Introduction
Greenhouse gas (GHG) observations from space provide unprecedented global
measurements of column GHG concentration, facilitating inference of carbon
fluxes on regional scales (Yoshida et al., 2011; Crisp et al., 2012).
However, atmospheric aerosol scattering has been shown to have a
considerable impact on the retrieval of GHGs from space-based observations
in the near infrared (Aben et al., 2007; Yoshida et al., 2011; O'Dell et
al., 2012). A study by O'Dell et al. (2012) showed that the error budget in
the satellite retrievals of column-averaged dry-air mole fraction of
CO2 (XCO2) is dominated by systematic errors due to imperfect
characterization of cloud and aerosol properties. Previous studies also
showed that the bias can be greatly mitigated by incorporating simple
aerosol properties into the retrieval state variables (Butz et al., 2009;
Guerlet et al., 2013). It is therefore crucial to characterize the aerosol
properties for the GHG retrieval algorithm. However, aerosols have complex
types and size distributions and are highly variable in number density.
Their optical properties are very difficult to measure directly (Seinfeld
and Pandis, 2006). The global ground-based aerosol monitoring
network, AERONET (Holben et al., 1998), has been providing
high-accuracy measurements of total aerosol optical depth (AOD) from the
ultraviolet to the near infrared, but is sparsely distributed, suggesting
that it would be useful to have further means of constraining aerosol
optical properties.
Water vapor (H2O) has absorption features across the electromagnetic
spectrum, from the ultraviolet to the infrared. In a non-scattering
atmosphere, H2O retrievals from different absorption bands would give
exactly the same value. However, due to wavelength-dependent aerosol
scattering (Eck et al., 1999; Zhang et al., 2015), the different light path
changes in different absorption bands result in discrepancies in H2O
retrievals. This variation in H2O retrievals from different bands can
therefore provide information on aerosol properties. Based on this
principle, we propose a novel approach to describe aerosol scattering
effects using the variation in H2O retrievals from multiple spectral
bands. This approach is illustrated using data from the California
Laboratory for Atmospheric Remote Sensing (CLARS), which continuously
collects high-resolution spectra in the near infrared. A two-stream-exact
single scattering (2S-ESS) radiative transfer (RT) model is used to simulate
the observations and explain the physical mechanism behind the proposed
approach. The additional information on aerosol optical properties gained by
applying this approach could be used to improve retrievals of GHGs from
space in the presence of aerosols.
In Sect. 2, a detailed description of the CLARS Fourier Transform
Spectrometer (CLARS-FTS) is presented. Section 3
introduces the 15 different bands chosen for retrieving H2O from
CLARS measurements. In Sect. 4, the correlation between variations in
H2O retrievals from different bands and corresponding AOD data from
AERONET-Caltech is discussed. In Sect. 5, the role of aerosol scattering
on the variations in H2O retrievals is illustrated using the 2S-ESS RT
model. Discussion and conclusions are presented in Sects. 6 and 7, respectively.
Schematic diagram of CLARS-FTS measurement geometries for western Pasadena
and the AERONET site at Caltech. CLARS-FTS has two modes of operation, including
Los Angeles Basin Survey mode (LABS; in solid red) and the Spectralon Viewing
Observation mode (SVO; in blue). An example of light path change due to aerosol
scattering along the path from the basin to the mountain top is illustrated (in
dotted red). Also shown is the light path of AERONET-Caltech (in green).
CLARS
The CLARS-FTS is located near the top of Mt. Wilson at an altitude of
1670 m a.s.l. overlooking 28 land target sites in Los Angeles
(Table S1 and Fig. S1 in the Supplement of Fu et al., 2014; Wong et al.,
2015, 2016). It offers continuous high-resolution spectral measurements from
4000 to 8000 cm-1. As shown in Fig. 1, CLARS-FTS has two operating
modes: the “Spectralon Viewing Observation” (SVO) mode and the “Los Angeles
Basin Surveys” (LABS) mode. In SVO mode, the FTS looks at a
Spectralon® plate adjacent to the FTS to observe
the reflected solar spectrum from the free troposphere above Mt. Wilson,
which is usually above the planetary boundary layer (PBL) and relatively
immune to aerosol scattering. In LABS, the FTS looks down at target sites in
the LA Basin to observe the reflected sunlight, which travels through a long
light path within the urban PBL (Fig. 1) and undergoes absorption and
scattering by gas molecules and aerosols. This observation geometry by LABS
mode makes CLARS-FTS measurements not only highly sensitive to the
atmospheric composition in the PBL but also very susceptible to influence by
aerosol scattering and absorption. The CLARS measurement technique from Mt.
