Introduction
Tropospheric nitrogen oxides (NOx= NO + NO2) are important
pollutants affecting ozone, aerosols, acid deposition, and climate. China
has become the top emitter of NOx due to its recent anthropogenic
emission growth along with reductions in North America and Europe (Richter
et al., 2005; van der A et al., 2008; Zhang et al., 2009; Lamsal et al.,
2011; Castellanos and Boersma, 2012; Lin et al., 2014a). High NOx
pollution not only has significant consequences for China's domestic
environment (Zhao et al., 2009; Lin et al., 2010a; Zhang et al., 2012), but
it has also raised concerns regarding long-range pollution transport to
downstream regions (Lin et al., 2008, 2014a; Cooper et al., 2010; Zhang et al., 2014; Jiang et al., 2015).
Vertical column densities (VCDs) of tropospheric nitrogen dioxide (NO2)
retrieved from the Ozone Monitoring Instrument (OMI) have been used
extensively to study Chinese NOx pollution (Stavrakou et al., 2008; Zhao
and Wang, 2009; Lin et al., 2010b; Mijling et al., 2013; Miyazaki and Eskes,
2013). The high spatiotemporal coverage of OMI is superior to ground-based in
situ measurements. However, NO2 retrievals from OMI and other
space-borne instruments are subject to large systematic and random errors due
to uncertainties in the conversion process from radiance to VCDs (Boersma et
al., 2011; Bucsela et al., 2013). In particular, current NO2 algorithms
take an implicit approach to accounting for aerosol optical effects, with no
explicit specification of aerosols in the retrievals of both NO2 VCDs
and ancillary cloud parameters. The rationales for this approach are that (1)
aerosols affect the retrieval of cloud parameters, so that the retrieved
cloud parameters are “effective” and implicitly contain certain aerosol
information, and (2) these effective cloud parameters at least partly
describe the effect of aerosols on NO2 air mass factors (Boersma et al.,
2004, 2011). This implicit treatment is supported by the good spatial
correlation (0.66) observed between coincident MODIS aerosol optical
thickness values (mostly due to scattering) and O2–O2 effective
cloud fractions over the eastern United States (Boersma et al., 2011).
Our previous study (Lin et al., 2014b) for several locations in the North
China Plain (NCP) has shown large changes in retrieved NO2 VCDs when
moving from an implicit to an explicit treatment of aerosols. In particular,
NO2 VCDs are reduced by 14 % on average but are changed by
(-90)–(+70) % for individual pixels when aerosol optical depth (AOD)
exceeds 0.8. In addition, current NO2 retrievals are often based on
monthly climatological surface albedo data, not accounting for the angular
dependence of surface reflectance and its interannual variability; the
corresponding effect on retrieved NO2 has been estimated at 0–20 %
for Europe (Zhou et al., 2010) and the NCP (Lin et al., 2014b) on average.
Despite the complex terrains (Fig. 1a), complex land use types (Fig. 1b),
and high aerosol loadings (Xin et al., 2007; Che et al., 2009; Wang et al.,
2011) over China, the effects of aerosol and surface reflectance treatments
are largely unknown.
(a) USGS GMTED2010 surface elevation and (b) ISCGM
GLCNMO land use types mapped to a 0.25∘ long. × 0.25∘
lat. grid. Provincial boundaries of China are shown. Urban land use type is
highlighted in (b), such that a grid cell is designated as “urban”
if at least 5 % of its land is covered by urban areas. Also indicated in
(a) are provinces (in white) and regions (in various colors)
mentioned in the text. Eastern China: 101.25–126.25∘ E,
20–46∘ N; western China: 80–101.25∘ E,
20–50∘ N; northern eastern China: 110–122∘ E,
29–41∘ N; Sichuan Basin: 103–108∘ E, 28–32∘ N;
Beijing–Tianjin–Hebei (BTH): 115.25–118.25∘ E,
38–41∘ N; urban Beijing: 116–116.75∘ E,
39.75–40.25∘ N; Yangtze River delta (YRD): 119–122∘ E,
29.5–32∘ N; urban Shanghai: 121.25–121.75∘ E,
31–31.5∘ N; Pearl River delta (PRD): 112.5–114.5∘ E,
21.5–23.75∘ N; and urban Guangzhou: 113–113.5∘ E,
23–23.25∘ N. Urban Beijing, urban Shanghai, and urban Guangzhou are
inside BTH, YRD and PRD, respectively. The North China Plain (NCP) indicated
in red represents the low-elevation (< 300 m) areas of northern
eastern China.
This study extends our previous work (for a few locations; Lin et al.,
2014b) to introduce an improved pixel-specific level-2 retrieval of
tropospheric NO2 VCDs over China (80–130∘ E,
20–53∘ N), Peking University OMI NO2 (POMINO).
Using a parallelized LIDORT-driven AMFv6 package (Palmer et al., 2001;
Martin et al., 2003; Lin et al., 2014b), we explicitly account for aerosol
optical effects, surface reflectance anisotropy, and their spatiotemporal
variability. We then evaluate the individual and combined effects of an
implicit aerosol treatment and changes in surface reflectance
characteristics. In particular, we show large seasonal and spatial
dependence of the effects of aerosol and/or surface reflectance treatments.
We further illustrate the influences on subsequent NOx emission
constraints, a popular application of OMI data. Our POMINO data are
available for 2004–2013 and will be updated to more recent times. Results
for 2012 are presented here, by aggregating level-2 data into monthly mean
values on a 0.25∘ long. × 0.25∘ lat. grid. Various
provinces and regions are defined in Fig. 1a to facilitate the present
analysis.
Section 2 presents our POMINO retrieval approach. Section 3 analyzes the
POMINO NO2 VCDs and the effects of various treatments of aerosols and
surface reflectance. Section 4 further shows the effects on subsequent
NOx emission constraints. Section 5 concludes the present study with a
further discussion on the applicability of our POMINO approach for a fast
global retrieval from future fine-resolution satellite instruments.
Methodology
General process to retrieve tropospheric NO2 VCDs
The OMI is a nadir-viewing imaging spectrometer onboard the polar-orbiting
Sun-synchronous EOS Aura satellite with an Equator crossing time at 13:45
(Levelt et al., 2006). For each of the 14 or 15 orbits per day, the sensor
measures UV-visible radiation reflected by the Earth from 60 across-track
pixels with a swath of 2600 km. The pixel size is small and varies with the
viewing angle (from 13 km × 24 km at nadir to
25 km × 150 km at the swath edge). The OMI pixels cover the globe
on a daily basis, but the coverage of valid data is reduced by cloud and
snow/ice contamination and by recent row anomaly issues (especially since
2009). Row anomaly affects the quality of the level 1B radiance data for some
viewing directions of OMI
(http://www.knmi.nl/omi/research/product/rowanomaly-background.php).
Retrieval of tropospheric NO2 VCDs from satellites normally undergoes a
three-step process (Boersma et al., 2011). The first step derives slant
column densities (SCDs) from satellite radiance data, and the second step
separates the contribution of the tropospheric from the stratospheric part of
the SCD. The final step involves an air mass factor (AMF) calculation to
derive tropospheric VCDs (i.e., VCD = SCD / AMF). The AMF calculation is
affected by surface reflectance, aerosol optical effects, cloud fraction
(CF), cloud top pressure (CP), and atmospheric profiles of pressure,
temperature and NO2 (Zhou et al., 2010; Boersma et al., 2011; Bucsela et
al., 2013; Lin et al., 2014b). Accurate knowledge of these parameters is an
important prerequisite for the NO2 retrieval. Over polluted regions like
many parts of China, the AMF calculation is the dominant error source of
retrieved tropospheric NO2 data. Hereafter, SCDs, VCDs and AMFs are
referred to as their tropospheric portions.
Before the main NO2 retrieval itself, CF and CP are normally retrieved
with the O2–O2 approach (Acarreta et al., 2004). Here, clouds are
treated simply as an isotropically reflecting surface at a certain level (CP)
with a Lambertian albedo of 0.8. As such, cloud optical properties are
constrained by the two parameters CF and CP that can be retrieved from OMI
data. Although O2–O2 cloud parameters are retrieved in a
consistent manner with the latest DOMINO v2 NO2 retrievals (same surface
albedo assumption, same radiative transfer model, same cloud model; Boersma
et al., 2011) to ensure that the effective radiative properties of the scene
are consistent between the cloud and NO2 retrievals, some
inconsistencies have recently come to light with respect to different
atmospheric pressure and temperature profiles and terrain heights
(Maasakkers, 2013; Lin et al., 2014b).
