACOL are classified as Class 2. Only Urumchi is in
Class 3. Note that even though Lanzhou and Datong were not selected due their
location, they should be classified as Class 4.
Overall, the ADV algorithm underestimates AODs at all sites but at different
levels, as demonstrated by the above categories. Serious underestimation
occurs over the sites in Class 3 and Class 4 in western China, where the
dominant land cover is a mixing of urban area and a large portion of
grasslands. For the sites in Class 2, differences exist between Beijing and
SACOL. SACOL is similar to the sites in Class 3 and Class 4, the main land
cover of which is grassland. Over the sites in Class 1, the algorithm
performs well with high R and low MBE, but there are no common features in
terms of surface conditions.
The ORAC product collocates most pairs of all of these products. Most
collocated pairs of the SU product and ADV product occur in March to
October, but the collocated pairs of the ORAC product occur during each
month over some sites in 2008. Because more matches suggest greater errors
for the determination of the outlier contribution to the overall
performance of the ORAC algorithm, we introduce the ratio of the individual
difference to average the differences for each site:
DR=τaero,i-τsate,i∑i=1nτaero,i-τSate,i/n,
where DR < 1 indicates a “relatively good” match,
3 > DR > 1 indicates a “relatively poor” match,
and DR > 3 is an outlier (see Table 7).
There are no obvious possible outliers in Ejina shown in Fig. 11. Most of
the DRs are in the range of 0 to 3, only two DRs are larger than 3, and the
maximum (overestimation) is 5.112. The retrieved AOD in March is a possible outlier because it is overestimated, whereas most are underestimated.
Another two sites dominated by barren or sparsely vegetated land cover are
Dunhuang (approximately 85 %) and Tazhong (100 %). The conditions in
Tazhong are complex, and there is no obvious relationship between the CARSNET
data and the ORAC AODs. Most of the DRs are less than 3, and a total of eight
DRs are larger than 3. The DR in February is an outlier because the
varying tendencies are different between the ORAC product and the
ground-based data, indicative of overestimation.
The ORAC product has the largest coverage, at the expense of accuracy,
especially in the presence of outliers, and only the ORAC product has
collocated validation pairs over some sites during each month in all 3 years. The ORAC algorithm underestimates AODs over Ejina, Tazhong, and
Dunhuang, but the possible outliers reduce the differences between the
CARSNET data and the ORAC product. Xilinhot, Urumchi, and SACOL share the
same main land cover of grassland. The problem is that the underestimations
over these sites are not at the same level.
It is worth noting that the ORAC algorithm has the ability to calculate high
AOD; however, most of the AODs have DRs larger than 3, indicating that the
estimation of high AOD is unstable and has large error, reducing the overall
precision.
Seasonal characteristics of three algorithms
Mainland China, covering about 60∘ of longitude and 30∘ of
latitude, is dominated by a monsoon-driven climate. In such a vast territory,
there are big differences in climate patterns between western and eastern China.
The main climate type in eastern and eastern coastal China is a monsoon
climate. For western China far from the ocean, the climate type is a hybrid
of a monsoon and continental climate. In dry seasons (winter, first half of
spring, and last half of autumn), poor vegetation coverage, loose surface, and winds in most northern China regions turn coarse particles (sea salt and
desert dust) into aerosol. Fine particles from coal combustion in winter and
soot from straw burning in autumn are also an important source of aerosol. In
rainy seasons (mainly in summer), high vegetation blocks dust being blown as aerosol and reduces surface reflectance in the visible wavelength. Table 8 shows
the seasonal distribution of the validation results of the three algorithms. For mainland China, which is located in the Northern Hemisphere from
20–55∘ N, the spring time starts on about March to May, the
summer time starts in about June to August, the autumn time starts in about September to November, and the winter time is from about December to
February of the next year.