Wilson mimics geostationary satellite observations of reflected sunlight,
which are governed by a sun-to-surface and surface-to-instrument optical
path geometry. The geostationary observations can be used to retrieve GHG
mixing ratios (e.g., Xi et al., 2015).
Slant column density (SCD), the total number of absorbing molecules per unit
area along the optical path, is retrieved for H2O using the version 1.0
operational retrieval algorithm of CLARS-FTS (Fu et al., 2014), developed
based on the gas fitting tool (GFIT) algorithm that has been widely used for
the Total Carbon Column Observing Network (TCCON) network (Toon et al.,
1992; Wunch et al., 2011). Surface reflection is included in the latest
version of GFIT but aerosol scattering is not taken into account. Thus,
errors in the SCD retrieval are largely due to light path changes caused by
aerosol scattering, and can be used as a proxy for aerosol loading. A
detailed description of CLARS-FTS, including the observation system, the
measurement sequence, the operational retrieval algorithms, and the
operational data products, can be found in Fu et al. (2014). The 15 different
bands used in this study for retrieving H2O SCDs from CLARS-FTS are
introduced in Sect. 3.
The normalized radiances, obtained by dividing the spectra by the
maximum radiance, of the 15 H2O absorption bands selected for retrieving
H2O SCDs from CLARS-FTS measurements. These radiances are spectral fits
using the CLARS-FTS measurements in western Pasadena on 1 March 2013 with a solar
zenith angle (SZA) of 41.45∘. Solid black curves are fits to the spectra,
including contributions of all trace gases and solar lines, from spectral
measurements by the FTS, and dashed blue curves are the estimated contribution
of H2O absorption to the spectral fits. Contributions of other species in
these spectral regions are not shown. Central wavelength and information
content (IC) value of each band used for retrieving H2O content are also indicated.
Spectral bands for H2O retrievals
H2O has absorption features across the electromagnetic spectrum
including many absorption lines in the near infrared as well as significant
continuum absorption. Within the spectral range of CLARS-FTS observations,
we carefully selected 15 spectral intervals containing dominant H2O
absorption lines, as shown in Fig. 2 and Table 1, across the spectral
range from 4000 to 8000 cm-1 for retrieving H2O SCD. Each interval
is 5 to 14 cm-1 wide and contains two to nine moderately strong
absorption lines. Rodgers (2000) shows that information analysis is a
powerful tool which can be used in the channel selection process. Therefore,
we conduct information analysis by calculating the information content (IC)
for each band selected for retrieving H2O SCD. IC is calculated as
-ln|I - A|/2, where I is
the identity matrix and A is the averaging kernel, a measure of the
sensitivity of the retrieval state to true state variables. A detailed
theoretical description of the information analysis can be found in Su et
al. (2016). The RT model used here is the numerically efficient 2S-ESS RT
model (Spurr and Natraj, 2011). The detailed settings for the 2S-ESS model
are described in Sect. 5. In the RT model, absorptions by the dominant gas
molecules in the atmosphere, including H2O, CO2, N2O, CO,
CH4, O2, N2, and HDO, are considered by using an a
priori atmospheric profile obtained from the National Center for
Environmental Prediction (NCEP)–National Center for Atmospheric Research (NCAR)
reanalysis data (Kalnay et al., 1996). AOD observations from AERONET,
and SSA and g from MERRA-GOCART are included in the RT model (details are
provided in Sect. 5). Here, we assume that only one state variable, i.e.,
H2O SCD, is retrieved. Therefore, the IC values shown in Fig. 2 are
for H2O retrieval only; they are indicators of the precision of
H2O retrieval in the selected bands. For any band with IC value equal
to x, as many as ex different atmospheric H2O states can
be distinguished (Rodgers, 2000). The purpose of the IC calculation is to
show the sensitivity of each absorption band to the variation of H2O
SCD. While fitting the CLARS-FTS measured spectra, we retrieve more state
variables including other trace gas abundances (CO2, CH4, CO, and
N2O) and continuum shape parameters (Fu et al., 2014). From the IC
results shown in Fig. 2, we can see that the average IC is high (about 6
on average) and similar among all bands, which indicates very high retrieval
precision compared with the a priori information (Rogers, 2000). H2O SCDs
retrieved from these 15 bands will be shown in the following section.
Central wavelengths and band widths of the 15 H2O absorption bands
and the corresponding information contents for retrievals of only H2O
SCD (IC_a) and for simultaneous retrievals of H2O SCD and AOD (IC_b).