Our POMINO retrieval approach
In this paper, our POMINO algorithm, referred to as case REF, largely follows
the method described by Lin et al. (2014b) with a few modifications. Here we
present a brief summary of the algorithm, placing emphasis on the latest
modifications. The reader is referred to Lin et al. (2014b) for a detailed
description. Our retrieval is focused on the derivation of tropospheric AMFs
to calculate tropospheric VCDs, taking the tropospheric SCD data (Dirksen et
al., 2011) from DOMINO v2 (Boersma et al., 2011). We adjust the calculated
layer AMFs to correct for the temperature dependence of the NO2
absorption cross section that is not accounted for in the SCD data (Boersma
et al., 2004). Following Lin et al. (2014b), we re-retrieve the prerequisite
(O2–O2-based) cloud properties by using a set of parameters
consistent with those in the retrieval of NO2. Our cloud retrieval is
focused on AMF calculations, starting with the O2–O2 SCDs from the
official cloud product OMCLDO2 v3 (Acarreta et al., 2004). Table 1 summarizes
our retrieval approach and parameters. Figure 2 briefly illustrates the
retrieval procedure.
Key tools and parameters used in POMINO and associated cloud
retrievals.
POMINO NO2 retrieval
Cloud (O2–O2)
AMF package
AMFv6; OpenMP parallelization
Same
RTM
LIDORT v3.6 (un-polarized, curved atmosphere); OpenMP parallelization
Same
Calculation for individual pixels
Pixel-specific radiative transfer modeling; no look-up table
Same
Surface reflectance
Land and turbid coastal ocean: BRDF at 440 nm, MCD43C2 Collection 5 (0.05∘); open ocean: OMLER v3 albedo
Same
Surface pressure
GEOS-5a; adjusted by elevationb
Same
Cloud fraction and cloud pressure
Derived here
–
Aerosol optical parameters
GEOS-Chem v9-02a; at 438 nm; model AOD is adjusted by MODIS Aqua AODc
Similar but at 475 nmc
Vertical profile of NO2
GEOS-Chem v9-02a,d
–
Vertical profiles of pressure and temperature
GEOS-5a,d
Same
a Resolution: 0.667∘ long. × 0.5∘ lat. horizontally
with 47 layers vertically and ∼ 10 layers below 1.5 km.
b Elevation information is from the GMTED2010 data set at
30 arcsec (http://topotools.cr.usgs.gov/GMTED_viewer/).
c See Sect. 2.3 for details.
d The pressure levels are re-calculated according to the
elevation-adjusted surface pressure, while the volume mixing ratios of
NO2 are not changed in individual layers (Zhou et al., 2009).
A diagram to illustrate our reference retrieval (case REF,
representing our POMINO product). Here Rs is surface reflectance, P is air
pressure, and T is air temperature. See Sect. 2.2 and Table 1 for detailed
descriptions of our retrieval approach and parameters.
We calculated the AMFs for O2–O2 (to derive CF and CP; at 475 nm)
and NO2 (at 438 nm) by using a newly improved AMFv6 package (Palmer et
al., 2001; Martin et al., 2003; Lin et al., 2014b) to coordinate the
retrieval process. The AMFv6 code makes use of the LIDORT v3.6 radiative
transfer model (RTM) (Spurr, 2008). Explicit radiative transfer is calculated
pixel by pixel. Both the AMFv6 and LIDORT v3.6 codes have been parallelized
via OpenMP. The parallelization has little overhead, in that, by using 16
computational cores in parallel, the speed-up is close to a factor of 16
relative to the single-core performance. With this speedup, a pixel-specific
radiative transfer calculation becomes feasible for a large-scale (large
domain, long time) retrieval, as in the present study. The traditional use of
a look-up table to interpolate the AMFs is then no longer needed.
We used data for vertical profiles of NO2, pressure and temperature on a
relatively high-resolution grid (0.667∘ long. × 0.5∘
lat.). The NO2 data were simulated by the GEOS-Chem chemical transport
model (CTM), and pressure and temperature data were taken from the GEOS-5
assimilated meteorological fields that were used to drive GEOS-Chem
simulations. GEOS-Chem has been shown to capture vertical profiles of
NO2 and ozone over the United States from aircraft measurements (Lin and
McElroy, 2010). Appendix A summarizes the CTM simulations. As we retrieve
clouds and NO2 pixel by pixel, model information at the grid cell
covering the pixel center is used. Although the size of our model grid cell
is larger than the size of an OMI pixel, our model grid cell size is much
smaller than that used in other OMI products [3∘
long. × 2∘ lat. for DOMINO (Boersma et al., 2011) and
2.5∘ long. × 2∘ lat. for OMNO2 (Bucsela et al.,
2013)]. In addition, we adjust the pressure profile for each pixel based on
the difference between pixel-specific surface elevation and grid cell average
elevation (Zhou et al., 2009; Lin et al., 2012, 2014b). The meteorological
and particularly NO2 profiles are subject to errors (Boersma et al.,
2011; Lin et al., 2014b). Further research is needed to evaluate these
profiles using available measurements over China.
Our retrieval explicitly accounts for the effects of spatiotemporally
varying aerosols and surface reflectance anisotropy on radiation. These two
factors have proved relevant for the NO2 retrieval (Zhou et al., 2010;
Lin et al., 2014b; Noguchi et al., 2014). Detailed information is presented
in Sects. 2.4 and 2.5.
Pixel selection, pixel-to-grid conversion, and sensitivity retrievals
For our present analysis, an OMI pixel is selected only when the ground is
free from snow and ice, and when the cloud radiance fraction (CRF) does not exceed 50 % (Boersma et al., 2011; Lin et al.,
2014b). Pixels with row-anomaly contamination are discarded. Data from valid
pixels are then converted to monthly mean values on a 0.25∘
long. × 0.25∘ lat. grid through an area-weighted
interpolation process. This process has been applied to NO2 and all
associated parameters. Section 3.5 discusses the number of days per month
with valid pixels on the gridded map.
In addition to our POMINO retrieval (case REF), three other retrievals
(cases SRF, AER, and S_A) were performed to evaluate the
sensitivity of retrieved NO2 VCDs and associated cloud parameters to
changes in aerosols and surface reflectance. These additional retrievals,
together with the standard DOMINO v2 product (referred to as case DOM)
(Boersma et al., 2011), were compared with case REF. These tests are
summarized in Table 2.
Retrievals of clouds and NO2 with different approaches∗.
REF = POMINO
SRF
AER
S_A
DOM
Surface reflectance
MCD43C2 BRDF
OMLER v3 albedo
MCD43C2 BRDF
OMLER v3 albedo
OMLER v1 albedo
Aerosol treatment
Explicit
Explicit
Implicit
Implicit
Implicit
∗ Retrieval procedures and parameters not mentioned here are the
same for cases REF, SRF, AER and S_A. Case DOM is adopted
from DOMINO v2.
There are notable differences in the representation of CRF between POMINO
and DOMINO. For POMINO, the CRF represents the fraction of the TOA radiance
caused by clouds alone (in the context of additional contributions from the
surface and aerosols). For DOMINO, however, the CRF applies to the fraction
of TOA radiance caused by both clouds and aerosols, with surface reflectance
represented by a geometry-independent surface albedo.
Different retrieval approaches lead to distinctive CRF values, which in turn
has consequences for the selection of valid data (Lin et al., 2014b) (see
discussions in Sect. 3.5). In Sects. 2 and 3, the pixels designated as
“valid” by case REF are selected for analysis, regardless of their
validity status in other retrievals. This choice ensures that the same set
of pixels is evaluated for all retrieval methods. For the emission
constraint study in Sect. 4, different sets of valid pixels specific to the
individual retrieval approaches are also analyzed, in addition to the set
determined by case REF.
Surface reflectance in POMINO (case REF)
Accurate knowledge of surface reflectance is of key importance for retrievals
of NO2 and ancillary cloud parameters. Surface reflectance depends both
on the ground conditions and on the state of the overlaying atmosphere (the
latter determining the relative amounts of diffuse vs. direct incident
radiance) (Lucht et al., 2000). Due to inhomogeneity in surface conditions,
the amount of reflected radiance relative to a given amount of incident
radiance depends on the incoming and outgoing angles. The degree of angular
dependence is determined by surface roughness and vegetation characteristics
(type, leaf area, and geometric shape). The angular dependence is most
prominent for direct incident radiance; angular effects on the reflected
radiation field are largely cancelled out for diffuse (isotropic) incident
radiance (Lucht et al., 2000). Normally the angular dependence is not
accounted for in the NO2 retrievals, with a few exceptions, e.g., Zhou
et al. (2010) for Europe, Noguchi et al. (2014) for Tokyo, and Lin et
al. (2014b) for several locations in China.
Case REF explicitly accounts for the angular dependence of surface
reflectance. It adopts the MODIS BRDF product as an approximate realization
of the complex dependence of surface reflectance on radiation geometry (Lucht
et al., 2000). This product is based on BRDF models using a linear
combination of three near-independent reflecting kernels: isotropic (no
angular dependence), volumetric (related to leaf area), and geometric
(related to vegetation shape). Each kernel contribution is regulated by a
parameter that varies with time and space. We used the snow- and ice-free
MCD43C2 Collection 5 data set (Lucht et al., 2000) that provides the kernel
parameters over land and turbid coastal ocean at 440 nm as 16-day average values on a 0.05∘
long. × 0.05∘ lat. grid. Kernel parameters are updated every
8 days, accounting for the seasonal and interannual variability in BRDF. To
cover missing values and reduce noise, the high-resolution data were passed
through a spatiotemporal interpolation and smoothing procedure. The data were
then mapped to each OMI pixel for subsequent AMF calculations. A detailed
description of these customized procedures is presented in Lin et
al. (2014b). Figure S1a–c in the Supplement presents the horizontal
distributions of BRDF kernel parameters on the original 0.05∘
long. × 0.05∘ lat. grid for the time period of 25 June
2012–10 July 2012.