Low MUs at 550 nm means that these retrievals are of high
quality in Fig. 2. Most Std_S are below 0.08, indicating high uniformity
of ADV products (see Fig. 2). Most collocated pairs of ADV AODs are
concentrated below the 1–1 line, and the RMB is 0.61, showing a tendency to
underestimate. This kind of underestimation has an impact on ADV algorithm
performances; for example, the RMSE is 0.19 in summer time, otherwise,
the corresponding RMB is 0.54, which makes the KAPPA coefficient the smallest
(0.26) of all seasons. The MBEs are from -0.12 in autumn to -0.16 in
spring in Table 8, which means that the ADV algorithm tends to underestimate
AOD in all seasons (except winter) over mainland China (See Fig. 13). For the monsoon climate, the main aerosol types in many parts of China are influenced
by coarse particles (dust from western China and sea salt from eastern
coastal China) in spring time. The performance in calculating aerosol
properties of a mixture of coarse particles is best in spring time, with the highest
KAPPA coefficient, even though there are some samples with high MUs and the
RMSE is 0.23.
The matches of the ORAC product collocated with reference data are
distributed discretely at two sides of the 1–1 line in Fig. 2. The best
performance, with a high KAPPA coefficient of 0.5, is in spring with no
underestimation, even though the RMSE is high (about 0.30). The KAPPA
coefficient in the autumn time is lower than in the spring time, even though
most evaluation metrics are better in the autumn. Note that only the ORAC
product of these three products collocated enough matches (more
than 30) with reference data in the winter time. The performance of ORAC in
winter is between that in spring and autumn without obvious underestimation
or overestimation. The limitation of the ORAC algorithm is the stability in
retrieving aerosol properties, as shown in Fig. 14: the magenta
mean ±2σ lines for each season in each range are longer
than those for the other two products.
The SU algorithm has better performances in 3 years, getting KAPPA
coefficients of 0.50. Most retrievals in matches are of high quality
collocated with reference data and most Std_S are lower than 0.08; i.e. the
sample quality is high and this coincides with an assumption of aerosol
property uniformity in a 50 km × 50 km area. The best
performance on retrieval is in the autumn, with the lowest RMSE being 0.16 and
the largest RMB being 0.74 in three seasons shown in Fig. 14. The magenta lines are
similar to those of the ADV product in corresponding seasons, showing same
level of stability in retrieving AOD. The SU algorithm has no obvious
differences in retrieving AOD in three seasons. One limitation of the SU and ADV
algorithms is less than 30 collocated matches in the winter time so that we
cannot evaluate its performance during that time.
The latest MODIS Collection 6 (C6) products were released in 2013, including
aerosol datasets produced by two “dark target” (DT) algorithms (one is for
retrieving over ocean and the other is for retrieving over land) and the Deep
Blue (DB) algorithm for retrieving over
bright or semi-arid surface (Levy et al., 2013). For over land, the DT
algorithm uses an updated cloud mask to allow the retrieval of heavy aerosol
compared to the algorithm employed in MODIS Collection 5. It is reported that
MODIS C6 products (produced by three algorithms) are of high quality (Sayer
at el., 2014). Here, we select both MODIS C6 DT
and DB 10 km × 10 km merged datasets as reference data for
cross-validation of AATSR L2 AOD products. The matches in Fig. 15 are
randomly chosen from MODIS and AATSR collocated AOD datasets. The ADV AOD has
a lowest RMSE of 0.11. The SU algorithm has the same performance as ORAC
(similar RMSE and KAPPA) but with a little underestimation, represented as
the magenta line in Fig. 15.
The aerosol Ångström exponent is an exponent that expresses the spectral
dependence of aerosol optical thickness with the wavelength of incident light
(Eck et al., 1999). The Ångström exponent is inversely related to the
average size of the particles in the aerosol: the smaller the particles, the
larger the exponent. Thus, the Ångström exponent is a useful quantity to
assess the particle size of atmospheric aerosols or clouds and the
wavelength dependence of the aerosol or cloud optical properties.
The ORAC product provides the Ångström exponent for 550–870 µm only
and the SU product provides the Ångström exponent for 550–870 µm only.