Central
Band
IC_a
IC_b
wavelength
width
(cm-1)
(cm-1)
4554.0
8.0
6.34
8.61
4568.0
12.0
6.44
8.47
4596.5
9.0
6.36
8.27
4632.0
8.0
5.89
6.72
4645.0
10.0
6.41
8.04
4703.0
14.0
6.44
9.55
5859.0
8.0
6.13
8.11
5910.0
6.0
5.60
6.29
6505.5
5.0
5.20
5.36
6534.0
12.0
6.15
7.58
6550.0
10.0
6.40
8.73
6618.0
10.0
6.52
9.40
7712.5
7.0
5.97
7.61
7738.0
12.0
6.20
8.23
7760.0
10.0
6.13
7.72
Daily Variations of CLARS H2O SCD retrievals for the western Pasadena
target and AOD measurements from AERONET-Caltech station on 1 March 2013 (left
column panels) and 28 September 2013 (right column panels). H2O SCD
retrievals from SVO mode are shown in (a), and from LABS mode for western Pasadena are shown in (b). The corresponding standard deviations of
H2O SCD retrievals, a measure of the degree of variation in the retrievals,
are shown in (c) and the AOD measurements from AERONET-Caltech are
shown in (d).
H2O SCD retrievals from CLARS and its correlation with AOD
Daily variation
Figure 3 shows examples of daily CLARS H2O SCD retrievals on 1 March 2013
and 28 September 2013. On these 2 days, there were dense
observations for the western Pasadena target retrieved using the 15 bands shown
in Fig. 2. Observations between 09:00 and 15:00 LT (local time) are shown, when
the solar zenith angles (SZAs) are less than 60∘, similar to the
majority of satellite observations. This time period also includes the local
overpass times at around 10:00 LT for SCIAMACHY (Bovensmann et al., 1999) and
around 13:30 LT for GOSAT, OCO-2, and TanSat (Liu et al., 2013). In particular,
we compare retrievals from SVO (Fig. 3a), which is above the PBL (Newman
et al., 2013) and therefore relatively unaffected by aerosol scattering,
with those from the western Pasadena target (Fig. 3b), a location in the
Los Angeles Basin that is influenced by aerosol scattering. The H2O SCD
retrievals from SVO are nearly identical across different wavelengths, and
the small differences (about 14.2 and 8.2 % of those from western Pasadena for 1 March and 28 September, respectively, in terms of standard
deviation of H2O SCD retrievals) may be attributed to the band-to-band
inconsistency of line parameters. However, H2O SCD retrievals for western Pasadena show significantly larger variation (about 7 times and 12 times
those from SVO for 1 March and 28 September, respectively, in terms of
standard deviation of H2O SCD retrievals) across different wavelengths.
These retrieval differences are much larger than can be attributed to
spectroscopic uncertainties alone, and reflect the wavelength dependence of
aerosol scattering in the boundary layer. To quantify the variation in the
H2O SCD retrievals from the 15 bands, the standard deviation (σ)
of the retrievals is calculated by the following:
σ=∑i=1nsi-s‾2/(n-1),
where n = 15 is the number of bands, si is the SCD retrieval for band i, and
s‾ = (∑i=1nsi)/n is the mean. The standard deviations
in Fig. 3c show that the variations in H2O SCD retrievals
monotonically increase throughout the day. As shown in Fig. 3d, the AOD
data on these 2 days from AERONET-Caltech show a typical pattern for the
LA Basin. AOD increases from the morning to the afternoon. Note here that
the AERONET station measures mainly in the visible and near-infrared
wavelengths from 340 to 1020 nm. However, we assume that these measurements
are also good proxies for AOD at longer wavelengths in the near infrared.
The increasing trend of AOD during the course of the day corresponds well to
that of the standard deviations of H2O SCD retrievals. This correlation
from daily measurements shows the potential of constraining the AOD using
the standard deviation of H2O SCD retrievals. However, apart from
aerosol scattering, the daily variation of standard deviations of H2O
SCD retrievals can also be influenced by differences in observation
geometry, such as SZA and relative azimuth angle (AZA), and variations in
planetary boundary layer height (PBLH). All of these parameters are changing
during the day. These effects are included in the discussion below.
Correlation between daily averaged standard deviation of H2O SCDs,
a measure of retrieval differences, from 12:00 to 14:00 LT, and the corresponding
averaged AOP, calculated by scaling AOD data (1020 nm) from AERONET based on
CLARS geometry, for two time periods in 2013. (a) Winter and spring,
including January to May, in which the coefficient of determination (R2)
is 0.53 and [slope, intercept] = [0.08 ± 0.03, -0.09 ± 0.21]
with 95 % confidence bounds from linear regression. No December data from
AERONET are available in 2013. (b) Summer and autumn from June to
November, in which R2 is 0.49 and [slope, intercept] = [0.04 ± 0.01,
-0.03 ± 0.16]. In total, there are 68 days of daily mean data available
in 2013, out of which 27 days are for winter–spring and 40 days are for summer–autumn.
