Introduction
Stratospheric aerosols play an import role in several atmospheric processes, including radiative forcing and ozone
depletion. For decades, monitoring of stratospheric aerosols from satellite observations was largely the domain of
occultation instruments such as Stratospheric Aerosol and Gas Experiment (SAGE) II. However, since the 2000s aerosol extinction has been retrieved from limb scatter
instruments such as the Optical Spectrograph and InfraRed Imaging System (OSIRIS) and references
therein, the SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY
(SCIAMACHY) and references therein and the Ozone Mapping and
Profile Suite Limb Profiler (OMPS-LP) . While limb scatter provides greatly improved
global coverage over occultation satellites, it requires additional assumptions and computationally expensive forward
models to perform the inversions. Despite the difficulties, comparisons between limb scatter and occultation measurements
generally agree favourably with mean biases in the 10–15 % range during volcanically quiescent periods. While this is
the average case, biases at certain latitudes and altitudes can be considerably larger. Additionally, biases after 2005
have not been well characterized due to the lack of baseline occultation measurements with which to compare.
This paper investigates the cause of the biases between the OSIRIS and SCIAMACHY aerosol extinction retrievals using
comparisons with SAGE II and a series of simulation studies. The two limb-scattering instruments and the inversion
techniques are described in Sect. . Also introduced here is the new version 1.4 SCIAMACHY aerosol
extinction product used in this work. Initially, a triple comparison among OSIRIS, SCIAMACHY and SAGE II is performed in
Sect. . As there was very little volcanic influence on the stratospheric aerosol load during the
overlap period, this serves as a baseline for the agreement seen between the limb scatter and occultation aerosol records
during volcanically quiescent times and motivates the investigation of error sources. Section
discusses the magnitudes of the errors that are expected from the assumptions in the OSIRIS and SCIAMACHY retrievals and
radiative transfer models through a series of simulation studies. Section applies the IUP and USask
retrievals to both datasets to investigate differences due to the inversion techniques and radiance products. Lastly,
conclusions and recommendations are discussed in Sect.
The aerosol retrievals
Generally, aerosol extinction retrievals for OSIRIS, SCIAMACHY and OMPS-LP limb-scattering instruments proceed in
a similar fashion. First, radiance profiles at one or more wavelengths are used to construct a single measurement vector
as a function of altitude. As this provides only one piece of information at each altitude, aerosol extinction is
typically chosen as the retrieved quantity, although this is not the only possibility. However, extinction is the natural
quantity retrieved from occultation instruments and allows for continuation of this historical record. Ideally, the
measurement vector would be dependent only on the desired aerosol extinction parameter, but in practice it is also affected
by the surface albedo, atmospheric density and aerosol optical properties including particle size, shape and composition.
Typically, an effective Lambertian surface reflectivity is retrieved concurrently with the aerosol extinction, while the
atmospheric density and optical properties are assumed using external information. Although atmospheric density is
provided at high resolution by ECMWF (European Centre for Medium-Range Weather Forecasts) or MERRA,
data on aerosol optical properties are much sparser and a notable
limitation in the current retrievals.
Although particle size information has been retrieved from limb instruments in the past with OMPS-LP and OSIRIS
and more recently with SCIAMACHY , the standard operational products
remain as extinction-only for the OSIRIS and OMPS-LP aerosol products. These extinction products have been used in
numerous studies and continue to contribute to the stratospheric aerosol record ,
highlighting the importance of accurately characterizing not only precision but also biases in the current operational
retrievals.
OSIRIS v5.07
OSIRIS was launched in 2001 aboard the Odin spacecraft
. The spectrograph produces limb-scattered radiance profiles from 280 to 810 nm, with typical
sampling every 2 km, a vertical resolution of 1 km and an altitude range from 7 to 75 km. Odin is in a near-terminator orbit with an equatorial crossing time of approximately 06:00 on the descending node, providing limb
measurements with a limited range of viewing geometries. Typically, solar scattering angles vary between 60 and
120∘ with the largest values occurring in the tropics, and little correlation between the mean scattering angle
and latitude. The OSIRIS measurements have been used in the inversions of multiple species with products now spanning over
15 years . The inversions use the SASKTRAN radiative transfer model
and a multiplicative algebraic reconstruction technique (MART) to retrieve ozone, NO2 and aerosol extinction at
750 nm. This paper uses the OSIRIS v5.07 aerosol data product retrieved with the algorithm discussed in
, which simplifies to the Chahine inversion technique for the choice
of tangent altitude weighting factors in the aerosol-specific portion of the MART retrieval. This algorithm will be
referred to as the USask retrieval in this paper. For the radiative transfer modelling, a unimodal lognormal distribution
is assumed with median radius, rg of 80 nm and distribution width, σg, of 1.6 as defined in
Eq. (). This distribution is consistent with mid-latitude optical particle counter (OPC) measurements during
volcanically quiescent periods , although the variability in the OPC measurements is large, as
discussed in Sect. . Mie theory is used to calculate the aerosol scattering properties with
a refractive index from assuming a 75 % concentration of H2SO4 and 25 % H2O.