The CARSNET dataset provides the Ångström exponent for 440–870 µm
only. As the ADV product provides the Ångström exponent for 550–670 µm only, we could not carry out a comparison for the ADV Ångström exponent. We
compared the Ångström exponent using both CARSNET and AERONET datasets
for SU and ORAC products. Figure 16 shows the comparisons of the Ångström exponent. In general, both the SU and ORAC algorithms
generate a similar quality of Ångström exponent values. There is no pattern linking the Ångström exponent with AOD values and uncertainty.
Conclusions
Satellite remote sensing of the atmosphere is one important aspect in the PEEX
scientific plan. Remotely sensed data could provide continuous spatial
coverage of aerosol property over the pan-Eurasian area. These three
algorithms (the SU algorithm, the ADV algorithm, and the ORAC algorithm)
display different performances in estimating AOD over mainland China in
2007, 2008, and 2010. However, none of the algorithms show an explicitly
better performance than the other two. The SU and ADV products have a higher
accuracy over most selected sites but less coverage, whereas the ORAC
product has greater coverage at the cost of accuracy.
All of these algorithms tend to underestimate AOD to some degree. The
underestimation becomes more severe with increasing AOD or aerosol loading.
The method of grouping helps to identify possible outliers in different
regions of aerosol loading.
The precision of the SU and ADV algorithms is at the same level over
different surfaces. However, the SU product has stricter quality control
than the ADV product, and it eliminates AODs to make the MBE less than 0.10
over different sites (de Leeuw et al., 2013). Over grassland and barren
vegetation, the SU displays a strong performance with slight underestimation
(MBE < 0.10). The limitations of underestimation and the applicability of
the ADV are more obvious over such sites. For complex surfaces with two or
more land cover types, the performances of these three algorithms are at the
same level. Note that Lanzhou and Datong are different from other sites, even
though the main land cover type is grassland. All of these algorithms
underestimated AOD at a high level, perhaps because these algorithms are not
sensitive to absorptive aerosols.
Only the ORAC product shows possible outliers identified by Eq. (2), which substantially decreases its accuracy. The most obvious feature of
the possible outliers is that the retrieved AODs are higher than the
ground-based measurements.
As reference data, AERONET L2 data have some limitations, including the
distribution and number of sites in mainland China. Most sites of AERONET are
distributed in eastern China and the coastal region of China for special
experimental use; as a result, sufficient reference data cannot be obtained
to validate the AOD product. The CARSNET data make up for this shortage
because there are more CARSNET sites in China, especially in western China,
where few AERONET sites have been constructed. Limited both by reference data
and satellite retrievals, most co-allocated pairs occur in March to November,
and a few occur in winter (December, January, and February).
Data availability
The ESA Aerosol_cci AOD research data (ADV, ORAC and SU) used in this
publication are publicly available at
http://www.icare.univ-lille1.fr/cci. The MODIS AOD data are available
via http://modis-atmos.gsfc.nasa.gov/MOD04_L2/. AERONET data have been
downloaded from http://aeronet.gsfc.nasa.gov/.
Acknowledgements
This work was supported in part by the Ministry of Science and Technology
(MOST) of China under grant nos. 2013CB733403 and 2016YFA0600302, the
National Natural Science Foundation of China (NSFC) under grant nos. 41471306
and 41590853, and the EU/FP7 MarcPolo project (Grant Agreement no. 606953).
Part of the work was conducted in preparation for the Aerosol_cci project
(ESA-ESRIN project AO/1-6207/09/I-LG), from which three AATSR AOD products
were provided. The data for uncertainty analysis and validation came from 34
AERONET sites and 8 CARSNET sites. We thank the PIs, investigators, and their
staff for establishing and maintaining the data for this study.
Edited by: A. Ding Reviewed
by: two anonymous referees
References
Adhikary, B., Kulkarni, S., Dallura, A., Tang, Y., Chai, T., Leung, L. R.,
Qian, Y., Chung, V., Ramanathan, C. E., and Carmichael, G. R.: A regional
scale chemical transport modelling of Asian aerosols with data assimilation
of AOD observations using optimal interpolation technique, Atmos. Environ.,
42, 8600–8615, 2008.
Andreae, M. O. and Rosenfeld, D.: Aerosol-cloud-precipitation interactions,
Part 1. The nature and sources of cloud-active aerosols, Earth-Sci. Rev.,
89, 13–41, 2008.
