(A)RC RADIATIV TRANSFER AND SST RETRIEVAL
Owen Embury, Chris Merchant
University of Edinburgh, Crew Building, The King's Buildings, West Mains Road, Edinburgh, EH9 3JN. UK
Email: [email protected]; [email protected]
ABSTRACT
(A)ATSR sea surface temperature retrieval coefficients
are currently derived from radiative transfer
simulations. This has contributed to the excellent
accuracy of the (A)ATSR SST products (errors ~ 0.2 to
0.5 K). However, the “(A)ATSR Re-analysis for
Climate” project has a target of reducing errors to less
than 0.1 K. We aim to achieve this partly by improving
the radiative transfer simulations on which retrieval
coefficients are based, using updated models and
spectroscopy, such that the observed biases better match
the expected prior and non-linearity errors. Another
improvements is in the method of retrieval, and we
identify an artefact of the current retrieval algorithm
which is causing biases ~0.1 K.
1.
INTRODUCTION
(A)RC – (Advanced) along-track scanning radiometer
(ATSR) Reanalysis for Climate – is a five year project
to produce a sea surface temperature (SST) record
suitable for climate change research from the three
(A)ATSR instruments flow on board the ERS and
ENVISAT satellites. In order to be suitable for such
research, the SST record must meet several stringent
requirements. It must span at least 15 years, with a
stability better than 0.05 K decade-1, have regional
biases < 0.1 K, and be independent of existing SST
analyses and in situ records to the greatest extent
possible. The (A)ATSR series of instruments have
several features which make them uniquely suitable to
this task. Firstly they are exceptionally well calibrated;
with two on-board blackbody reference targets,
precisely measured spectral response functions, and low
levels of noise. Secondly their dual-view capabilities
allow an improved correction for atmospheric effects.
Finally, the instruments have been operating for a
sufficient length of time, with adequate overlap periods.
Current SST retrievals from (A)ATSR data have a
scatter of 0.2 – 0.5 K and regional biases up to 0.3 K
[1]. In order to meet the requirements of the (A)RC
project the SST retrieval must be significantly
improved. There are several aspects to this task:
improvements to the cloud screening, robustness to
tropospheric aerosol, correction for prior error and nonlinearity, and refinement of the linear retrieval
coefficients. Here we focus just on the linear retrieval
coefficients.
_____________________________________________________
Proc. ‘Envisat Symposium 2007’, Montreux, Switzerland
23–27 April 2007 (ESA SP-636, July 2007)
The SST retrieval scheme used for the (A)ATSR
instruments is a linear equation where the observed
brightness temperatures are multiplied by a set of
coefficients. These coefficients are calculated based on
radiative transfer (RT) simulations of a range of
atmospheric and surface conditions sampled from a
numerical weather prediction (NWP) model. In order to
have the best available set of coefficients, the radiative
transfer model (RTM) must be as accurate as possible,
and the sample of atmospheric profiles must be
representative of real conditions.
In this paper we describe the new RTM used for the
(A)RC project and the new atmospheric profile dataset
in Section 2. Section 3 discusses the effect of scattering
– important to account for the effects of stratospheric
and marine aerosols. The RTM output is compared with
actual (A)ATSR observations in Section 4. Finally, the
linear retrieval is discussed in Section 5 and 6.
2.
RADIATIVE TRANSFER MODELLING
2.1. Line-by-Line Absorption Calculations
The current (operational) SST retrieval coefficients
were derived from RT simulations performed using a
specially written RTM which used look-up tables of
atmospheric absorption at fixed pressures and
temperatures based on the 1996 version of the
HITRAN[2] spectroscopic database.
Due to the increase in computing power over
the last 15 years it is now feasible to use a full line-byline model for all the RT simulations. We are using the
Reference Forward Model (RFM) – originally
developed as the reference forward model for the
MIPAS instrument. RFM is based on GENLN2 and
both portable (written in standard FORTRAN-77), easy
to use, and actively maintained at the University of
Oxford, UK.
We use RFM to calculate monochromatic
level-to-space absorption spectra at a high spectral
resolution of 0.01 cm-1. These are used along with the
surface emissivity model (see Section 2.5) to calculate
top-of-atmosphere (ToA) clear-sky radiance spectra.
Finally the radiance spectra are convolved with the
(A)ATSR spectral response functions (SRFs) to
determine channel integrated radiances and brightness
temperatures (BT).