This produces a refractive index of 1.427 + i7.167 × 10-8 at 750 nm and 1.432 + i0.0 at
470 nm. The USask measurement vector is defined as
yjk′=lnI(λk,j)I(λref,j)-1N∑jref=mm+N-1lnI(λk,jref)I(λref,jref),
where the measurement vector, yjk′ at wavelength k and altitude j is the radiance, I, normalized by
a reference altitude, jref, and shorter wavelength, λref, that is generally less sensitive to
aerosols. To reduce noise at the reference altitude N measurements are used, beginning at tangent height
jref=m. To improve the convergence speed of the relaxation technique ,
a modelled measurement vector assuming a molecular atmosphere is also used as a normalization, yielding the measurement
vector
yjk=yjk′-yjkmol,
where ymol is computed using Eq. (), with the modelled radiances assuming an
aerosol-free atmosphere. As this acts as a constant offset, it does not affect the sensitivity of the measurement vector
to aerosols. However, in addition to improving convergence, this normalization also helps to identify the region of
interest for the aerosol retrieval; after normalization by the molecular signal, the dominant components remaining are
aerosol at lower altitudes and stray light at higher altitudes. The reference altitudes are chosen as the point, or
points, where the measurement vector is at a minimum within the measurement noise, i.e. where both the stray light and
aerosol signals are smallest. This produces a normalization that varies scan to scan, but typically produces reference
altitudes between 25 and 40 km with lower altitudes near the poles. For the USask retrieval, 750 nm is used as the long
wavelength, λk, and 470 nm is used as the reference, λref. Atmospheric data for pressure and
temperature are interpolated to the OSIRIS scan location from the ECMWF operational analysis.
SCIAMACHY v1.4
SCIAMACHY
was a national contribution to the payload on ESA's Envisat Satellite, which was
launched in March 2002. Envisat was placed in a sun-synchronous orbit at 800 km altitude with an equatorial crossing
time of 10:00 on the descending node. In the limb mode the SCIAMACHY instrument scans across the flight direction
with the total swath of 960 km and the centre of the scan displaced by a few degrees westwards from the flight direction.
This results in solar scattering angles ranging from 30∘ in the high northern latitudes to 150∘ in the
high southern latitudes with a strong latitudinal dependence. SCIAMACHY operation started in August 2002 and ended with
a sudden loss of communication with the Envisat satellite in April 2012. SCIAMACHY performed measurements in eight spectral
channels covering a wide spectral range from 214 to 2380 nm with a resolution varying from 0.2 to 1.5 nm. During its
mission, SCIAMACHY measured the solar radiation in nadir, limb scatter and solar–lunar occultation geometries and provided
daily measurements of the solar spectral irradiance that have been used to retrieve a variety of species including
aerosols, clouds, ozone, BrO, NO2 and water vapour. For this study stratospheric aerosol retrievals are performed
using the data from the limb scatter viewing geometry, where measurements are provided every 3.3 km with a vertical
resolution of 2.6 km in the altitude range from approximately 0 to 100 km.
The stratospheric aerosol extinction retrieval algorithm used in this study is an updated version of the algorithm
described by and . The SCIAMACHY v1.4 retrievals, herein referred to as the IUP
retrievals, use the newer version 8 SCIAMACHY Level 1 radiance data. Atmospheric pressure and temperature background
profiles from ECMWF operational analysis data from the specific date,
time and location of each SCIAMACHY limb measurement are used. In comparison to the previous version of the algorithm
and the USask retrieval algorithm, the updated v1.4 algorithm drops the shorter,
470 nm wavelength normalization to reduce the uncertainties related to measurement noise and lower sensitivity to
aerosols. The new measurement vector is given by
yjk=ln(I(λk,j))-ln(I(λk,jref)).
To reduce noise on the measurements, all measured wavelengths within ±2 nm of λk are used in the
retrieval. For the v1.4 extinction product the aerosol profiles are retrieved at 750 nm. The retrieval uses measurements
in the altitude range from around 12 to 35 km (depending on the latitude and season) with a reference tangent altitude of
about 38 km. The v1.4 aerosol extinction retrieval is performed on the measurement altitude grid, and the values below
and above the retrieval range are fixed to the a priori. Effective Lambertian albedo of the underlying surface is
concurrently retrieved based on the limb radiances near the reference tangent height to reduce the influence of clouds
below the field of view, although clouds within the field of view remain an issue. To reduce their impact, extinction
values greater than 0.001 km-1 are considered cloud contaminated and filtered after the retrieval is performed.