Over China, the angular dependence of surface reflectance is more important
over parts of the west, southwest, southeast and northeast having complex
terrains (Fig. 1a), vegetated lands (Fig. 1b), and relatively low aerosol
loadings (Fig. 3, first row). Over the NCP and many other polluted regions
of China, high aerosol loadings (Fig. 3, first row) mean that most incident
radiation to the ground is diffuse, so that the angular dependence is
reduced for the radiance reflected from the ground. Nonetheless, the
seasonal and interannual variability in surface reflectance may still be
important for China, especially considering rapid land use change due to
urbanization, industrialization and agricultural activities (Liu et al.,
2014).
Over the open oceans, there are no applicable BRDF data; therefore, we used the
surface albedo data from the OMLER v3 product on a 0.5∘
long. × 0.5∘ lat. grid (Kleipool et al., 2008). This albedo
data set is a 5-year (2005–2009) mean monthly climatology, an update of
OMLER v1 adopted by DOMINO v2. The oceanic pixels are included for
completeness in our product; exclusion or inclusion of these pixels has
little effect on our present analysis. Figure S1d in the Supplement shows the
OMLER v3 data in July.
Figure 4 (first row) presents the horizontal distribution of annual and
seasonal bi-directional reflectance factor (BRF) values in 2012, as
representative of MODIS BRDF over land and turbid coastal ocean and OMLER v3 albedo over the open ocean.
The data have been sampled from valid OMI pixels and mapped to a
0.25∘ long. × 0.25∘ lat. grid, as described in
Sect. 2.3. The BRF here is the ratio of reflected radiance to the π-divided direct incident irradiance (along the geometric light path from the
Sun to the ground and then to the OMI) (Schaepman-Strub et al., 2006).
Especially for incident radiation with strong direct and weak diffuse
contributions, the value of BRF is critical for the total amount of radiation
received by OMI. The BRF data in Fig. 4 (first row) imply that the spatial
and temporal variations in solar zenith angle are taken into account, in
addition to changes in the ground characteristics. The choice of BRF for
presentation purposes follows Zhou et al. (2010) and Lin et al. (2014b).
Figure 4 (first row) shows that in all seasons, the BRF reaches maximum
values of 0.1–0.3 over western and northwestern China with desert or bare
land. The minimum values at 0.02–0.04 occur in parts of eastern China where
there is stronger absorption by vegetation (Fig. 1b). Especially over eastern
China, the relatively high (low) BRF values are often coincident with high
(low) AOD (comparing Figs. 3 and 4, first rows), likely indicating the
presence of aerosol contamination in the BRDF data. The BRF data exhibit
significant seasonal variation. Over eastern China, the BRF often reaches
maximum values in summer and minima in fall. For example, over much of the
NCP, the BRF varies from 0.06–0.08 in summer to 0.04–0.06 in fall. Over the
west, the summer season has the lowest BRF values, likely a result of a lower
solar zenith angle.
Aerosol optical properties in POMINO (case REF)
POMINO explicitly accounts for the optical effects of aerosols, given the
high aerosol loadings over China (Fig. 3, first row). Lin et al. (2014b) has
a detailed description of the implementation of aerosol optical properties in
the retrieval process. Here we emphasize the modifications to POMINO needed
to facilitate a large-scale retrieval (i.e., for a large domain in all
seasons, as compared to several spot locations investigated by Lin et al.,
2014b).
Day-to-day varying aerosol optical properties (AOD, single scattering albedo
(SSA), phase functions, and vertical profiles) are taken from the GEOS-Chem
v9-02 simulations on a 0.667∘ long. × 0.5∘ lat. grid.
The model is updated from an earlier version (v8-03-02) used by Lin et
al. (2014b). See Appendix A for model descriptions. GEOS-Chem simulates
various aerosol types, including secondary inorganic aerosols (sulfates,
nitrates, and ammoniums), organic aerosols, black carbon, dust and sea salts.
For a given OMI pixel, aerosol data at the grid cell covering the pixel
center are used during the retrieval process. Aerosol optical properties at
two wavelengths are implemented to retrieve NO2 (438 nm) and clouds
(475 nm), respectively.
To constrain the AOD, CTM-modeled AOD at 550 nm is adjusted to match MODIS
Aqua data on a monthly basis (Appendix B). The AOD adjustment is then carried
over to other wavelengths (438 and 475 nm) based on species-specific size
distributions, refractive indices and hygroscopic growth rates as assumed in
GEOS-Chem. The same procedure was used in Lin et al. (2014b).
Figure 3 (first row) presents the horizontal distribution of AOD at 550 nm
used in POMINO. Both annual and seasonal mean data are shown. High AOD values
are apparent over the NCP (the annual mean is 0.8–1.0), the Sichuan Basin
(0.8–1.0), and parts of southern China (0.6–0.8) due to significant
anthropogenic sources. The high AOD values over the Sichuan Basin are also a
result of a long-lasting stagnant atmosphere. Large AOD values are also
present over the western deserts, especially in spring (0.9–1.2), which has
the highest dust emissions. Over eastern China, AOD values are higher in
spring and summer than in fall and winter. These spatial and temporal
patterns are generally consistent with previous findings (Xia et al., 2007;
Xin et al., 2007; Wang et al., 2011; Lin et al., 2014c).
AOD and SSA at annual and seasonal scales. Provincial boundaries of
China are shown. Data are sampled from valid pixels of case REF. AOD values
exceeding a value of 1.2 are shown in black. Missing values are shown in
grey. Filled circles in the second row indicate the SSA estimates for 2005
with AOD > 0.4 by Lee et al. (2007). The embedded numbers are our
SSA values minus Lee et al. (±1 standard deviation); our data are
sampled at grid cells covering their sites.
Figure 3 (second row) shows the SSA at 550 nm. The SSA is largest over
western and northwestern China, where there are few black carbon sources.
Over the west, the SSA varies between 0.92 and 0.98 in all non-winter
seasons. In winter, the SSA is reduced to 0.90–0.92 over large parts of
Xinjiang. Over eastern China, the SSA experiences even larger seasonal and
spatial variability, from ∼ 0.80 over parts of the NCP in winter to
0.94–0.96 over most of eastern China in summer. The seasonality of SSA is
mostly a consequence of black carbon emissions reaching their maximum values
in winter and minimum amounts in summer (see Appendix A for the
implementation in GEOS-Chem).
Figure 3 (second row, filled circles) shows SSA values for 2005 estimated by
Lee et al. (2007) from MODIS top-of-atmosphere radiance and ground AOD
networks. Their estimates are only for situations with AOD > 0.4,
and have a root mean square error of 0.03. Differences between our SSA values
and Lee et al. are highly season- and location-dependent. On a seasonal
basis, our SSA values, sampled at grid cells covering their sites, are most
consistent with Lee et al. in winter (mean difference across China is
-0.02 ± 0.05), followed by spring, summer and fall. Note that these
comparisons are qualitative, given the inconsistency in data sampling.
Several limitations constrain our ability to improve aerosol modeling. Model
aerosol optical properties (AOD, SSA, phase functions) and vertical profiles
are subject to errors (Drury et al., 2010; Ford and Heald, 2012; van
Donkelaar et al., 2013). We used MODIS AOD data to constrain CTM-derived
AOD, even though MODIS data are not free of errors (Wang et al., 2007; Wang
et al., 2010; Hyer et al., 2011). No adequate observations are available to
constrain other aerosol optical parameters at a regional scale with high
spatial and temporal resolutions. Observation-based estimates of SSA are
essentially lacking at the scale considered here, and the few results in the
literature contain large uncertainties (±0.03) (Lee et al., 2007).
Although the CALIOP instrument provides information of aerosol vertical
profiles (Winker et al., 2009), the CALIOP profiles are limited by their
spatiotemporal coverage and data quality (especially near the ground) (Ford
and Heald, 2012; van Donkelaar et al., 2013). Note that since the same
vertical mixing and convection schemes were used to simulate aerosols and
NO2, the height of aerosols relative to NO2 (relevant to our
study) may be subject to smaller errors than the absolute height of
aerosols. Future work is needed to better understand and constrain aerosol
properties and evaluate how they affect the NO2 retrieval.
First row: surface reflectance in case REF = POMINO (MODIS BRF
data) at annual and seasonal scales. Second row: surface reflectance in case
SRF (OMLER v1 albedo) minus reflectance in case REF (MODIS BRF). Provincial
boundaries of China are shown. Data are sampled from valid pixels of case
REF. Values outside the upper (lower) bound of color intervals are shown in
black (purple). Missing values are shown in grey. Color intervals are
nonlinear to better present the data range; an interval without labeling
represents the mean of the adjacent two intervals.