Bevan, S., North, P., Los, S., and Grey, W.: A global dataset of atmospheric
aerosol optical depth and surface reflectance from AATSR, Remote Sens.
Environ., 116, 199–210, 2012.
Bilal, M., J. E. Nichol, and Chan, P. W.: Validation and accuracy assessment
of a Simplified Aerosol Retrieval Algorithm (SARA) over Beijing under low
and high aerosol loadings and dust storms, Remote Sens. Environ., 153,
50–60, 2014.Bloch, D. A. and Kraaemer, H. C.: 2×2 Kappa
coefficients: Measures of agreement or association, Biometrics, 45, 269–287,
1989.Che, H., Zhang, X., Chen, H., Damiri, B., Goloub P., Li, Z., Zhang, X., Wei,
Y., Zhou, H., Dong, F., Li, D., and Zhou, T.: Instrument calibration and
aerosol optical depth validation of the China Aerosol Remote Sensing Network,
J. Geophys. Res., 114, D03206, 10.1029/2008JD011030, 2009.Che, Y., Xue, Y., Xu, H., Mikusauskas, R., and She, L.: The inter-comparison
of AATSR aerosol optical depth retrievals from various algorithms, in
Proceedings of the IEEE International Geoscience and Remote Sensing
Symposium held in Milan, Italy, 27–31 July 2015, 2230–2233,
10.1109/IGARSS.2015.7326249, 2015.Chu, D. A., Kaufman, Y. J., Ichoku, C., Remer, L. A., Tanré, D., and
Holben, B. N.: Validation of MODIS aerosol optical depth retrieval over land,
dust, Geophys. Res. Lett., 29, MOD2-1–MOD2-4, 10.1029/2001GL013205,
2002.Climate Change Initiative: http://www.esa-aerosol-cci.org/, last access: 15
January 2016.
Cohen, J. A.: A coefficient of agreement for nominal scales, Educ. Psychol.
Meas., 20, 37–46, 1960.
de Haan, J. F., Bosma, P. B., and Hovenier, J. W.: The adding method for
multiple scattering calculations of polarized light, Astron. Astrophys., 183,
371–391, 1987.
de Leeuw, G., Holzer-Popp, T., Bevan, S., Davies, W. H., Descloitres, J.,
Grainger, R. G., Griesfeller, J., Heckel, A., Kinne, S., Klüser, L.,
Kolmonen, P., Litvinov, P., Martynenko, D., North, P., Ovigneur, B., Pascal,
N., Poulsen, C., Ramon, D., Schulz, M., Siddans, R., Sogacheva, L.,
Tanré, D., Thomas, G. E., Virtanen, T. H., von Hoyningen-Huene, W.,
Vountas, M., and Pinnock, S.: Evaluation of seven European aerosol optical
depth retrieval algorithms for climate analysis, Remote Sens. Environ., 162,
295–315, 2013.
Eck, T. F., Holben, B. N., Reid, J. S., Dubovik, O., Smirnov, A., O'Neill,
N. T., Slutsker, I., and Kinne, S.: Wavelength dependence of the optical
depth of biomass burning, urban, and desert dust aerosols, J.
Geophys. Res.-Atmos., 104, 31333–31349, 1999.
Flowerdew, R. J. and Haigh, J. D.: Retrieval of aerosol optical thickness
over land using the ATSR-2 Dual-Look Satellite Radiometer, Geophys. Res.
Lett., 23, 351–354, 1996.
Foody, G. M.: Status of land cover classification accuracy assessment,
Remote Sens. Environ., 80, 185-=-201, 2002.
Grey, W. M. E., North, P. R. J., Los, S. O., and Mitchell, R. M.: Aerosol
Optical Depth and Land Surface Reflectance from Multiangle AATSR
Measurements: Global Validation and Intersensor Comparisons, IEEE T. Geosci.
Remote, 44, 2184–2197, 2006.
Griggs, M.: Satellite Observation of Atmospheric Aerosols During the EOMET
Cruise, J. Atmos. Sci., 36, 695–698, 1979.