To solve the inverse problem an iterative regularized inversion approach similar to that described by
is used. As in it is assumed that the errors are uncorrelated, and the noise
covariance matrix is chosen to be diagonal. The signal-to-noise ratio is set to 200 for all tangent heights. For the a priori
covariance matrix the non-diagonal elements drop off exponentially with a correlation radius of 3.3 km and the
diagonal elements correspond to a relative standard deviation (SD) of 1.
Forward modelling, as well as retrievals, is done using the radiative transfer model with the retrieval code SCIATRAN-3.7
. The scattering phase functions are calculated using Mie scattering theory, assuming spherical
sulfate aerosol particles with a unimodal, lognormal size distribution. The refractive indices are calculated using the
OPAC database . At 750 nm the real component of the index of refraction is 1.427, and the
imaginary component is 7.170 × 10-8. The stratospheric aerosol parameters are defined from 12 to 46 km, where
it is assumed to consist of sulfuric droplets with 0 % relative humidity in the surrounding atmosphere. The previous
version 1.1 algorithm used a lognormal particle size distribution with a median radius of 110 nm
and width of 1.37, also consistent with in situ observations by . Although there is no evidence to
prefer either the particle size distribution used in the USask retrieval or that used by , using
different distributions complicates the comparison of limb-scattering retrievals, and so it is beneficial to make
a consistent choice for this work. Therefore, the version 1.4 SCIAMACHY product uses the same lognormal assumption as the
v5.07 OSIRIS product (rg = 80 nm, σg = 1.6). While a full validation of the version 1.4 is currently
ongoing, initial comparisons with version 1.1 show smaller uncertainty of the individual retrievals, reduced profile
oscillations and better parameterized upwelling radiation (resulting also in less sensitivity to underlying clouds) due
to the retrieval of albedo.
Coincident comparisons with SAGE II
The SAGE II was launched in 1984 and operated until November 2005, providing
one of the longest continuous records of stratospheric aerosols. As an occultation instrument, the SAGE II aerosol
retrieval is insensitive to many of the assumptions required in the limb scatter retrievals, making for a robust,
independent comparison. This work uses the version 7.00 SAGE II aerosol extinction data at 525 and 1020 nm
. Several improvements have been made since version 6.2 that have resulted in aerosol extinction
decreasing more quickly at higher altitudes. As both the OSIRIS and SCIAMACHY aerosol products are produced at 750 nm,
the SAGE II data are interpolated to this wavelength using the Ångström coefficient derived from the 525 and
1020 nm channels. Although this is not a perfect conversion, as the wavelength dependence is not strictly linear in
log-wavelength log-extinction space, the error is generally limited to less than 10 % for most particle sizes
. To test agreement between the three instruments a coincident comparison is performed when all
instruments have collocated measurements. Measurements are used when OSIRIS and SCIAMACHY observations are within
±5∘ latitude, ±20∘ longitude and ±24 h of the SAGE II tangent point. As limb measurements
have approximately 200 km path lengths through the atmosphere, and scanning of a vertical profile typically occurs over
a few degrees latitude, tightening these criteria does not generally improve agreement. To minimize the impact of clouds
in the analysis extinction, values greater than 0.0025 km-1 have been excluded. Due to the relatively
eruption-free period of this comparison, this has minimal effect on the comparisons removing approximately 3 % of scans
above 15 km and none above 20 km. This criterion provides 2580 coincident measurements between 2002 and 2005, when all
three instruments were operating. The comparison is broken into 20∘ latitude bins to better distinguish biases
related to latitude and solar geometry conditions. Results are shown in Fig. . In general, all
instruments agree to within approximately 15 % for most regions. Exceptions to this are at high altitudes and
latitudes (such as panels a, b and h) where both OSIRIS and SCIAMACHY retrieve lower values than SAGE II by
up to 40 % at 30 km. At latitudes above 40∘ N SCIAMACHY shows systematically higher results than SAGE II
for all altitudes below 30 km. This effect increases with latitude up to approximately 40 % at the highest northern
latitudes and is visible in panels g and h of Fig. . Although the largest clouds have been removed,
both limb scatter instruments are likely to still contain some cloud contamination near and below the tropopause and the
differences compared to SAGE II show large SDs in these regions.
Coincident comparison between OSIRIS and SCIAMACHY measurements compared to SAGE II. Difference computed as (Instrument - SAGE II)/SAGE II × 100 %. Shaded regions indicated one SD of the differences from the median.
Several factors are expected to contribute to the differences between the aerosol extinction retrieved from the
measurements of the occultation and limb scatter instruments, as well as the different biases between OSIRIS and
SCIAMACHY. Limb scatter inversions use complex forward models which are not identical in their assumptions or approaches.
The inversions themselves also differ in several ways, with SCIAMACHY using a regularized inversion technique and OSIRIS
using MART. A priori assumptions, such as the choice of aerosol particle size distributions and extinction profiles, also
affect the retrievals. The importance of these effects depends on the viewing geometry of the instrument. OSIRIS and
SCIAMACHY have significantly different viewing geometries as a result of the Envisat and Odin orbits. The following
sections explore the significance of these different effects.