Seasonal variation
To quantitatively compare the variations in H2O SCD retrievals and AOD,
we choose the daily mean of the data between 12:00 and 14:00 LT when
there is generally less haze or fog. This time is also coincident with the
local crossing time of the two currently operating GHG observation
satellites, GOSAT (Yoshida et al., 2011) and OCO-2 (Crisp et al., 2012). We
focus on a 2 h time period to limit the effect of other parameters on
the H2O retrieval standard deviation, such as solar geometry and PBLH.
The closest (temporal and spectral) AERONET AOD data are used as the
concurrent AOD data for the CLARS retrievals. If the measurement time
difference is more than 30 min, then the CLARS data are not used.
AERONET-Caltech has AOD measurements in seven wavelengths from 340 to 1020 nm,
and we choose the data at 1020 nm, which is the closest to the
wavelengths used for retrieving H2O in this study (from 1289 to
2196 nm). Furthermore, aerosol optical path (AOP) values are calculated by
multiplying the vertical path AOD data by air mass factors, which are
derived from the SZA and viewing zenith angle of CLARS at western Pasadena
(83.1∘). Following this procedure, 68 daily mean data pairs are
available in 2013 after excluding days (1) that are cloudy according to the
images from the visible camera looking at the target and (2) in which there
are fewer than three valid observations. We separate the data into two
different time periods, the winter–spring season (December to May) and the
summer–autumn season (June to November) by considering the different
dominant wind directions between winter and summer in Pasadena (Conil and
Hall, 2006; Newman et al., 2016). In the summer, the prevailing winds come
from the southwest across the basin; in the winter, the winds come from the
northeast across the mountains and deserts. Different wind patterns (Newman
et al., 2016) suggest that the dominant aerosol types during these two time
periods may be different. For each of these two time periods, we expect the
CLARS observation geometry and PBLH at noon time to be similar across the
days, and therefore assume their effects on H2O retrievals to be
similar. In fact, it will be clear from the RT model simulation in Sect. 5
that the contributions to the variation in H2O retrievals from other
factors, such as solar geometry and PBLH, are actually much smaller than
those from aerosol scattering, even though the geometry and PBLH change a
lot during a day.
Figure 4 shows significant linear correlations between the AOP value and the
standard deviation of H2O SCDs for the two time periods, both with
R2 around 0.5. However, the slopes from linear regression between the
AOP and the standard deviation of H2O SCD retrievals are different. In
summer–autumn, the regression slope, an indicator of the degree of light
path change due to aerosol scattering relative to the change in aerosol
loading, is about one half of that in winter–spring, indicating that for the
same AOD, the variation in H2O SCD retrievals in summer–autumn is about
twice that in winter–spring. H2O abundance shows significant seasonal
variation in the LA Basin, and previous studies have shown, for example,
that aerosol optical properties change dramatically with relative humidity
(e.g. Thompson et al., 2012). In fact, results from control experiments
using the RT model, shown in Sect. 5, show that the difference in H2O
abundance and aerosol phase function, indicated by the asymmetry parameter,
are responsible for the difference in slope between the winter–spring and
summer–autumn periods.
Simulations using 2S-ESS RT model
In this section, we use the 2S-ESS RT model to simulate the measurements and
quantify the role of aerosol scattering in the variation in H2O SCDs
retrieved from CLARS. The 2S-ESS model performs an exact computation of the
single scattering using all moments of the phase function, while the
multi-scattered radiation is calculated using the two-stream
approximation. This model has been used for GHG remote sensing in several
previous studies (Xi et al., 2015; Zhang et al., 2015, 2016).
Monthly-averaged climatological aerosol composition (as percentages
of total optical depth averaged over the 15 H2O absorption bands) for the
five composite MERRA aerosols (black carbon, organic carbon, sulfate, dust, and
sea salt) in the daytime (13:00 LT).