Holben, B. N., Eck, T. F., Slutsker, I., Tanré, D., Buis, J.
P., Setzer, K. A., Vermote, E., Reagan, J. A., Kaufman, Y. J., Nakajima, T.,
Lavenu, F., Jankowiak, I., and Smirnov, A.: AERONET – A Federated Instrument
Network and Data Archive for Aerosol Characterization, Remote Sens.
Environ., 66, 1–16, 1998.Holzer-Popp, T., de Leeuw, G., Griesfeller, J., Martynenko, D., Kluser, L.,
Bevan, S., Davies, W., Ducos, F., Deuze, J. L., Graigner, R. G., von
Hoyningen-Huene, W., Kolmonen, P., Litvinov, P., North, P., Poulsen, C. A.,
Ramon, D., Siddans, R., Sogacheva, L., Tanre, D., Thomas, G. E., Vountas,
M., Descloitres, J., Griesfeller, J., Kinne, S., Schulz, M., and Pinnock,
S., A.:Aerosol retrieval experiments in the ESA Aerosol_cci project, Atmos. Meas. Tech., 6, 1919–1957, 10.5194/amt-6-1919-2013, 2013.Ichoku, C, D. A. Chu, Mattoo, S., Kaufman, Y. J., Remer, L. A., Tanré,
D., Slutsker, I., and Holben, B. N.: A spatio-temporal approach for global
validation and analysis of MODIS aerosol products, Geophys. Res. Lett., 29,
MOD1-1–MOD1-4, 10.1029/2001GL013206, 2002.
IPCC: Intergovernmental Panel on Climate Change: Climate Change 2013: The
Physical Science Basis, Contribution of Working Group I to the Fifth
Assessment Report of the IPCC, edited by: Stocker, T. F., Qin, D., Plattner,
G. K., Tignor, M., Allen, S. K., Boschung, J., Nauels, A., Xia, Y., Bex, V.,
an Midgley, P. M., Cambridge University Press, Cambridge, UK, New York, NY,
USA, 2013.Kahn, R. A., Gaitley, B. J., Martonchik, J. V., Diner, D. J., and Crean, K.
A.: Multiangle Imaging Spectroradiometer (MISR) global aerosol optical depth
validation based on 2 years of coincident Aerosol Robotic Network (AERONET)
observations, J. Geophys. Res., 110, D10S04, 10.1029/2004JD004706,
2005.Kahn, R. A., Li, W. H., Moroney, C., Diner, D. J., Martonchik, J. V., and
Fishbei, E.: Aerosol source plume physical characteristics from Space-based
multiangle imaging, J. Geophys. Res., 112, D11205,
10.1029/2006JD007647, 2007.
Kahn, R. A., Nelson, D. L., Garay, M. J., Levy, R. C., Bull, M. A., Diner, D.
J., Martonchik, J. V., Paradise, S. R., Hansen, E. G., and Remer, L. A.: MISR
aerosol product attributes and statistical comparisons With MODIS, IEEE T.
Geosci. Remote, 47, 4095–4113, 2009.
Kahn, R. A., Garay, M. J., Nelson, D. L., Levy, R. C., Bull, M. A., Diner,
D. J., Martonchik, J. V., Hansen, E. L., Remer, L. A., and
Tanré, D.: Response to “Ward Unified Satellite Climatology
of Aerosol Properties, 3. MODIS versus MISR versus AERONET”, J. Quant.
Spectrosc. Ra., 112, 901–909, 2011.
Kaufman, Y. J., Tanré, D., Remer, L. A., Vermote, E. F., Chu, A.,
Holben, B. N.: Operational remote sensing of tropospheric aerosol over land
from EOS moderate resolution imaging spectroradiometer, J. Geophys. Res.,
102, 17051–17067, 1997.Kinne, S., Lohmann, U., Feichter, J., Schulz, M., Timmreck, C., Ghan, S.,
Easter, R., Chin, M., Ginoux, P., Takemura, T., Tegen, I., Koch, D., Herzog,
M., Penner, J., Pitari, G., Holben, B., Eck, T., Smirnov, A., Dubovik, O.,
Slutsker, I., Tanre, D., Torres, O., Mishchenko, M., Geogdzhayev, I., Chu, D.