Simulation study
To test the sensitivity of the retrievals to assumed parameters and retrieval settings, a series of simulation studies is
performed. The 2580 near-coincident scans from the SAGE II comparison are used as the test cases. These scans cover the
full range of OSIRIS and SCIAMACHY geometries. While these scans are limited to pre-2006, the simulations use a range of
atmospheric scenarios consistent with both background and volcanically perturbed conditions. Four factors are
investigated in this study: the impact of different radiative transfer models, a priori extinction profile and particle
size assumptions and choice of measurement vectors.
Comparisons of the radiative transfer models. Panel (a) shows the differences in radiance computed using SASKTRAN and SCIATRAN.
Panel (b) shows the difference in measurement vectors. Panel (c) shows the difference in retrieved profiles.
Differences in panels (a) and (b) are computed as (SASKTRAN - SCIATRAN)/(SASKTRAN + SCIATRAN) × 200 %.
Extinction error is computed as (retrieved - true)/true × 100 %.
Radiative transfer modelling
It is difficult to decouple the retrieval algorithms from the radiative transfer models entirely due to differences in
languages, input formats, and interfaces. However, differences between the IUP and USask retrievals due to the radiative
transfer models can still be estimated by simulating measurements using one model and retrieving with the other. For
this test, the SASKTRAN radiative transfer model is used to generate radiances that simulate the OSIRIS measurements.
These synthetic radiances are then used in the IUP retrieval which uses the SCIATRAN radiative transfer model. The same
is then performed with the SCIATRAN simulated radiances and the USask retrieval using SASKTRAN, again on OSIRIS
measurements. Although this is not a test of “correctness” of either model, nor a test of how well the radiative
transfer models could agree, it provides an indication of the magnitude of differences that should be expected due to the
configuration of the radiative transfer models as used in the retrievals. Figure shows the
differences in the modelled radiances and retrievals. Panel a shows the differences in the radiances at the 470 and
750 nm wavelengths. The radiances have systematic differences of approximately 5 %, with SCIATRAN producing larger
radiance values than SASKTRAN. Some of this difference is due to model resolution settings. SASKTRAN simulations are
performed at a higher vertical resolution of 1 km, and when both models use this higher-resolution vertical
grid the
agreement is improved to within 2–5 %. However, because the IUP retrieval is performed on a 3.3 km grid, the higher
resolution is not required for SCIAMACHY retrievals. Although the variation in radiances between the models can
occasionally reach 15 %, the normalizations used in the measurement vectors cancel much of the systematic differences.
This can be seen in panel b, where differences in the measurement vectors, computed using the two different models, are
shown. In this panel the red curve shows the percent difference between the IUP retrieval vectors defined in
Eq. () when computed from SASKTRAN vs. SCIATRAN radiances. The blue curve shows the same,
except
computed using the USask measurement vector definition from Eq. (). The high-altitude
normalization used in the IUP retrieval decreases the differences between the models to less than 2 % at most
altitudes. If the short wavelength normalization is included the difference is larger, typically near 5 %, since the
wavelength-dependent modelling differences vary more with altitude. How this difference translates to the retrieved
extinction is shown in panel c. Here, the red curve shows the difference in the IUP retrieved extinction using SASKTRAN
generated radiances compared to the true state. Similarly, the blue curve shows the same for USask retrieved extinction
using SCIATRAN-generated radiances. The IUP retrieval produces errors in the retrieved extinction less than 5 % for
most of the aerosol layer, with a SD close to 5 % as well. The larger differences in the USask measurement vector
lead to larger differences in the USask the retrieved extinction, although errors are still typically less than 10 %.
The exception to this is below 17 km and above 30 km, where the sensitivity to aerosol is low, and therefore small
changes in the radiative transfer cause large changes in the extinction. This highlights that the high-altitude
normalization is effective not only in minimizing errors due to uncertainties in unknown physical quantities such as
albedo but also in reducing errors due to model assumptions. Conversely, the short wavelength normalization has the
potential to introduce additional error if the radiative transfer model biases change with wavelength.
A priori profiles
The effect of the a priori profile on the retrieval is an important consideration and one that has the potential to vary
substantially between retrieval methods. Although the MART relaxation used in the USask retrieval has no regularization,
and the IUP retrieval is only weakly constrained by the a priori, the effect of the a priori at altitudes above the
retrieval range can still play an important role. The aerosol here can couple to the lower altitudes due to the high-altitude normalization of the measurement vectors. While this normalization has many benefits, it has the drawback of
coupling the error at high altitudes to all altitudes below. The USask retrieval scales the a priori above the retrieval
range, at each iteration of the retrieval to match the top retrieved value and thus avoid sharp discontinuities in the
retrieved profile. Therefore, the absolute error above the retrieval range depends on the shape of the a priori profile
at and above the normalization and the retrieved aerosol just below the normalization. Conversely, the IUP extinction is
fixed to the a priori value above the retrieval altitudes and so will depend less on the shape of the chosen a priori and
more on the absolute value in the normalization range.