Nevertheless, our present study, at the very least, reveals the importance of
an explicit aerosol treatment for NO2 and associated cloud-parameter
retrievals at a regional scale, especially given the lack of such an explicit
treatment in current satellite products. In support of our work here, Lin et
al. (2014b) showed that, by explicitly accounting for aerosols with just the
AOD values constrained by observations, there is excellent correlation
between retrieved NO2 VCDs and independent MAX-DOAS data (R2= 0.96 in day-to-day variability across the few locations being studied).
Section 3.3 further shows large changes in retrieved NO2 VCDs from an
explicit to an implicit treatment of aerosols, and Sect. 4 illustrates the
consequences for subsequent NOx emission constraint. Therefore, we
expect that the explicit inclusion of aerosols will improve the NO2
retrieval, especially if more comprehensive observations become available to
constrain model aerosols.
OMI NO2 retrievals and complex influences
of aerosols and surface reflectance
General characteristics of POMINO NO2 (case REF)
Figure 5 (first row) shows annual and seasonal mean NO2 VCDs from POMINO
(case REF) on a 0.25∘ long. × 0.25∘ lat. grid. At
this fine resolution, hotspots of NO2 pollution across China are clearly
visible. The formation of hotspots is also a result of the short lifetime of
NO2 (2–3 h to 1 day, depending on chemical activity). Large pollution
covers much of the NCP, with annual mean NO2 VCDs exceeding
15 × 1015 cm-2, due to emissions from both urban and
regional sources. Annual mean NO2 VCDs are below 1015 cm-2
over much of western China due to a lack of anthropogenic influences. The
spatial gradient of NO2 pollution is greatest in summer due to its
shortest lifetime, but the NCP still has a large inter-connected area of high
NO2. This regional-scale high pollution highlights the severity and
extensiveness of China's environmental problems.
First row: tropospheric NO2 VCDs retrieved from case REF at
annual and seasonal scales. Second–fifth rows: changes in NO2 VCDs from
case REF to other cases as a percentage fraction of case REF. Provincial
boundaries of China are shown. Data are sampled from valid pixels of case
REF. Values outside the upper (lower) bound of color intervals are shown in
black (purple). Missing values are shown in grey. Color intervals are
nonlinear to better present the data range; an interval without labeling
represents the mean of the adjacent two intervals.
Figure 5 (first row) also shows a large seasonal variation in NO2. Over
eastern China, NO2 VCDs reach maxima in winter and minima in summer. The
maximum to minimum ratio is about 3.6 for all regions east of
101.25∘ E. The large seasonality in NO2 VCDs mainly reflects
the seasonality in the species' lifetime (Lin, 2012). Over most of western
China with few anthropogenic emissions, NO2 VCDs are largest in summer
due to a peak in natural (lightning and soil) sources that overcompensates
for the shortest lifetime. This seasonal pattern is most notable over Tibet
and Qinghai. For western China (west of 101.25∘ E) as a whole, the
ratio of summer peak to winter minimum is about 1.4. Over much of Xinjiang
and Inner Mongolia, the growth in anthropogenic influences has meant a winter
maximum and a summer minimum, reversing the seasonality typical for western
China.
Figure 5 (fifth row) presents the difference in NO2 VCDs between the
cases DOM and REF, as a percentage fraction of REF. Case DOM is taken from
DOMINO v2 (Boersma et al., 2011) and sampled from pixels valid in case REF,
irrespective of whether these pixels are flagged as valid in the DOMINO v2
product. (Note that in Sect. 4, valid pixels from case REF and DOMINO v2 are
both evaluated for the derivation of NOx emissions.) Figure 5 (fifth
row, first panel) shows that at an annual scale, results from case DOM exceed
those from REF by 0–60 % over central eastern China (consistent with
POMINO NO2 columns being ∼ 45 % lower than DOMINO, as reported
in Lin et al., 2014) and much of the west. Case DOM results are smaller than
those from REF over parts of the south and north. Seasonal dependence is
significant (Fig. 5, fifth row). Over the NCP, case DOM is greater than REF
by 10–40 % in summer, while the signs of difference are
location-dependent in winter. Over most of Tibet, case DOM is similar to REF
in the fall season, but greatly exceeds REF (by up to 40 %) in other
seasons. These differences reflect the dissimilar AMF approaches in the two
retrievals. It is beyond the scope of this study to fully elucidate the
differences between case REF (POMINO) and DOMINO v2. Instead, the following
sections analyze the effects of surface reflectance and aerosols on retrieved
NO2 VCDs.
Effects of surface reflectance on NO2 and cloud retrievals
The MODIS BRDF data account for spatial and temporal variability in surface
reflectance as well as its angular dependence. Here we evaluate the
sensitivity of retrieved NO2 VCDs to surface reflectance, by adopting an
alternate surface reflectance data set to repeat the retrieval process.
Case SRF adopts the OMI monthly climatological albedo data (OMLER v1, at
440 nm) from the DOMINO v2 product and re-derives cloud properties and
NO2 VCDs; other retrieval procedures, including aerosol treatments, are
unchanged. Unlike the MODIS BRDF data, the OMI albedo data are monthly climatology
(October 2004–October 2007 average) with no interannual variability. The OMI
albedo data set means isotropic reflectance with no angular dependence. The
horizontal resolution of OMI albedo data is 0.5∘
long. × 0.5∘ lat., compared to the high-resolution MODIS BRDF
data (at 0.05∘ long. × 0.05∘ lat.). The use of OMI
albedo data tests the sensitivity of retrieved NO2 to large changes in
surface reflectance.
Figure 4 (second row) compares the OMI albedo with the MODIS BRF over China.
The OMI albedo is normally within ±0.05 of the MODIS BRF. Over most of
eastern China, the OMI albedo exceeds the MODIS BRF in all seasons with a
difference of 0.01–0.04. Over northeastern China, however, the OMI albedo is
lower than the MODIS BRF in spring and winter. Over most of western China,
the OMI albedo is smaller than the MODIS BRF with a difference of 0.01–0.06.
The OMI albedo greatly exceeds the MODIS BRF at various locations in the west
and north (by 0.10 or more).
The diagram in Fig. 6 illustrates how a change in surface reflectance affects
the pre-NO2 cloud retrieval. In the cloud-property O2–O2
algorithm, higher reflectance leads to lower effective CF, since fewer clouds
are needed to reflect a given amount of radiation to the outer space. Effects
of changing surface reflectance on CP are multifold. Higher reflectance means
an enhanced AMF in the clear-sky portion of the OMI pixel (AMFcr), which can
be compensated for by a decrease in CP. In addition, the reduction in CF
caused by enhanced surface reflectance has a secondary effect on CP. A
decrease in CF may lead to a further reduction in CP if the AMF of
O2–O2 in the cloudy portion of the OMI pixel, AMFcl, is smaller
than AMFcr (this is a “screening” effect of clouds on radiation). This
effect occurs in most situations where the cloud top is distant from the
ground and there is alow above-cloud O2–O2 concentration.
Occasionally, a decrease in CF may result in an enhancement in CP, when the
cloud top is close to the ground and thus the value of AMFcl exceeds AMFcr
(an “albedo” effect of clouds on radiation).
A diagram of how changes in surface reflectance (Rs), aerosol
scattering (ASOD) and aerosol absorption (AAOD) affect the retrievals of CF
and CP in the O2–O2 algorithm. Grey colors illustrate the indirect
influence of CF changes on the CP; the influence depends on the “screening”
or “albedo” effects of clouds on radiation. Thin grey lines indicate that
the “albedo” effect of clouds is occasional. The figure is updated from
Fig. 5 of Lin et al. (2014b).
Figure 7 contrasts case REF and SRF for the CF, cloud radiance fraction (CRF)
and CP at an annual scale. A negative correlation is apparent between changes
in CF (and CRF) from case REF to SRF and changes in surface reflectance. Over
land, the CP tends to change in the opposite direction to the change in
surface reflectance. Over much of the oceans away from Chinese coasts,
however, the CP decreases with declining surface reflectance from case REF to
SRF; this is due to the “albedo” effect of clouds (CP = 800–900 hPa).
Overall, the CP changes from case REF to SRF with an opposite sign as surface
reflectance for about 78 % of all grid cells and months. Changes are
similar across different seasons for individual cloud parameters
(Figs. S2–S4 in the Supplement, second rows).
Annual mean (a) cloud fraction, (b) cloud top
pressure and (c) cloud radiance fraction retrieved via different
approaches. Provincial boundaries of China are shown. Data are sampled from
valid pixels of case REF. Values outside the upper (lower) bound of color
intervals are shown in black (purple). Missing values are shown in grey.
Color intervals in (a) are nonlinear to better present the data
range; an interval without labeling represents the mean of the adjacent two
intervals.