We use the line parameter database available from
Atmospheric and Environmental Research Inc
$% %$&# '
The current SST retrieval coefficients[3] are based on
RT calculations using a dataset of 1358 atmospheric
profiles sampled from the ERA-40 database. These were
extracted from 16 global snapshots comprising analyses
at four dates and times (1st at 0000, 7th at 1200, 15th at
0600, and 23rd at 1800) from four months (October
1991, January 1992, April 1992, July 1992). Profiles on
a 15 degree horizontal grid were selected where they:
- corresponded to ocean grid cells
- had surface temperature greater than 271.35 K
- had maximum relative humidity less than 95%
- had minimum relative humidity greater than 0%
There are a number of ways in which this profile set is
improved within (A)RC. Firstly, the sampling strategy
leads to an under-representation of high-latitude cases
due to the surface temperature constraint (this is
discussed further in the next Section). Secondly, the
coarse 15 degree grid means there are no profiles from
certain geographical areas such as the Mediterranean
Sea, Red Sea, and the Gulf of Mexico. Finally, the
profiles were extracted from a database containing data
on only 16 fixed pressure levels, rather than the full 60
levels used in the ERA-40 model.
For the (A)RC project we use the 60L-SD
database of ERA-40 profiles described in [4] as a
starting point. This comprises 13,495 profiles extracted
from 1991 and 1992 of the reanalysis. The sampling
strategy employed by [4] ensured that the full range of
variability of temperature and water vapour in the
ECMWF reanalyses is represented in the output dataset.
Furthermore, the profiles were extracted at the full
resolution of the model grid.
Clear-sky open-ocean profiles were selected from
the 60L-SD dataset using a similar strategy to [3],
except we replace the surface temperature constraint
with:
- sea ice coverage less than 95% (see section 2.3 for
further details)
The resulting profile set had fewer high-latitude winter
profiles than high-latitude summer ones. In order to
ensure that all trace gas conditions were sufficiently
represented (see Section 2.4), further ERA-40 profiles
were extracted to give an even latitude and monthly
distribution of profiles.
( )* #' %$&#
The surface temperatures reported in the ERA-40
dataset are grid-cell averages. In polar regions this
includes contributions from both the open ocean and sea
ice, such that the average temperature is often less than
271.35 K. This means such polar profiles were rejected
by the sampling scheme and high-latitude conditions
were poorly represented in the profile datasets.
Reference [5] investigates using the sea-ice
coverage to calculate a realistic ocean SST using an
empirical relationship tuned to observations. Eq. 1
shows the formula used in [5].
1 − ice n1 n2
Ts = T +
+
0 T −T 5
5
0 1
(1)
Where T0=2 °C and T1=-1.8 °C are the end points of the
interpolation representing ice-free and ice-covered
conditions respectively; ice is the sea ice coverage (0 to
1); n1 and n2 are random variables sampled from a
Poisson distribution with mean of 5. An example
distribution of generated SSTs is shown in Fig. 1.
280
278
SST / K
(http://rtweb.aer.com).
This
is
equivalent
to
HITRAN2001 with updates to the CH4, CO, O3,
HOOCH, O2 parameters, and the HITRAN2004 HNO3
lines. This is used in preference to HITRAN2004 as the
MT_CKD water vapour continuum model used in RFM
was built using the AER line parameter database.
276
274
272
0.2
0.4
0.6
Ice Coverage
0.8
Figure 1. Distribution of ocean skin temperatures for
sea ice contaminated profiles
+ ' & ,' *
In addition to water vapour several other atmospheric
gases can have a significant impact on ToA BTs. For
the (A)RC project we include any gas which has an
effect greater than 0.001 K on any of the (A)ATSR
infra-red channels. These gases were identified by
running the RFM model using the mid-latitude day-time
MIPAS model atmosphere (similar to the US Standard
FASCODE profile) and comparing results where each
gas was included/excluded in turn.
Several of these gases also have significant
geographic or seasonal variations which cause
detectable changes in ToA BTs. Furthermore, the
average concentration of several of these gases – such
as carbon dioxide and CFCs – have changed during the
last 15 years. It is important to include these changes in
the RT modelling, otherwise they will cause biases in
the retrieved SSTs. Variable gases were identified by
+ $ '$,'
Although geographic/seasonal variation of CO2 does not
significantly affect ToA BTs, it is included as a variable
gas because increases in average concentration during
the lifetimes of the three (A)ATSR instruments do have
an impact ~0.02K on the 11 and 12 micron channels.