The range of the true state aerosol profiles is shown as the shaded region. The USask a priori is shown in blue and the IUP in red.
The effect of the a priori above the retrieval range is tested through a simulation study where the true high-altitude
aerosol profile (i.e. the input profile used to generate the synthetic measurements) differs from that assumed in the
retrievals. For this test an exponentially varying aerosol profile above 30 km is taken to be the truth. The slope of
the exponential profile is then varied for each simulated OSIRIS and SCIAMACHY scan. The range of exponential profiles
used as true states in the simulations is shown as the grey shaded region in Fig. . The USask
and IUP a priori values are shown as the blue and red lines respectively. The shape of the a priori profile below 30 km,
as well as all other aerosol parameters such as particle size, is assumed correctly in the simulated retrievals to avoid
introducing errors due to other retrieval parameters. The simulated data were then used to retrieve the extinction profile
using the USask and IUP retrievals under two conditions. First, both retrievals are initialized with the USask a priori
profile and, second, both are initialized with the IUP a priori profile.
Relative error in the OSIRIS data retrievals at 20 km as a function of the absolute error in the true extinction at the reference point. The solid lines show the least squares fit to the data.
Figure shows the relationship between errors at the reference altitude to errors lower in
the profile for four cases. The top row shows results for the USask retrieval with the bottom row showing the IUP
retrievals. The left column shows results when the USask a priori profile is used for the retrievals, with the right
column showing results when the larger IUP a priori is used. The solid line shows a linear best fit to the data.
Generally, if aerosol is overestimated in the normalization range, due to an a priori profile that decays too slowly with
altitude, the aerosol is overestimated for the entire retrieval. This is because the modelled vector is normalized by
an overly large value, decreasing the magnitude in the retrieval range; as a result, extra aerosol is added to compensate.
The error in the retrieved aerosol is very well correlated with the error in the normalization range, with little
dependence on whether the USask or IUP retrieval is used. This holds well for all geometries tested and for both
retrieval algorithms. However, higher altitudes are more sensitive to aerosol loading, and so show a larger error in the
retrieved profile for a similar absolute error in the a priori as the normalization altitude is increased. This can be
seen in the larger sensitivity to a priori errors in the IUP retrieval, which uses a 38 km reference height, as opposed
to the USask retrieval that used 35 km. The same error of 10-6 km-1 at a normalization altitude of
38 km will cause approximately twice the error that it does at 35 km. At low altitudes, less than approximately 14 km,
the sensitivity to aerosol is very low and the retrievals no longer show a clear relationship between the retrieval error
and the a priori error.
Percent error in the OSIRIS data retrievals as a function of altitude relative to an extinction error of 10-6 km-1 at 35 km. Solid lines show values computed from the best fit line from the simulation studies shown in Fig. . Dashed line shows the error expected from the linear error analysis of Eq. ().
The altitude dependence of the retrieved error, normalized by the error at 35 km is shown in
Fig. . We note that normalizing the IUP retrieval by the error at 35 km is not
strictly correct as the reference altitude is at 38 km. However, this allows for a consistent comparison between the two
algorithms, and due to the relatively linear nature of the error it is not expected to introduce large biases. The
retrieval error is smallest at around 22 km, where the aerosol loading is highest, and the measurement sensitivity is
still quite good, with error increasing above and below this altitude. The error can also be estimated without simulating
the full retrieval using the equation
δk=Gδy,
where δk is the error in the retrieved extinction, G is the gain matrix or the sensitivity of
the retrieved extinction to variations in y, and δy is the error in the measurement vector. In this
case, δy is the error in the measurement vector due to errors in the assumed aerosol at the normalization
altitude and above. As the retrieval error is quite linear with respect to errors in the high-altitude profile,
δy in the retrieval range can be calculated directly from the Jacobian matrix, K. This
analysis as applied to the USask retrieval is plotted in Fig. as the dashed line.
Agreement between the analytic method and simulation study is excellent over the full range of values
tested. As G and K are typically readily available from the inversion method, this can also be
applied on an operational basis if estimates of the extinction error at the normalization point are known.
Particle size
In the standard extinction retrievals the aerosol optical properties are not retrieved and must therefore be assumed when
retrieving extinction. Of primary importance in the IUP, USask and OMPS retrievals is the assumption of a fixed particle
size. All three retrievals assume lognormal distributions that correspond to typical background conditions as measured by
, albeit with somewhat different lognormal parameters. The lognormal distribution used in the
retrievals is given by the equation
n(r)=N2πln(σg)rexp-(ln(rg)-ln(r))22ln2(σg),
where rg is the median radius, σg the distribution width and N the number density. During background
conditions the median radius is generally larger than 40 nm but less than 200 nm, depending on altitude. However, after
volcanic eruptions, a second mode of particles with median radii up to a few microns may be present, further complicating
the analysis. The effect of this constant unimodal particle size assumption was estimated to a degree by
, but a limited number of geometries and cases were tested. More recently,
estimated the impact of particle size assumptions based on estimates of the phase function, but they did not fully propagate
the error through the retrievals. This work extends these previous analyses to additional conditions and geometries and
estimates the impact on the retrieved extinction.