As shown in Eq. (1) (the IPA or independent pixel approximation), changes in
CRF affect the relative weights of NO2 AMFcl vs. AMFcr, while changes in
CP affect the absolute magnitude of NO2 AMFcl. In addition, an increase
in surface reflectance results in an enhancement in NO2 AMFcr. These
factors together determine the effects of surface reflectance on the NO2
retrieval.
AMF=AMFcl⋅CRF+AMFcr⋅(1-CRF)VCD=11VCDcl⋅CRF+ 1VCDcr⋅(1-CRF)
Figure 5 (second row) shows that the effects of surface reflectance on
retrieved NO2 VCDs are largely region- and season-dependent. Over most
Chinese regions except central and southeastern China, replacing the MODIS
BRDF with OMI albedo tends to increase the retrieved NO2 VCDs by
0–40 % (mostly less than 15 %). Over central and southeastern China,
however, the use of OMI albedo reduces NO2 VCDs by 0–20 % in
spring, fall and winter, but with a slight enhancement in summer. This
spatial and seasonal divergence in NO2 changes reflects the complex
influences of surface reflectance on retrieved cloud properties and NO2
AMFs. The magnitude of NO2 changes is comparable to previous studies
(Zhou et al., 2010; Lin et al., 2014b).
Figure 8b presents the percentage changes in NO2 VCDs from case REF to
SRF as a function of AOD values (binned at intervals of 0.05) and changes in
surface reflectance (i.e., OMI albedo minus MODIS BRF, binned at intervals
of 0.01). Here the percentage changes from all grid cells and months with
respect to each bin of AOD and surface reflectance change are averaged; the
frequency of data located in each bin is shown in Fig. 8a. Figure 8c and d
also separate the effects of surface reflectance on NO2 VCDs for the
pixel clear-sky and cloudy portions. Here, a clear-sky VCD (VCDcr) is
derived as the SCD divided by AMFcr, and a cloudy-sky VCD (VCDcl) represents
the SCD divided by AMFcl; Eq. (2) shows the relation between VCD, VCDcr
and VCDcl. Pixels with no clouds are excluded from calculations of AMFcl and
VCDcl. The VCDcr and VCDcl for individual pixels are aggregated to
respective monthly mean values in the same way as was done for the total
VCD.
(a) Frequency of occurrence from all months and grid cells
for each bin of AOD (x axis, bin size 0.05) and surface reflectance
difference (OMI albedo in case SRF minus MODIS BRF in case REF, y axis, bin
size 0.01). Data are sampled from valid pixels of case REF.
(b) Percentage changes in NO2 VCDs from case REF to SRF
averaged over all data in each bin of AOD and surface reflectance difference;
also embedded are the correlations between percentage VCD changes and AOD
values as a function of surface reflectance differences (black line) and
between VCD changes and surface reflectance differences as a function of AOD
values (red line). Similar panels are drawn for (c) clear-sky VCDs
and (d) cloudy-sky VCDs.
Figure 8c shows that, when the OMI albedo is larger (smaller) than the MODIS
BRF, we get a reduction (enhancement) in VCDcr. The relative changes in
VCDcr are within 40 % (mostly < 10 %), and a greater change in
surface reflectance tends to result in a larger change in VCDcr. In any
given AOD bin, the magnitude of correlation between changes in surface
reflectance and changes in VCDcr exceeds 0.8 (embedded red line). For a
given bin of surface reflectance, by comparison, there is no apparent
dependence of VCDcl changes on the amounts of AOD. The effects of surface
reflectance on NO2 VCDs are similar to the effects on VCDcr (Fig. 8b)
due to the low CRF values on average (Fig. 7c, second column).
Figure 8d shows complex effects of surface reflectance perturbations on
VCDcl. This is because the CP does not always change in the same direction as
surface reflectance (see discussions above). When the OMI albedo is within
±0.05 of the MODIS BRF, the VCDcl tends to increase from cases REF to
SRF. There is greater scatter in VCDcl changes when the changes in surface
reflectance are greater. In addition, VCDcl undergoes a much greater change
in magnitude than VCDcr, reflecting the strong sensitivity of AMFcl to the
CP.
Influences of implicit aerosol treatment on NO2 and cloud retrievals
Case AER tests the effects of aerosols on NO2 and cloud retrievals by
setting AOD to zero during the retrieval process, thus mimicking the
traditional treatment (Boersma et al., 2011; Bucsela et al., 2013). This
procedure leads to changes in CF and CP that implicitly affect the
subsequent NO2 retrieval. If this implicit treatment results in the
same NO2 VCDs as case REF, then an explicit treatment of aerosols is no
longer strictly desirable, as it is more expensive computationally.
Figure 7a and c show that at an annual scale, the exclusion of aerosols
results in significant enhancements in CF and CRF. This is because aerosols
reflect solar radiation to space, and exclusion of aerosols is compensated
for by an increase in effective CF and CRF (see Fig. 6) (Boersma et al.,
2011; Lin et al., 2014b; Castellanos et al., 2015). Over China, the CF is
enhanced from 0.04–0.15 in case REF to 0.06–0.30 in case AER. The increase
is greatest over parts of eastern China, with large AOD values in case REF.
The CRF in case AER is greater than 0.35 over most of eastern China, compared
to the values of 0.15–0.30 in case REF. Over the Sichuan Basin, the annual
mean CRF exceeds 0.50 in case AER, more than doubling the CRF value (about
0.25) in case REF. Overall, the correlation between the amounts of AOD
neglected in case AER and the level of CF (CRF) augmentation reaches a high
value of 0.75 (0.82). Analyses of individual seasons also show large
enhancements in CF and CRF as a result of neglecting aerosols (Figs. S2 and
S4 in the Supplement, third rows). Correlation between effective CF and AOT
has been found for the eastern United States (0.66) (Boersma et al., 2011)
and South America (Castellanos et al., 2015).
Figure 7b shows that excluding aerosols leads to an increase in CP from case
REF to AER, in order to compensate for an otherwise reduction in the
O2–O2 AMF. Over eastern China, the CP is increased from
700–900 hPa in case REF to 750–950 hPa in case AER. The CP enhancement is
smaller over western China, and is smallest (0–20 hPa) over the Tibetan
Plateau due to the lowest aerosol loadings. However, the correlation between
the amounts of AOD neglected in case AER and the amounts of CP increase is
only 0.19, reflecting the complex effect of aerosols on the CP retrieval (Lin
et al., 2014b). The CP enhancements are apparent in all seasons (Fig. S3 in
the Supplement, third row).
Figure 5 (third row) shows the horizontal distribution of percentage NO2
changes from cases REF to AER. In all seasons, case AER is larger than REF by
0–40 % over most of China. The overestimate is most obvious in central
and southern China, especially in winter (by 15–40 %). By comparison,
case AER leads to lower NO2 VCDs by 0–20 % over parts of the NCP in
spring, many places in the north in winter, and parts of the west in
non-summer seasons. At an annual scale, case AER leads to larger NO2
VCDs than case REF by 0–40 % over most regions. The magnitudes of
NO2 changes are weakly correlated with the AOD or SSA (comparing Fig. 5,
third row with Fig. 3). The degree of divergence in NO2 changes is
consistent with that found in Lin et al. (2014b).
The red solid lines in Fig. 9 present the AER to REF ratio for NO2 VCDs
on a monthly basis. Several representative regions of China are considered,
including eastern China, western China, northern eastern China, the Sichuan
Basin, Beijing–Tianjin–Hebei (BTH), the Yangtze River delta (YRD), the
Pearl River delta (PRD), and urban areas of Beijing, Shanghai and Guangzhou.
These regions are defined in Fig. 1a. Figure 9 shows that for NO2 VCDs
averaged over a large region such as eastern or western China, the AER to REF
ratio is close to 1 (0.9–1.1) in all months. For smaller regions, the
deviation in the AER / REF ratio increases dramatically. The ratios vary
between 0.8 and 1.4 across the 12 months for BTH, YRD and PRD, and between
0.8 and 1.6 for the urban areas of Beijing, Shanghai and Guangzhou. Whether
case AER leads to larger or smaller NO2 VCDs than case REF depends
strongly on location and season. For urban Beijing, the maximum AER / REF
ratio over the 12-month period is about twice as much as the minimum
AER / REF ratio. Case AER has larger NO2 VCDs than case REF in most
months over the Sichuan Basin and the PRD, while the AER / REF ratios are
more seasonally variable in other regions. These regional and seasonal
features call for comprehensive independent measurements to validate
satellite retrievals.
Ratios of cases AER and DOM to case REF for regional mean NO2
VCDs in each month of 2012. Regions are defined in Fig. 1a. For cases AER
and DOM, data can be sampled from pixels valid in respective retrieval
approaches (dotted lines) or valid in case REF (solid lines). Also shown are
some statistical quantities.
Figure 10b shows the percentage changes in NO2 VCDs from case REF to
case AER as a function of AOD and SSA. Here the percentage changes from all grid
cells and months with respect to each bin of AOD (bin size = 0.05) and SSA
(bin size = 0.01) are averaged; the amount of data in each bin is shown in
Fig. 10a. Whether a larger AOD value corresponds to a greater enhancement in
NO2 VCD from case REF to case AER depends in a complex manner on the
SSA. This effect is also found by Castellanos et al. (2015) for biomass
burning aerosols over South America. The largest increase (by 40–80 %)
from case REF to case AER occurs with high AOD of 1–2 and low SSA of
∼ 0.90. The dependence of NO2 changes on SSA is weak, with
positive correlations when AOD is below 1.5 and negative correlations when
AOD exceeds 1.5.