A., and Kaufman, Y.: Monthly averages of aerosol properties: a global
comparison among models, satellite data, and AERONET ground data, J. Geophys.
Res., 108, 4634, 10.1029/2001JD001253, 2003.
Kinne, S., O'Donnel, D., Stier, P., Kloster, S., Zhang, K., Schmidt, H.,
Rast, S., Giorgetta, M., Eck, T. F., and Stevens, B.: MAC-v1: A new global
aerosol climatology for climate studies, J. Adv. Model Earth Syst., 5,
704–740, 2013.
Kokhanovsky, A. A. and de Leeuw, G.: Satellite Aerosol Remote Sensing Over
Land, Springer, Berlin, 2009.
Kolmonen, P., Sogacheva, L., Timo, H., Virtanen, de Leeuw, G., and Kulmala,
M.: The ADV/ASV AATSR aerosol retrieval algorithm: current status and
presentation of a full-mission AOD dataset, Int. J. Digit. Earth, 9,
545–561, 2016.
Kulmala, M., Alekseychik, P., Paramonov, M., Laurila, T., Asmi, E., Arneth, A.,
Zilitinkevich, S., and Kerminen, V.-M.: On measurements of aerosol particles
and greenhouse gases in Siberia and future research needs, Boreal Env. Res.,
16, 337–362, 2011.Kulmala, M., Lappalainen, H. K., Petäjä, T., Kurten, T., Kerminen,
V.-M., Viisanen, Y., Hari, P., Sorvari, S., Bäck, J., Bondur, V.,
Kasimov, N., Kotlyakov, V., Matvienko, G., Baklanov, A., Guo, H. D., Ding,
A., Hansson, H.-C., and Zilitinkevich, S.: Introduction: The Pan-Eurasian
Experiment (PEEX) – multidisciplinary, multiscale and multicomponent
research and capacity-building initiative, Atmos. Chem. Phys., 15,
13085–13096, 10.5194/acp-15-13085-2015, 2015.Levy, R. C., Remer, L. A., and Dubovik, O.: Global aerosol optical properties
and application to Moderate Resolution Imaging Spectroradiometer aerosol
retrieval over land, J. Geophys. Res., 112, D13210,
10.1029/2006JD007815, 2007.Levy, R. C., Remer, L. A., Kleidman, R. G., Mattoo, S., Ichoku, C., Kahn, R.,
and Eck, T. F.: Global evaluation of the Collection 5 MODIS dark-target
aerosol products over land, Atmos. Chem. Phys., 10, 10399–10420,
10.5194/acp-10-10399-2010, 2010.Levy, R. C., Mattoo, S., Munchak, L. A., Remer, L. A., Sayer, A. M.,
Patadia, F., and Hsu, N. C.: The Collection 6 MODIS aerosol products over
land and ocean, Atmos. Meas. Tech., 6, 2989–3034,
10.5194/amt-6-2989-2013, 2013.
Li, M. and Zhang, L.: Haze in China: Current and future challenges, Environ.
Pollut., 189, 85–86, 2014.Martonchik, J. V., Diner, D. J., Kahn, R., and Gaitley, B.: Comparison of
MISR and AERONET aerosol optical depths over desert sites, Geophys. Res.
Lett., 31, L16102, 10.1029/2004GL019807, 2004.
McMurry, P. H.: A review of atmospheric aerosol measurements, Atmos.
Environ., 34, 1959–1999, 1999.
Mei, L., Xue, Y., Xu. H., Guang, J., Li, Y., Wang, Y., Ai, J., Jiang, S., And
He, X.: Validation and analysis of aerosol optical thickness retrieval over
land, Int. J. Remote Sens., 33, 781–803, 2012.Misra, A., Jayaraman, A., and Ganguly, D.: Validation of MODIS derived
aerosol optical depth over Western India, J. Geophys. Res., 113, D04203,
10.1029/2007JD009075, 2008.