We simulate the spectral radiance observed by CLARS-FTS for the western Pasadena target. The settings for this model are largely the same as those
used by Zhang et al. (2015). Some essential settings and modifications in this RT model are
as follows:
the a priori atmospheric profile is obtained from NCEP–NCAR reanalysis data (Kalnay et al., 1996). The
profile has 70 layers from the surface up to 70 km;
absorption coefficients for all absorbing gas molecules are derived from the HITRAN
version 2008 database (Rothman et al., 2009);
the optical depth for each layer is calculated using the Reference Forward Model (Dudhia et al., 2002);
the surface reflection is assumed to be Lambertian with a surface albedo
of 0.23, as measured by Fu et al. (2014) for western Pasadena;
Rayleigh scattering by air molecules is considered;
the observation geometry, including the viewing zenith angle for the western
Pasadena target, the daily SZA, and AZA on 1 March 2013, are included in the model;
the aerosol scattering phase function in the model is assumed
to follow the Henyey–Greenstein type phase function (Henyey and Greenstein,
1941). Climatological aerosol compositions, as percentages of total optical
depth, for five types of aerosol (black carbon, organic carbon, sulfate,
dust, and sea salt) are obtained from the Modern Era Retrospective analysis
for Research and Applications (MERRA) aerosol reanalysis database (Rienecker
et al., 2011; Buchard et al., 2015). A more detailed description of the data
can be found in Connor et al. (2016). The aerosol single scattering
properties, including single scattering albedo (SSA) and asymmetry parameter (g),
were computed using the GOCART model (Colarco et al., 2010; Chin et
al., 2002). These properties were tabulated for black carbon, organic
carbon, sulfate, dust, and sea salt aerosol types, with hygroscopic effects
included where appropriate. The average values of the scattering parameters
are shown in Table 2. Using the compositions of the five composite MERRA
aerosols and their scattering properties, the effective SSA and g of aerosol
scattering are calculated, respectively, as the composition-weighted sum,
and then incorporated into the 2S-ESS RT model. Figure 5 shows the
monthly-averaged climatological aerosol compositions for the five composite
MERRA aerosols. We can see that, in general, sea salt dominates in the
summer–autumn period while dust dominates in the winter–spring period;
unlike Zhang et al. (2015), in this study, the average hourly PBLH data
measured over late spring in 2010 in LA (Newman et al., 2013) are used.
Unlike CO2, the H2O mixing ratio varies dramatically with altitude
in the atmospheric column (Seinfeld and Pandis, 2006). In the LA Basin, a large portion of H2O is concentrated within the PBL.
Therefore, the PBLH is an important parameter in modeling the effects of
scattering on the H2O retrieval. The model output radiance is convolved
using the CLARS-FTS instrument line shape with full width at half maximum
(FWHM) = 0.022 cm-1 (Fu et al., 2014). The spectral resolution is set
to be 0.06 cm-1, and the corresponding instrument maximum optical path
difference is 5.0 cm. The signal-to-noise ratio is assumed to be constant at 300.
We perturb the simulated spectra with Gaussian white
noise.
Aerosol single scattering properties, computed using the GOCART
model*, for the five types of aerosols (black carbon, organic carbon,
sulfate, dust, and sea salt) used in the study. Single Scattering Albedo (SSA)
and asymmetry parameter (g) are averaged values over the 15 H2O absorption bands.
Black
Organic
Sulfate
Dust
Sea
carbon
carbon
salt
SSA
0.03
0.77
0.97
0.94
0.99
g
0.10
0.25
0.35
0.71
0.80
* Colarco et al. (2010) and Chin et al. (2002).
(a) Case I: scaling factors for H2O SCDs retrieved from
the simulated synthetic spectral radiance of the 15 chosen bands using the
2S-ESS RT model with AOD data from AERONET-Caltech on 1 March 2013. The scaling
factors are mean-centered by subtracting the mean to clearly show the variations
in the retrievals; (b) Case II: same as (a) except that the
AOD is fixed at the clear-day level, in which the lowest AOD in 2013 is used for
all hours across the day; (c) Case III: same as (a) except that
the AOD is set to be zero for all hours across the day; (d) Comparison
between CLARS measurements and results from the three RT model experiments
in (a)–(c) in terms of standard deviations, a measure of
variations in H2O SCDs retrieved from the 15 chosen bands. The standard
deviations are normalized to be between 0 and 1 for both measurements and
simulations. The half-hourly mean of the CLARS data is calculated to obtain the
maximum and minimum for the normalization.