(a) Frequency of occurrence from all months and grid cells
for each bin of AOD (x axis, bin size 0.05) and SSA (y axis, bin size
0.01). Data are sampled from valid pixels of case REF.
(b) Percentage changes in NO2 VCDs from case REF to SRF
averaged over all data in each bin of AOD and SSA; also embedded are the
correlations between percentage VCD changes and AOD values as a function of
SSA values (black line) and between VCD changes and SSA values as a function
of AOD values (red line). Similar panels are drawn for (c) clear-sky
VCDs and (d) cloudy-sky VCDs.
Figure 10c presents the percentage changes in NO2 VCDcr as a function of
AOD and SSA. With a fixed AOD value, lower SSA tends to result in a smaller
increase or a large decrease in VCDcr from case REF to AER; for AOD of
0.2–1.3, the correlation between VCDcr changes and SSA values exceeds 0.7.
This is because stronger absorption diminishes the radiation that could
otherwise be absorbed by NO2, and leads to a consequent reduction in
AMFcr. The dependence of VCDcr changes on AOD is highly SSA-sensitive. In
many situations, the correlation between NO2 changes and AOD or SSA
values is weak (within ±0.5), owing to the indirect effects of aerosol
vertical profiles and other factors. These features reflect the complex
effects of aerosol extinction on the radiation absorbed by NO2 under
cloud-free conditions, as has been found in GEOS-Chem simulations (Lin et
al., 2012). This contrasts with the effect of changing surface reflectance on
NO2 VCDcr such that an increase in surface reflectance reduces the
amount of VCDcr (Fig. 8c).
Figure 10d shows that the effect of excluding aerosols on VCDcl differs
significantly from the effect on VCDcr. Case AER leads to lower VCDcl values
by 0–60 % for a wide range of AOD and SSA, mostly because the increase in
CP leads to a consequent enhancement of AMFcl.
We further elucidate how the exclusion of aerosols affects retrieved NO2
VCDs. Changes in NO2 VCD are determined by CRF, VCDcr and VCDcl (Eq. 2).
Excluding aerosols leads to a general increase in CRF (and thus the weight of
VCDcl) that is compensated for by a decrease in VCDcl (due to an increase in
CP). The change in VCDcr is more complex in sign, and to a lesser extent in
magnitude. In most situations, the magnitude of VCDcr is much smaller than
VCDcl (Fig. S5 in the Supplement); this is because clouds are normally above
the NO2-concentrated layer, producing a “screening” effect on the
radiation absorbed by NO2. These factors explain the distinctive
patterns of changes in NO2 VCDcr, VCDcl and VCD from case REF to AER.
For example, for AOD ∼ 1.5 and SSA ∼ 0.90, both VCDcr and VCDcl
are reduced from case REF to AER, while the VCD is enhanced because CRF is
much enhanced and VCDcl greatly exceeds VCDcr.
In summary, inclusion or exclusion of aerosols has distinctive effects on
three independent factors, CRF (or CF), VCDcr, and VCDcl (or CP). Section 3.5
will show the effect of excluding aerosols on the choice of “valid” data
based on the CRF criterion. It follows that an implicit treatment of aerosols
cannot fully account for the complex influences of aerosols on retrieved
NO2 VCDs.
Coupled effects of aerosols and surface reflectance on NO2 retrieval
This section evaluates the coupled effects of perturbing aerosols and surface
reflectance on retrieved NO2 VCDs. For this purpose, we use case S_A,
which simultaneously adopts the OMI albedo from DOMINO v2 (following case
SRF) and excludes aerosol information (following case AER) in the retrieval
process. For this case, cloud parameters are also re-derived.
Figure 5 (fourth row) shows that, at an annual scale, case S_A leads to
higher NO2 VCDs than case REF over central eastern China, southwestern
China, Tibet and Qinghai. NO2 VCDs are reduced by about 10 % over
parts of the NCP. Seasonal dependence is large for the NO2 changes.
Reductions over the NCP are most significant in spring (by 10–40 %),
followed by winter and fall. In summer, case S_A leads to general NO2
enhancements by 0–40 % over most of China, including the NCP. Over the
Sichuan Basin, NO2 VCDs in fall are increased by 20–40 % from case
REF to S_A, while the signs of change are more location-dependent in winter.
Figure 5 (fourth and fifth rows) shows that although case S_A follows the
DOMINO v2 assumptions regarding surface reflectance and aerosols, in general
it does not explain the differences between case REF and DOM. Therefore,
other factors are also important in differentiating case REF from DOM. They
include NO2 vertical profiles, cloud parameters, the pixel-specific
radiative transfer calculation, and air pressure (Lin et al., 2014b).
Figure 11 further shows that the effect of changing surface reflectance on
retrieved NO2 VCDs interacts with the effect of excluding aerosol
information. The figure presents the difference between (SRF – REF) + (AER
– REF) and (S_A – REF) as a percentage fraction of REF. Positive values
occur over most of China; i.e., the effect of simultaneously changing surface
reflectance and aerosols is smaller than the summed effect of changing the
two parameters individually. The magnitude of differences greatly depends on
seasons and locations, and is largest in winter (by 5–20 % over much of
eastern China) and smallest in summer. These differences reflect the
nonlinear influences of individual factors in retrieving NO2, such that
the effect of a particular parameter depends on other parameters (Lin et al.,
2014b).
Differences in NO2 VCDs between (SRF – REF) + (AER – REF)
and (S_A – REF) as a percentage fraction of REF. Provincial boundaries of
China are shown. Data are sampled from valid pixels of case REF. Values
outside the upper (lower) bound of color intervals are shown in black
(purple). Missing values are shown in grey. Color intervals are nonlinear to
better present the data range; an interval without labeling represents the
mean of the adjacent two intervals.
Discussion on the sampling of “valid” pixels and implications for analysis of pollution severity
Figure 12a shows the number of days per month with “valid” pixels based on
the current criteria for cloud cover, snow/ice cover, and row-anomaly
contamination. Averaged over China, fewer than half of the calendar days are
included for analysis. For most of northern China, about 12–20 days per
month of 2012 are selected. The number of available days per month is below 8
over much of the south due to higher cloud coverage.
(a) Number of days per month of 2012 with valid pixels in case REF
on a 0.25∘ long. × 0.25∘ lat. grid. (b) Changes in the
number of days per month with valid pixels from REF to other cases.
Provincial boundaries of China are shown.
Changes in aerosols and surface reflectance both have a consequence for CF
and CRF. The changes in CF and CRF in turn affect the number of retrieved
NO2 results that are determined “valid” under the usual criterion of
CRF < 50 %. With an implicit aerosol treatment (AER, S_A, and
DOM), the CRF < 50 % criterion designates pixels as invalid when
more than half of the TOA radiance is from the combination of clouds and
aerosols. With an explicit aerosol treatment (REF and SRF), however, the CRF
< 50 % criterion means that pixels are excluded if the clouds'
contribution to the TOA radiance exceeds 50 %. Thus, the soundness of
this criterion depends on how well the aerosol optical effects are quantified
in the retrieval process, a critical factor in our explicit and physically
more realistic aerosol treatment.
Case SRF has similar data coverage as case REF (Fig. 12b). Compared to case
REF, case AER discards an additional 2–10 days per month (or 15–60 %)
over most of eastern China because the aerosol-affected CRF exceeds 50 %
(Fig. 12c). This is consistent with the loss of 25 % found by Lin et
al. (2014b). Cases S_A and DOM also have much smaller numbers of valid days
than does case REF (Fig. 12d and e).
For analyses of all retrievals in Sects. 3.1–3.4, we have evaluated
retrieval results from the same set of pixels determined by case REF. In
practice, however, information is not available on the pixels that could
have been included in certain retrieval approaches but are instead discarded
in other approaches. In such situations, a question is raised whether
selecting different sets of satellite pixels will affect the evaluation of
pollution severity and its subsequent applications. Of particular interest
here is the loss of valid data by not explicitly accounting for aerosol
optical effects (case AER vs. REF).
Figure 9 contrasts NO2 VCDs in case AER derived from pixels “valid” in
case REF (solid red lines) vs. from pixels “valid” in case AER (dashed red lines). Average VCDs
for various representative regions of China are shown. A detailed horizontal
distribution of their differences can be inferred by contrasting Fig. 5
(third row) and Fig. S6 in the Supplement (third row). Figure 9 shows a
significant reduction in retrieved NO2 VCDs by using the reduced set of
“valid” data from case AER relative to the set determined by case REF. The
reduction is about 10 % averaged over eastern China, but exceeds 50 %
in some regions and months. Similar results are found for case DOM that also
treats aerosol optical effects implicitly (Fig. 9, solid green vs. dashed
green lines). These results arise because high aerosol loadings are often
coincident with high NO2 pollution, due to similar emission sources and
meteorological influences. Discarding pixels with high aerosols (that lead to
significant increases in CRF from case REF to AER and DOM) tends to exclude
high NO2 pollution situations as well (Lin et al., 2014b). This sampling
bias is critical for determining the severity of NO2 pollution from
satellite remote sensing.