North, P. R. J., Briggs, S. A., Plummer, S. E., and Settle, J. J.: Retrieval
of land surface bidirectional reflectance and aerosol opacity from ATSR-2
multiangle imagery, IEEE T. Geosci. Remote, 37, 526–537, 1999.
Prins, E. M., Feltz, J. M., Menzel, W. P., Menzel, W. P., and Ward, D. E.: An
overview of GOES-8 diurnal fire and smoke results for SCAR-B and 1995 fire
season in South America, J. Geophys. Res., 103, 31821–31836, 1998.
Remer, L. A., Kaufman, Y. J., Tanré, D., Mattoo, S., Chu, D. A., Martins,
J. V., Li, R.-R., Ichoku, C., Levy, R. C., Kleidman, R. G., Eck, T. F.,
Vermote, E., and Holben, B. N.: The MODIS aerosol algorithm, products, and
validation, J. Atmos. Sci., 62, 947–973, 2005.Rosenfield, G. H. and Fitzpatrick-Lins, K.: A coefficient of agreement as a
measure of thematic classification accuracy, Photogramm. Eng. Rem. S., 52,
223–227, 1986.
Samet, J. M., Dominici, F., Curriero, F. C., Coursac, I., and Zeger, S. L.:
Fine particulate air pollution and mortality in 20 US cities, 1987–1994, New
Engl. J. Med., 343, 1742–1749, 2000.
Sayer, A. M., Munchak, L. A., Hsu, N. C., Levy, R. C., Bettenhausen, C., and
Jeong, M.-J., MODIS Collection 6 aerosol products: Comparison between Aqua's
e-Deep Blue, Dark Target, and “merged” data sets, and usage
recommendations, J. Geophys. Res.-Atmos., 119, 13965–13989, 2014.
Stehman, S. V.: Selecting and interpreting measures of thematic
classification accuracy, Remote Sens. Environ., 62, 77–89, 1997.Takahashi, Y., Higashi, M., Furukawa, T., and Mitsunobu, S.: Change of iron
species and iron solubility in Asian dust during the long-range transport
from western China to Japan, Atmos. Chem. Phys., 11, 11237–11252,
10.5194/acp-11-11237-2011, 2011.
Thomas, G. E., Carboni, E., Sayer, A. M., Poulsen, C. A., Siddans, R., and
Grainger, R. G.: Oxford-RAL Aerosol and Cloud (ORAC): Aerosol Retrievals from
Satellite Radiometers, in: Aerosol Remote Sensing Over Land, edited by:
Kokhanovsky, A. and de Leeuw, G., Springer, Berlin, 2009.
Torres, O., Bhartia, P. K., Herman, J. R., Sinyuk, A., Ginoux, P., and
Holben, B.: A long-term record of aerosol optical depth from TOMS
observations and comparison to AERONET measurements, J. Atmos. Sci., 59,
398–413, 2002.
Tzanis, C. and Varotsos, C. A.: Tropospheric aerosol forcing of climate: a
case study for the greater area of Greece, Int. J. Remote Sens., 29,
2507–2517, 2008.Varotsos, C.: Airborne measurements of aerosol, ozone, and solar ultraviolet
irradiance in the troposphere, J. Geophys. Res., 110, D09202,
10.1029/2004JD005397, 2005
Varotsos, C., Ondov, J., Tzanis, C., Ozturk, F., Nelson, M., Ke, H., and
Christodoulakis, J.: An observational study of the atmospheric ultra-fine
particle dynamics, Atmos. Environ., 59, 312–319, 2012.
Veefkind, J. P., de Leeuw, G., and Durkee, P. A.: Retrieval of aerosol
optical depth over land using two-angle view satellite radiometry during
TARFOX, Geophys. Res. Lett, 25, 3135–3138, 1998.
Xue, Y., Xu, H., Guang, J., Mei, L., Guo, J., Li, C., Mikusauskas, R., and
He, X.: Observation of an agricultural biomass burning in central and east
China using merged aerosol optical depth data from multiple satellite
missions, Int. J. Remote Sens., 35, 5971–5983, 2014.