The range of particle sizes tested as a function of altitude. Panel (a) shows the fine-mode parameters and panel (b) the coarse mode. The blue lines show the USask and IUP a priori values assumed in the retrievals. The grey shaded region shows the range of values used in the simulations.
To estimate errors due to particle size assumptions two sets of simulations are performed. First, a study to estimate
errors in the retrieved extinctions during relatively quiescent periods is done, when only a fine mode of aerosols is
present. For these simulations, the fine-mode lognormal parameter profiles as measured by the OPC in Wyoming by
between 2001 and 2014 are used as inputs for the simulated data. This provides 44 unique particle size
profiles. To avoid noise and high-frequency oscillations the OPC profiles are smoothed to a vertical resolution of
approximately 3 km. The extinction profile was set to twice that of the a priori assumption, with no change in the shape
to avoid including a priori errors in this portion of the study. The second set of simulations covers conditions more
representative of those after volcanic eruptions, when an additional mode of larger particles is present. For this case,
the smoothed coarse mode as measured by the OPC is also added to the true extinction profile. The number densities of the
fine and coarse modes are set such that the coarse mode accounts for 70 % of the total extinction. In each case, the
coincident OSIRIS and SCIAMACHY scans were simulated using a random OPC particle size profile and a random albedo between
zero and one as the true state. Figure shows the range of median radii, widths and
Ångström exponents (calculated between 525 and 750 nm) used in the simulations, as well as the a priori values.
Error in the retrieved USask extinction for the simulated SCIAMACHY measurements grouped by the scattering angle for four different cases.
The top two rows show results when only a fine mode of aerosols are present; the bottom two rows have both a fine and coarse mode.
In both cases retrievals are done with a short wavelength normalization (a and c) and without (b and d).
The colour of the lines indicates the Ångström coefficient.
The standard USask algorithm was then used to retrieve extinction with the simulated data. These retrievals were also
repeated using the USask algorithm but without the short wavelength normalization to determine its effect. The top row of
Fig. shows the relative error in the retrieved extinction for the standard USask retrieval when
only a fine mode of particles is present, grouped by scattering angle. The colour of the line indicates the Ångström coefficient. Only the SCIAMACHY geometries are shown here, as the OSIRIS results are very similar, but with
a reduced range of scattering angles. Generally, errors are largest in the strongly forward and backscattering cases,
with a strong dependence on the Ångström coefficient. The assumed size distribution has an Ångström
coefficient of 2.3, and consequently when the true state is near this value the retrieval has little error. At altitudes
above 25 km, however, this assumption is consistently too high and leads to large errors, particularly in strongly
forward scattering conditions.
The second row of Fig. shows the same, but when the retrieval does not use a short wavelength
normalization. In this case, the error is reduced in forward scattering conditions but increased in backscatter,
particularly at lower altitudes, where sensitivity to aerosol is poor. The third row shows retrievals when the true state
includes a second coarse mode of particles. In this case the assumption of an Ångström coefficient of 2.3 is
generally more accurate at higher altitudes, and so the error above 20 km is reduced compared to the fine-mode-only case.
However, the dependence on Ångström coefficient is weaker for the bimodal distributions, with many different
particle sizes producing comparable errors. The effect of normalization is also not as clear under these more volcanic
conditions, with only strongly forward scattering geometries showing a clear preference for no wavelength normalization.
Error in the retrieved USask extinction as a function of Ångström coefficient at 20 km.
The colour of the points shows the solar scattering angle. The top row shows the error for conditions when only a fine mode of aerosol is present.
The bottom row shows the error when there is both a fine- and coarse-mode distribution. The black dashed line indicates the Ångström
coefficient corresponding to the particle size distribution used in the retrievals.