The wavelength range covered by AERONET-Caltech measurements is from 340 to
1020 nm; however, the wavelengths of the 15 H2O absorption bands used
in this study, ranging from about 1280 to 2200 nm, are outside the AERONET
wavelength range. For the sake of calculations in the 2S-ESS RT model, the
AOD data in these 15 bands were extrapolated using the Ångström
exponent law (Seinfeld and Pandis, 2006; Zhang et al., 2015). This law is given by the following:
ττ0=λλ0-k,
where λ and τ are the wavelength and the corresponding AOD to
be interpolated, respectively; λ0 and τ0 are the
reference wavelength and the corresponding AOD from AERONET, respectively;
and k is the Ångström exponent. The k value is obtained by applying linear
regression using the logarithmic form of Eq. (2), on the AERONET AOD
measurements in the six different bands (340, 380, 440, 500, 870, and 1020 nm).
Examples of applying this law to the AERONET AOD
measurements are shown in Fig. S1, from which we can
see that the wavelength dependence of the total AOD, a combination of
different types of aerosols, generally follows the above exponent law. The
extrapolated AOD data on 1 March 2013 for the 15 bands are included in the
RT model, assuming non-zero AOD is evenly distributed vertically and
horizontally in the PBL.
Simulations of daily variation
To quantify the influence of aerosol scattering on the H2O SCD
retrievals, we simulate the bias observed by CLARS-FTS by (1) using the
2S-ESS RT model to generate synthetic spectral radiance data for the
15 chosen bands, and (2) fitting the synthetic spectral data and retrieving
H2O SCD based on Bayesian inversion theory (Rodgers, 2000) using the
forward 2S-ESS RT model with the same configuration, but with AOD set to
zero and held constant, as in Zhang et al. (2015). This approach
approximately simulates the influence of neglecting aerosol scattering on
the retrieved H2O SCDs by CLARS. The fitting process employs the
Levenberg–Marquardt algorithm (Rodgers, 2000). The state vector element to
be retrieved from the inversion approach is the scaling factor, which is the
ratio of retrieved H2O SCD to the assumed “truth” data obtained from
NCEP reanalysis. Figure S2 is a schematic diagram of the retrieval algorithm
which is based on the 2S-ESS RT model and Bayesian inversion theory.
We perform three experiments to demonstrate the effect of aerosol scattering
on the variations in H2O SCD retrievals (Fig. 6). In the first
experiment, denoted as Case I (Fig. 6a), the AOD data vary during the
day in the same way as the AERONET-Caltech measurements, while in the second
(control) experiment, denoted as Case II (Fig. 6b), the AOD data are
fixed at the clear-day level, in which the lowest AOD across the year is
used for all hours across the day. In the third (also control) experiment,
denoted as Case III (Fig. 6c), the AOD is set to be zero for all hours
across the day. In these three cases, the changes of AOD in the forward
modeling are made only when we generate the synthetic spectral radiance (see
also Fig. S2). For inverse modeling (when we use
the RT model to produce the Jacobian matrix), we keep the AOD constant at
zero to approximately simulate the influence of neglecting aerosol
scattering on the retrieved H2O SCDs by CLARS.
The results for simulated H2O SCD retrieval scaling factors are shown
in Fig. 6a–c, for these three experiments, respectively. The
scaling factors (f) are mean-centered by subtracting the mean to
clearly show the variations in the data (scaling factors before
mean-centering are shown in Fig. S3). The
mean-centered scaling factor is calculated as follows:
f^i = fi - f‾ for
i = 1 to 15, where fi is the scaling factor for
band i, f^i is its mean-centered scaling factor, and
f‾ is the mean of the scaling factors. From Fig. 6a, we can see
that the variation in the simulated H2O SCD retrievals increases with
increasing AOD from the morning to the afternoon, similar to what we see
from the CLARS observations in Fig. 3b and c. The scattering effect
is stronger in the afternoon than in the morning, as shown in Fig. 6b,
even though the AOD is the same for all hours; this effect is probably due
to the changes in SZA and AZA from the morning to the afternoon. However,
since the results from the control experiments show much smaller diurnal
changes, seen in Fig. 6b and c, than the first experiment, aerosol
scattering must be the dominant cause of the variations in H2O SCD retrievals.
Further confirmation of this is provided by Fig. 6d, which shows the
comparison between CLARS measurements and the three RT model experiments in
terms of normalized standard deviations of H2O SCDs (data before
normalization are shown in Fig. S4). As with the
CLARS measurements, the standard deviations of the scaling factors are
calculated using Eq. (1). To emphasize the relative change in
variations of H2O SCDs between measurements and simulations, their
standard deviations (σ) are normalized to lie between 0 and 1.