Explicit inclusion of aerosols in the retrieval process (as in case REF or
POMINO) will avoid the sampling bias caused by removal of strong-pollution
situations, especially if aerosol information can be adequately constrained.
In publishing the POMINO data, we elect to include pixels with high aerosol
loadings, although we note that these pixels may contain larger uncertainties
for NO2 than those with low aerosol content, since we do not have fully
accurate aerosol information. This choice is supported in part by our
previous comparisons against MAX-DOAS NO2 data (Lin et al., 2014b) and
by the fact that the AOD values are relatively well constrained by MODIS on a
monthly basis. We publish the AOD and SSA values together with NO2 VCDs,
so that users can choose whether or not to include high-aerosol situations.
Impacts of aerosols and surface reflectance on OMI-based
NOx emission constraint
Sections 3.2–3.5 show the sensitivity of retrieved NO2 VCDs to various
treatments of aerosols and surface reflectance. This section further analyzes
the influences on subsequent NOx emission constraint, a popular
application of OMI data. NO2 VCDs retrieved via different approaches are
used to optimize NOx emissions. We focus on anthropogenic emissions in
China. As discussed in Sect. 3.5, the set of “valid” OMI pixels depends on
the retrieval approach. Thus, this section also identifies the effects of
selecting various sets of “valid” pixels.
Top-down emissions are derived for individual months in 2012 on a
0.25∘ long. × 0.25∘ lat. grid, by scaling the a
priori emissions (assumed in GEOS-Chem) with the ratio of OMI-derived to
CTM-modeled NO2 VCDs (Martin et al., 2003). Model NO2 values (at
0.667∘ long. × 0.5∘ lat.) are sampled at times and
locations coincident with valid OMI pixels, and are then re-mapped to the
0.25∘ long. × 0.25∘ lat. grid; see Fig. S7 in the
Supplement for the horizontal distribution of model NO2. The scaling
approach here assumes local mass balance (Leue et al., 2001). It does not
fully account for the horizontal transport of NOx or the dependence of
NOx lifetime on emissions (Turner et al., 2012); these two factors are
likely to result in a small difference (within 10 %) in top-down total
Chinese emissions (Lin, 2012). Also, the same scaling is applied to both
anthropogenic and natural emissions. Lin (2012) presents an approach that
takes advantage of the distinctive seasonality in various emission sources to
better distinguish anthropogenic from natural sources. They found similar
changes (from a priori to top-down) in anthropogenic and natural emissions
when summed over China.
The a posteriori emissions are calculated as a weighted average of a priori
and top-down emissions, by assuming normal distributions of errors in these
emissions. Following Lin (2012), errors in anthropogenic emissions are taken
as 60 % for a priori and 52 % for top-down for combined errors in
model simulations (∼ 40 %, Lin et al., 2012; Yan et al., 2014),
satellite NO2 retrievals (∼ 30 %, Boersma et al., 2011; Lin et
al., 2014b), and emission inversion procedures (∼ 12 %, Lin, 2012).
The same errors are assigned to all grid cells, following Lin (2012). This
leads to an error of 39 % in the a posteriori emissions. Although the
actual errors may be larger for individual locations, there is no such
detailed information for emission constraint.
(a) The a priori anthropogenic emissions of NOx on a 0.25∘
long. × 0.25∘ lat. grid. (b) The a posteriori emissions constrained by case
REF. (c) Differences between a priori and REF as a percentage fraction of REF.
(d–g) Changes in a posteriori emissions from REF to other cases as a percentage
fraction of REF; OMI pixels are selected only when valid in case REF. (h–k)
are similar to (d–g) but with respect to pixels valid in individual
retrieval approaches. In (c–k), areas with emissions below 0.5 × 1015 cm-2 h-1
are masked. Provincial boundaries of China are shown.
Values outside the upper (lower) bound of color intervals are shown in black
(purple). Color intervals are nonlinear to better present the data range; an
interval without labeling represents the mean of adjacent two intervals.
Figure 13b shows the horizontal distribution of annual mean a posteriori
emissions in 2012 derived from case REF. The spatial pattern is close to that
of the a priori emissions (Fig. 13a) with largest values located at cities.
The a posteriori emissions are lower than a priori values by 0–40 % over
most of eastern China, with enhancements at many high-emission hotspots and
over the southern coastal provinces (Fig. 13c).
Table 3 shows that for China as a whole, the a posteriori emissions in case
REF are 9.05 TgN yr-1, about 9.3 % smaller than the a priori
emissions at 9.78 TgN yr-1. For eastern China
(101.25–126.25∘ E, 20–46∘ N), the REF a posteriori
emissions for 2012 total 8.43 TgN yr-1, about 19 % higher than the
estimate for 2006 by Lin (2012). This increase likely reflects the recent
growth of NOx pollution in China after its recovery from the economic
downturn (Lin and McElroy, 2011).
A priori, top-down and a posteriori emissions of NOx∗.
China (80–130∘ E, 20–53∘ N)
Eastern China (101.25–126.25∘ E, 20–46∘ N)
REF
SRF
AER
S_A
DOM
REF
SRF
AER
S_A
DOM
A priori
9.78
9.78
9.78
9.78
9.78
9.11
9.11
9.11
9.11
9.11
With OMI data sampled at valid pixels specific to each case
Top-down
8.50
8.20
8.63
8.21
8.75
7.92
7.65
8.03
7.65
8.19
A posteriori
9.05
8.88
9.12
8.88
9.19
8.43
8.27
8.49
8.27
8.58
With OMI data sampled at valid pixels by REF
Top-down
8.50
8.23
8.91
8.53
8.80
7.92
7.68
8.33
7.97
8.27
A posteriori
9.05
8.90
9.28
9.06
9.22
8.43
8.29
8.66
8.46
8.63
∗ Units: TgN yr-1
Figure 13h–k present the percentage differences in a posteriori emissions
comparing results from case REF to those from the other retrieval cases. OMI
pixels are selected when they are deemed to be “valid” according to
criteria specific to individual retrieval cases; here each retrieval case has
a distinctive set of “valid” pixels. Locations with emissions lower than
0.5 × 1015 cm-2 h-1 (totaling 11 % of Chinese
emissions) are masked to highlight the polluted areas. Overall, the magnitude
of emission changes is highly region-dependent. Across China with few
exceptions, case SRF produces emissions similar to REF. Differences are much
larger between cases REF and AER. Compared to case REF, case AER produces
lower emissions by 5–10 % over the NCP but higher emissions by
5–30 % over many other regions; the greatest enhancements occur over the
Sichuan Basin and the PRD. Compared to REF and AER, case S_A results in
smaller emissions over the NCP. Case S_A produces emissions slightly larger
than case REF in the Sichuan Basin, PRD, and many other southern provinces.
By comparison, case DOM (exactly the same as DOMINO v2) leads to larger a
posteriori emissions than does case REF, by 0–40 % over most of China.
Table 3 shows that total Chinese emissions in all cases are within 3 % of
values from case REF; this closeness is due mainly to compensation between
region- and time-dependent positive and negative differences during
spatiotemporal averaging.
Figure 13d–g show the percentage differences in a posteriori emissions
comparing case REF to other retrieval cases, but now based on a single set of
“valid” pixels determined by case REF. Locations with emissions lower than
0.5 × 1015 cm-2 h-1 are masked to highlight the
polluted areas. Case SRF leads to similar emissions to case REF. Case AER
exceeds REF by 0–30 % over most regions, but with reductions by
0–5 % over parts of northern China. Case S_A is also larger than REF
over much of China. Case DOM exceeds REF by 0–40 % especially over the
Sichuan Basin and parts of northern China.
Differences between Fig. 13e–g and i–k are apparent – a result of
different pixel sampling. The differences over the NCP indicate that, since
the implicit treatment of aerosols in cases AER, S_A and DOM leads to
exclusion of situations with high aerosol loadings (and coincidently high
NO2 pollution; see Sect. 3.5), there is a consequent underestimate in
the a posteriori emissions.
Figure 14a, d and g shows the contrasts between cases REF and AER for the
ratio of maximum to minimum monthly a posteriori emissions. Locations are
shown only with emissions greater than
0.5 × 1015 cm-2 h-1. For case REF (Fig. 14a), the
max / min ratios range from 1 to 10, but are about 2–5 at most locations.
Figure 14d (based on REF-determined valid pixels) shows enhancements in the
max / min ratio from case REF to case AER over most of China. By comparison,
Fig. 14g (based on case-specific sets of valid pixels) shows that case AER
leads to much larger max / min ratios over most of southern China, while
the changes in max / min ratios are more location-dependent over the NCP.