This dependence on Ångström coefficient and scattering can be seen more clearly in Fig. ,
which shows a cross section of the results in Fig. at 20 km, as well as the results from OSIRIS
geometries. Each panel shows the relative error in the retrieved extinction as a function of the true Ångström
coefficient at 20 km. The colour of each point indicates the scattering angle of the measurement. Panels a and b show
results for the fine-mode-only simulations, while c and d show results from bimodal cases. Panels a and c shows results
from OSIRIS geometries, and those from SCIAMACHY geometries are presented in panels b and d. The retrievals without the
short wavelength normalization are shown in the right four panels. If only fine-mode particles are included in the
simulated atmosphere, the error in the retrieval can be well parameterized by the Ångström coefficient and the
solar scattering angle of the measurement. When the Ångström coefficient is assumed correctly the error in the
retrieval is less than 10 %, nearly independent of the particular lognormal parameters. As the error in the
Ångström coefficient increases, so does the error in the retrieval, up to 100 % for OSIRIS geometries. For
SCIAMACHY geometries the range of scattering angles and errors can be larger due to larger variations in the aerosol
phase function at extremely large and small angles. With a short wavelength normalization the retrievals show errors that
are mostly symmetric around zero. While this will help to reduce biases over longer periods of time when a large range of
scattering angles are sampled, seasonal biases are still to be expected as different scattering angles are sampled over
the course of a year. Similarly, latitudinal biases are likely in the SCIAMACHY data as scattering angle depends strongly
on latitude. Without a short wavelength normalization the general spread and shape of the errors is similar; however, the
errors are not centred around zero with aerosol being overestimated more often than not. In this case, the error is
minimized during forward scattering conditions when scattering angles are near 60∘. When short wavelength
normalization is used the error is at a minimum near 90∘; subsequently the error for forward scatting geometries
is increased, while it is decreased for backscattering geometries.
When coarse-mode particles are included, the phase functions can vary more widely for a given Ångström
coefficient, leading to less of a clear relationship in the retrieved error. This can be seen in panels c and d of
Fig. , where much weaker correlation between the Ångström coefficient, solar scattering angle
and extinction error is visible. Even when the Ångström coefficient is assumed correctly, differences in the
lognormal parameters can induce errors of 30 % in the retrieval for OSIRIS geometries and 50 % for SCIAMACHY
geometries. While the error is less correlated, errors are not systematically larger than during volcanically quiescent
periods, but do have a tendency to introduce low biases in the retrieved results for most geometries and particle sizes.
Additionally, while backscatter can still have large biases, they are not as large at the extreme scattering angles as
during fine-mode-only conditions. During bimodal conditions the error in both the normalized and non-normalized
retrievals is comparable, except during strongly forward scattering conditions when the short wavelength normalization
increases the error. In general, this shows that the short wavelength normalization is beneficial during background
periods under backscattering conditions, but generally increases the error during forward scatter. Additionally, in
forward scatter both the 470 and 750 nm wavelengths are positively sensitive to aerosol, so the wavelength ratio will
tend to decrease the sensitivity to aerosol and decrease the retrieved precision due to measurement noise as well.
Retrieval study
In Sect. the sensitivity to retrieval assumptions and radiative transfer modelling was estimated. In
this section, we explore the applicability of the USask retrieval to the SCIAMACHY measurements and vice versa, both to
confirm the simulation studies and to better understand the sensitivity of the retrievals to differences in the radiance
products. The same set of coincident SAGE II scans is used for this study, with comparisons performed in the same way as
those presented in Sect. .
Coincident comparison with SAGE II when both OSIRIS and SCIAMACHY measurements have been processed with the USask algorithm.
Figure shows the USask retrieval applied to both instruments. Retrievals using the
SCIAMACHY measurements agree very well with those using OSIRIS and show many of the same biases with respect to SAGE II.
Both instruments show underestimation with respect to SAGE II at high altitudes and latitudes. If this was due to
inaccuracies in the assumed particle size the error would be expected to change signs between hemispheres as the SCIAMACHY
solar scattering angle goes from backscattering to forward scattering, which is not the case. Instead, these high-altitude errors are more likely to be caused by errors in the assumed a priori extinction profile at high altitudes where
the measurements are normalized, as the effect of this is nearly independent of solar geometry. From
Fig. errors of 3 × 10-6 km-1 in the reference altitude range
could explain biases of -30 % at high altitudes. Additionally, both instruments have some stray light at these
higher altitudes that increases the radiance signal. This changes the shape of the aerosol measurement vector and is
likely a contributing factor to the low biases at high altitudes and latitudes. Unfortunately, both a priori and stray
light errors have similar systematic biases on the profile making them difficult to separate except in simulation, and
errors in the a priori can either help to cancel or exacerbate errors due to stray light. The shift in the SCIAMACHY
measurements from low biases in the Southern Hemisphere to high biases in the Northern Hemisphere is present, as was seen
in the IUP retrieval in Fig. , again suggesting a particle size error. In the USask retrieval this
shift is approximately 20–30 % between hemispheres, which from Fig. would be consistent with an
overestimation of the Ångström coefficient by approximately 0.3, i.e. an assumption that particles are too large at
the high latitudes.
Coincident comparison with SAGE II when both OSIRIS and SCIAMACHY measurements have been processed with the IUP algorithm.
The IUP retrieval applied to both the SCIAMACHY and OSIRIS data is shown in Fig. . OSIRIS
solar scattering angles do not vary as strongly between the northern and southern hemispheres, and so the OSIRIS
retrievals do not exhibit the same shift from low biases in the south to high biases in the north that are seen in the
SCIAMACHY measurements. The impact of the a priori choice can also be seen here. For the OSIRIS retrievals the USask
a priori was used without scaling, resulting in low aerosol values in the normalization range and leading to lower aerosol
values at all altitudes. However, if the IUP a priori is used the retrievals are substantially higher when compared to
SAGE II (not shown). This is consistent with the results from Sect. , in that larger a priori values in
the normalization range lead to larger values at all altitudes.