The normalized standard deviations are calculated as σ̃t = (σt - σmin)/(σmax - σmin), where
σt is the standard deviation, σ̃t is the normalized standard deviation at time t, and
σmax and σmin are
the maximum and minimum standard deviations, respectively, throughout the
day. This normalization is independently implemented for measurements and
simulations. When normalizing the CLARS measurements, the half-hourly means
of the data are calculated to obtain the maximum and minimum. From Fig. 6d,
we can see that the increasing trend from measurements is very similar
to that from the simulations from the first experiment, while the trend is
much smaller for the control experiment. The major difference between the
results from the different cases can be attributed to the effect of aerosol
scattering in the PBL. Hence, we conclude that aerosol scattering is the
dominant factor contributing to the variations in H2O SCD retrievals,
and the correlation between AOD and standard deviations of H2O SCD
retrievals from measurements is robust. Therefore, the H2O SCD
retrievals from CLARS can potentially be used for constraining the aerosol
properties in the LA basin.
Simulations of seasonal variation
To investigate the sensitivity of the correlation between variations in
H2O SCDs and the corresponding AOP (as shown in Fig. 4) to aerosol
scattering properties, we reproduce the correlation using the 2S-ESS RT
model with input AERONET AOD data, daily aerosol compositions (at 13:00 LT;
monthly means shown in Fig. 5) derived from MERRA aerosol
reanalysis data, and aerosol scattering properties (shown in Table 2),
including SSA and g, computed by the GOCART model for the 68 days (discussed
in Sect. 4.2) at 13:00 LT. The retrieval algorithm is the same as
that described in Sect. 5.1. After obtaining the retrieved scaling factor
for H2O SCD for each of the 68 days, the simulated H2O SCD
retrieval is calculated as the product of the scaling factor and the
H2O abundance truth, which is set to be the mean of the H2O SCD
retrievals from the 15 bands used by the CLARS-FTS operational retrieval algorithm.
The result is shown in Fig. 7a. We can see that the strong linear
correlations between the standard deviations of H2O SCDs and the
corresponding AOP are well reproduced with higher R2 for both
winter–spring and summer–autumn periods. However, the slopes from the linear
regression in the simulations are slightly overestimated (by about 20 %)
relative to those for the CLARS data (as shown in Fig. 4), probably due to
the uncertainties of the input climatology aerosol compositions or single
scattering properties. Interestingly, the ratio of the slope from the
winter–spring to that from the summer–autumn periods from the 2S-ESS model,
which is about 2.0, agrees well with that from CLARS observations. On the
other hand, the difference between the two slopes becomes very small in
Fig. 7b. This suggests that the difference in H2O abundance between
the winter–spring and summer–autumn periods is responsible for the slope difference.
Similar control experiments implemented to examine the impact of variations
of SSA and g are shown in Fig. 7c and d, respectively. We can see
that the results are similar to those in Fig. 7a, suggesting that the
variations of SSA and g are not the key contributors to the variations of
the standard deviations of H2O SCDs. However, the difference between slopes
in Fig. 7d are slightly smaller compared to those in Fig. 7a,
indicating that g might contribute to the small difference which still
exists in Fig. 7b even when the AOD does not vary. To further
investigate the impact of g on the slope, Fig. 7e shows the result from
the 2S-ESS model with the same settings as those in Fig. 7a but with g
reduced by 50 %; we can see that the slopes for the two time periods
become smaller (by about 13 %), indicating that the scattering effects
become stronger when g is reduced. This is because, when g is smaller, the
aerosol scattering phase function becomes less forward peaked and more light
is scattered at scattering angles around 40∘, which is the scattering
angle for CLARS at noon time. Small aerosols, such as organic carbon, black
carbon, and sulfate from urban pollution and biomass burning, have much
smaller SSA and g, as shown in Table 2, suggesting that the slope will also
be small during events that are dominated by these small aerosols. From
these results, we can conclude that the strong linear correlation between
AOD and variations in H2O SCDs is robust, and that g is the main
contributor to the ratio of the changes (the slope in linear regression) between them.
Reproduction of the correlation, as shown in Fig. 4, between daily
averaged standard deviation of H2O SCDs at noon and the corresponding
averaged AOP for two time periods in 2013, using the 2S-ESS RT model. The input
AOD data are obtained from AERONET, aerosol compositions are derived from MERRA
aerosol reanalysis data, and the related scattering properties, including single
scattering albedo (SSA) and asymmetry parameter (g), are computed using the
GOCART model. The slopes and R2 from the linear regression for the winter–spring
period (solid line) and summer–autumn period (dash line) are shown on the
upper-left and bottom-right corners, respectively. Results from five experiments
are presented, including (a) the linear correlation reproduced by the
2S-ESS model with averaged AOD data between 12:00 and 14:00 LT and compositions
for the five composite aerosols (black carbon, organic carbon, sulfate, dust, and sea
salt) at 13:00 LT, for the 68 days;
(b) same as (a) except that the H2O SCD truth is fixed
for all days at the mean value for the 68 days in 2013, in order to examine the
impact from variability in H2O abundance; (c) same as (a)
except that the SSA is fixed for all days at the mean value for the 68 days
in 2013, in order to examine the impact from SSA variations; (d) same
as (a) except that g is fixed for all days at the mean value for the
68 days in 2013, in order to examine the impact from g variations;
(e) same as (a) except that g for each of the 68 days is
set at half of the original value, in order to examine its contribution to the
linear correlation.