Figure 14 further contrasts cases REF and AER for the months of maximum and
minimum emissions. For case REF, minimum emissions often occur in
May–September in the NCP, in contrast to December–March in many southern
areas (Fig. 14b). Maximum emissions occur in November–February over most of
China (Fig. 14c). Case AER leads to location-dependent shifts in the months
of maximum and minimum emissions (Fig. 14e, f, h, and i); furthermore, these
shifts are more apparent when OMI pixels are selected with respect to
individual retrieval approaches (Fig. 14h and i). These differences highlight
the seasonal and spatial dependence of constrained emissions on the NO2
retrieval approaches.
(a) The maximum to minimum ratio for monthly a posteriori
anthropogenic emissions in case REF, and the months of (b) minimum
and (c) maximum emissions. (d–f) Differences between cases
AER and REF; OMI pixels are selected only when valid in case REF.
(g–i) are similar to (d–f) but with respect to pixels
valid in individual retrieval cases. Provincial boundaries of China are
shown. Values outside the upper (lower) bound of color intervals are shown in
black (purple). In (d–i), areas with emissions below
0.5 × 1015 cm-2 h-1 are masked. In (e),
(f), (h) and (i), the month proceeds in loop;
i.e., November (January) is 1 month behind (ahead of) December.
Conclusions
This paper presents an improved retrieval, the POMINO algorithm,
of tropospheric NO2 VCDs over China in 2012, by explicitly accounting
for temporally and spatially varying aerosol optical effects and surface
reflectance anisotropy. Prerequisite cloud optical parameters (CF and CP) are
retrieved with the same treatments for aerosols and surface, thus eliminating
any algorithm inconsistency in retrieving NO2 and clouds. Our NO2
retrieval is focused on calculations of tropospheric AMFs, taking the
tropospheric SCDs from DOMINO v2. Our cloud retrieval starts from the SCDs of
O2–O2 in OMCLDO2 v3. Aerosol vertical profiles and optical
properties are first taken from GEOS-Chem simulations, with subsequent AOD
constraints based on monthly MODIS Aqua data. Surface reflectance data over
land are taken from the MODIS BRDF product. Retrievals are performed for
individual OMI pixels via online radiative transfer calculations without the
need for look-up tables, thanks to the newly parallelized AMFv6 code driven
by LIDORT v3.6. Sensitivity tests are performed to evaluate the effect of an
explicit (as opposed to an implicit) treatment of aerosols and the effect of
changes in surface reflectance characteristics on retrieved NO2 VCDs.
Additional analyses are done for subsequent OMI-based NOx emission
constraints. Results are presented as monthly mean values on a 0.25∘
long. × 0.25∘ lat. grid.
POMINO NO2 VCDs undergo strong seasonal and spatial variability. Large
NO2 VCDs are located in eastern China due to significant anthropogenic
emissions. On an annual basis, NO2 VCDs vary from
15–25 × 1015 cm-2 over the NCP to below
1015 cm-2 over much of western China. Over the polluted regions,
NO2 VCDs reach maxima in winter and minima in summer, due mostly to the
seasonal variation in NOx lifetime. Over cleaner regions (e.g., much of
western China), NO2 VCDs peak in summer with minima in winter, because
strong natural emissions in summer overcompensate for the short NO2
lifetime. The maximum to minimum ratio in monthly mean NO2 VCDs is about
3.6 for regions east of 101.25∘ E as a whole and about 1.4 over the
west. A POMINO-based NOx emission constraint leads to a posteriori
Chinese anthropogenic emissions at 9.05 TgN yr-1, an increase from
2006 (Lin, 2012) by about 19 %.
In one sensitivity test, we re-retrieved clouds and NO2 VCDs, by
adopting the monthly climatological OMI albedo data (OMLER v1) in place of
the MODIS BRDF data. Surface reflectance is greatly enhanced over most of
eastern China and reduced over the west. Changes in surface reflectance
result in changes of opposite sign in CF with more complex effects on CP. As
a consequence, annual mean NO2 VCDs are decreased by 5–15 % over
central and southern China, with enhancements over many other regions. The
reductions in southern China are largest (15–40 %) in fall, but are
replaced by slight enhancements in summer. On an annual basis, changes in
surface reflectance have a small effect (within 5 %) on constrained
NOx emissions over most Chinese locations with anthropogenic emissions
greater than 0.5 × 1015 cm-2 h-1.
In another sensitivity test, an implicit treatment of aerosols mimics the
traditional algorithms by excluding aerosol information in retrieving clouds
and NO2 VCDs. The implicit treatment greatly enhances effective CF (the
spatial correlation is 0.75 between neglected annual mean AOD values and CF
enhancements), CRF (correlation 0.82), and CP (correlation 0.19). The low
correlation for CP highlights the complexity of aerosol effects. Changes in
NO2 VCDs from an explicit to an implicit treatment of aerosols depend in
a complex manner on AOD, SSA, heights of aerosols relative to NO2, and
other factors; these factors affect CRF, VCDcl and VCDcr that determine
NO2 VCDs. The annual mean NO2 VCDs are enhanced by 15–40 %
over much of eastern China. The seasonal and spatial variability of NO2
changes is apparent. NO2 VCDs are reduced by 10–20 % over parts of
the NCP in spring and over northern China in winter, whereas these reductions
are replaced by general enhancements in summer and fall. The effect on
subsequently constrained annual NOx emissions varies between -5 and
30 % over eastern China with apparent seasonal and regional dependence.
For the usual criterion of CRF < 50 % to select “valid” pixels,
the large enhancements in CRF due to an implicit aerosol treatment result in
significant reductions in the number of days with “valid” pixels over
polluted regions (by 2–10 days per month, or 15–60 %, on average). This
also leads to a likely sampling bias due to the exclusion of high-aerosol
days that often experience high NO2 pollution, consistent with the
findings of Lin et al. (2014b).
The effect of an implicit aerosol treatment on retrieved NO2 VCDs also
interacts with the effect of changes in surface reflectance. NO2 changes
obtained by simultaneously excluding aerosol information and adopting the OMI
albedo are smaller than the overall NO2 changes obtained by making these
two adjustments individually and summing the results. Although these
adjustments (no aerosols + OMI albedo) are present in the DOMINO v2
approach, the resulting NO2 VCDs still differ from those of DOMINO v2,
indicating the importance of differences between other factors assumed in the
retrieval processes (e.g., NO2 profile shape, atmospheric profile in
O2–O2 retrieval, pixel-by-pixel radiative transfer calculation vs.
look-up table, and air pressure).
There are no sufficiently comprehensive independent measurements available
to systematically evaluate the various NO2 retrieval approaches.
Current MAX-DOAS measurements are very limited over China, with few sites
and short operation periods (Irie et al., 2012; Ma et al., 2013; Hendrick et
al., 2014; Kanaya et al., 2014). In situ measurements are rare for vertical
profiles of aerosols and NO2. Our results show that the effects of
aerosols and surface reflectance are highly season- and location-dependent.
This clearly indicates the need for a comprehensive measurement network to
validate satellite data. Nonetheless, our present study and that of Lin et al. (2014b) point the way forward for a physically more realistic
NO2
retrieval by explicit inclusion of aerosol effects.
Currently, our POMINO NO2 data are available for 2004–2013
(http://www.atmos.pku.edu.cn/acm/acmModel.html#POMINO). Daily level-3
data on a 0.25∘ long. × 0.25∘ lat. grid are provided
on the webpage for general users, and level-2 data (including averaging
kernels) can be provided for advanced users upon request. We simultaneously
provide daily AOD, SSA and surface reflectance (MODIS BRF over land and turbid coastal ocean and OMLER
v3 albedo over the open ocean) data for users. We elect to include daily NO2
data with high aerosol loadings to reduce the potential sampling bias, in
line with our choice of an explicit aerosol treatment and as supported by Lin
et al. (2014b). We note that these NO2 data with high aerosol pollution
may be subject to larger uncertainties, since aerosols are not fully
accurately constrained by observations. Users can make their own judgment on
whether to include the NO2 data with high aerosol pollution. With more
aerosol data from satellite, ground and in situ measurements in the future,
aerosol optical effects can be better constrained to reduce the associated
uncertainties in NO2 retrievals.
Our parallelized LIDORT-driven AMFv6 code is also available for public use.
With 16 computational cores (Intel (R) Xeon (R) CPU X7550 at 2.00 GH)
running in parallel, it takes about 3 h of wall-clock time to retrieve both
cloud parameters and NO2 VCDs over China for a month; the retrieval
process includes pixel-specific radiative transfer calculations and explicit
treatments of aerosols and surface reflectance anisotropy. The excellent
scalability of our parallelized code means that additional computational
cores can be employed to further speed up the retrieval process. Such
retrieval efficiency enables a fast global retrieval that will be
particularly important for future fine-resolution satellite instruments such
as TropOMI (which is expected to have a data rate ∼ 8 times that of
OMI) and GEMS (which will be onboard a geostationary satellite with hourly
measurements at a horizontal resolution of 5 × 15 km2).