This highlights the sensitivity to the chosen a priori and reference altitudes and the limitations of both the USask and
IUP approaches. The USask technique of scaling an a priori profile that decays rapidly with altitude works with both
instruments provided the normalization altitude is chosen to minimize stray light. The variable normalization altitude
ensures there is sufficient aerosol signal to determine the scaling, while the quickly decaying profile ensures the
measurement vector is only weakly dependent on the scaling applied. However, while this provides a relatively robust
retrieval it is likely to cause the aerosol to be underestimated at the normalization point, leading to low biases in the
retrieved extinction, particularly at high altitudes. Conversely, the larger fixed a priori used in the IUP retrieval
works well for SCIAMACHY when an appropriate reference altitude is chosen and can reduce biases at high altitudes.
However, it yields poor results when applied to the OSIRIS measurements, illustrating the necessity of properly matching
the normalization altitudes with the stray light characteristics and choice of a priori when using a fixed a priori
profile. Together, the stray light, choice of normalization altitudes and a priori profile in the normalization range
have a complex interplay. This can be seen panels a, b, g and h, where the OSIRIS biases at
low altitudes are reduced compared to the USask retrieval (Fig. ), despite not improving
the retrievals at high altitudes. Conversely, the biases are increased elsewhere (panels c–f). Unfortunately, without
more detailed knowledge of the stray light and error in the extinction in the normalization altitudes, the relative
contribution of each cannot be determined.
Conclusions
The updated SCIAMACHY v1.4 aerosol extinction product shows good agreement with coincident SAGE II measurements, typically
within 20 % for most regions. Exceptions to this include high northern latitudes where larger positive biases of
20–40 % are present and altitudes above 25 km in the southern high latitudes where negative biases are present.
The differences between the limb and occultation measurements are well explained by two primary causes. First, the choice
of a priori profiles is important in the limb retrieval due to the high-altitude normalization. If the shape of the
a priori profile is assumed incorrectly in the USask retrieval the scaling applied to the profile in the retrievals will
produce incorrect aerosol in the reference altitude, resulting in biases at all altitudes. The IUP retrieval fixes the
aerosol profile above the retrieval range to the a priori value and errors couple similarly to lower altitudes. For both
retrievals extinction errors in the reference altitude of 10-6 km-1 lead to errors in the retrieved
extinction of 5 % near the aerosol peak and up to 20 % just below the reference altitude. Second, incorrect
particle size generally shows a small mean difference when averaged over a range of scattering angles, but can have large
differences of 100 % or more for individual cases, particularly for strongly forward and backscattering viewing
conditions. This is especially important for orbits that systematically sample solar scattering conditions as a function
of latitude. Simulations including a coarse mode of particles suggest a low bias in the retrieved extinctions during
volcanically perturbed periods is likely for most geometries. However, the magnitude of the error is not expected to be
systematically larger than the during background conditions on a profile-by-profile basis. Additionally, while the USask
and IUP retrievals use the same particle size assumptions, the biases are different for both the instruments and retrieval
algorithms due to the difference in viewing geometries and definition of the measurement vectors. The error due to
particle size can be reduced in backscatter geometries through the short wavelength normalization. However, this
normalization has the opposite effect in strongly forward scattering conditions, where it makes the retrievals more
sensitive to particle size assumptions and measurement noise. Differences in SASKTRAN and SCIATRAN radiative transfer
models can cause systematic differences of up to 10 % between the retrieved products and may explain some of the
vertical structure in the comparisons, but they are not expected to be a primary driver of the differences.
Future retrievals would benefit from improved a priori estimates of the aerosol extinction above 30 km and particle size
distributions. In particular, OSIRIS retrievals could benefit from larger assumed a priori values at higher latitudes to
reduce low biases compared to SAGE II. SCIAMACHY retrievals would benefit most from improved particle size estimates to
reduce north–south biases. However, if this information remains limited, careful use of wavelength normalization (and
the lack thereof) for specific viewing geometries has the potential to reduce retrieval biases. Additionally, although
the USask and IUP approaches to aerosol in the normalization range of the measurements are different (scaling vs. fixed to
a priori respectively), both show comparable errors in the retrieved product for a given error in the normalization range.
Robust measurements of high-altitude aerosol are therefore needed to establish whether a fixed a priori or a scaled one
leads to less error at these altitudes. In summary, this study investigates the retrieval of extinction from the limb
viewing observations of scattered solar radiance by the satellite borne instruments OSIRIS and SCIAMACHY. It provides
a detailed analysis of our understanding of the systematic errors associated with these data products and biases with
respect to the SAGE II measurements of extinction.