Retrieval of AOD using the H2O absorption bands
The purposes of this section are to (1) demonstrate that accurate AODs can
be retrieved from the spectra data using the 15 H2O absorption bands,
and (2) investigate the impact of AOD retrieval on the process of retrieving
H2O SCDs in the H2O absorption bands. Given the fact that the
operational retrieval algorithm for CLARS-FTS, developed based on the GFIT
algorithm, does not take aerosol scattering into account, we make the above
calculations by conducting a realistic numerical simulation study using the
2S-ESS RT model. The advantages of this numerical simulation study are that
(1) the truth state vector is known and we can directly assess the accuracy
of the retrievals, and (2) control experiments can be conducted by
minimizing the influences from other factors except for the one that is
investigated. The calculations are implemented using noon data for the 68 days
shown in Fig. 4.
Retrieval of AOD and H2O SCD simultaneously based on a realistic
numerical simulation study using the 2S-ESS RT model. (a) Scatter plots
between true and retrieved AOD from the simultaneous retrieval experiment. The
mean and standard deviation of the difference between them are -0.0018 and 0.0051,
respectively. The black dotted line is the one-to-one line; (b) retrieval
of AOD and H2O SCD scale factors averaged over the 15 H2O absorption bands
from the simultaneous retrieval experiment; the one sigma error bar is also
shown; (c) H2O SCD scale factor retrievals from three different
cases; in Case A, aerosol scattering is not considered; in Case B, H2O
and AOD are simultaneously retrieved; in Case C, AOD is perfectly known.
Simultaneous retrieval of H2O SCDs and AODs
In the retrieval algorithm illustrated in Fig. S2,
we first produce synthetic spectra using the 2S-ESS RT model with input AOD
from AERONET and SSA and g from MERRA-GOCART. We then set H2O SCD scale
factor and AOD as the state variables, and simultaneously retrieve them
using the 15 H2O absorption bands. The ICs for the 15 H2O
absorption bands are shown in Table 1. The retrieval results are shown in
Fig. 8a and b. From the comparison between retrieved and true AODs
(Fig. 8a), we see that the retrieved AOD agrees well with the truth in
the H2O absorption bands for all months (R2 = 0.93;
RMSE = 0.0051). From the simultaneously retrieved H2O SCD scale factor
and AOD averaged over the 15 bands (Fig. 8b), we see that their
retrieval errors show a similar pattern; when the H2O SCD scale factor
diverges from unity, the difference between the retrieved and true AOD also
increases. When H2O SCD becomes smaller, the observed radiance
increases since fewer photons are absorbed in the H2O bands. On the
other hand, when AOD becomes smaller, the observed radiance decreases, since
the aerosol scattering effect becomes weaker and fewer photons will be
scattered to the observer in the H2O bands. Therefore, we can see that
the AOD and H2O SCD retrievals show the same pattern, whereby smaller
AOD retrievals coincide with smaller H2O SCD retrievals, so that their
effects on the observed radiance largely cancel out.
Retrieval of H2O SCDs when AODs are perfectly known
We set the H2O SCD to be the only state variable and assume that the
AOD is perfectly known. Comparison of the retrieved H2O SCD scale
factors from three different cases are shown in Fig. 8c. In Case A,
aerosol scattering is not considered; in Case B, H2O and AOD are
simultaneously retrieved; in Case C, the AOD is perfectly known. We can see
that (1) H2O SCD can be accurately retrieved if we have perfect
knowledge of AOD, and (2) when we retrieve AOD and H2O SCD
simultaneously, the variations in retrieved H2O SCD scale factors are
largely reduced compared to the case when aerosol scattering is not considered.
Based on the above results from a realistic numerical simulation study, we
conclude that (1) accurate AODs and H2O SCDs can be simultaneously
retrieved from the spectral data in the 15 H2O absorption bands, and
(2) variations in retrieved H2O SCD scale factors are largely reduced
when we retrieve AOD simultaneously compared to that when aerosol scattering
is not considered.