The MIPAS mid-latitude profile is scaled to a
concentration appropriate for the year being simulated.
+ -$
Variation in ozone concentrations only has a very small
effect on (A)ATSR BTs (<0.01 K). However, as profiles
are provided with the ERA-40 data they are used in
preference to a fixed profile.
+ ( $ $,' ' These two gases have a small impact on the 3.7 micron
channel: ~0.04 and 0.02 K respectively. Both gases have
higher stratospheric concentrations at the equator
causing a 0.05 K difference between polar and
equatorial conditions. There was little significant
seasonal variation and interpolation between the
different MIPAS model profiles based on latitude was
found to be represent their variability.
+ + '
Variation in HNO3 concentration can cause differences
up to 0.1 K in the 12 micron channel and 0.2 K in the 11
micron channel – larger than any other gas except water
vapour. Concentrations vary both with latitude and
season, from ~4 ppbv near the equator to over 20 ppbv
in polar winters.
In order to assess the expected variation in
HNO3 a climatology was created using data from the
EOS microwave limb sounder (MLS) instrument on the
Aura satellite [6] which has retrieved stratospheric
HNO3 profiles since August 2004. Fig. 2 shows the
maximum climatological concentrations both monthly,
and quarterly. Comparison of the profiles retrieved by
MLS, and the MIPAS model atmospheres showed good
agreement. Profiles for use in (A)RC RT simulations are
generated by interpolating between the MIPAS equator
10
15
5
5
5
5
Jan
Apr
Jul
2005
Oct
Jan
Apr
Jul
2006
Oct
DJF MAM JJA
Season
10
10
10
15
15
20
10
Oct
2004
10
-60
15
Latitude
5
5
0
20
Table 1. Additional gases included in RT simulations.
10
10
30
10
Variable
NH3 OCS H2CO N2 C2H6 F22
F113 F114 CCl4 HNO4
CO2 O3 N2O CH4 HNO3 F11 F12
15
15
60
-30
Fixed
10
and MIPAS polar winter model profiles based on the
climatological HNO3 concentration.
10
comparing simulations where the trace gas profile was
replaced with another MIPAS model atmosphere (polar,
equatorial etc.) while the temperature / water vapour
data were kept fixed.
The gases included in the (A)RC RT
simulations are listed in Tab. 1, the variable gases are
discussed in more detail in the following sections.
SON
PPBV
0
5
10
15
20
Figure 2. Peak stratospheric HNO3 concentrations from
EOS MLS. Left: time series. Right: seasonal climatology
+ . Two chlorofluorocarbons (CFC-11 and CFC-12) absorb
strongly in the long wavelength window, with latitude
variations causing ~0.02 K differences in the 11 and 12
micron channels. For these gases we interpolate
between different MIPAS model profiles based on
latitude
. & / $'#
The surface emissivity along with the temperature
determines how much radiance is emitted by the sea
surface. As we wish to retrieve the temperature from
measurements of the radiance, it is important to know
the emissivity as accurately as possible. Previous
(A)ATSR RT simulations [3] used the emissivity model
described in [7]. This model accounts for multiple
reflection effects in the forward view, where radiance
emitted by the surface may have been reflected by
another part of the surface into the satellite view
direction.
Here we use the same geometric approach to
emissivity modelling, but with the following
modifications:
- Emissivities are calculated at the same spectral
resolution as the line-by-line transmittance
modelling, rather than channel integrated values.
- Emissivity is calculated as a function of
temperature (in addition to wavelength, wind speed
and view angle).
Refractive index data from [8] and [9] are used to
calculate the temperature dependent emissivities.
( ! ( * The RFM atmospheric transmittance model discussed in
Section 2 does not include scattering and therefore
cannot account for the effect of aerosols. In order to
model these effects, we use the discrete ordinates
(DISORT) radiative transfer model [10]. The required
(A)RC BTs
4
4
4
3
3
3
2
2
1
3.7f-12f
5
1
2
1
0
0
0
-1
-1
-1
-2
260
-2
260
270
280
290
300
310
270
280
11
290
300
-2
260
310
RAD7 BTs
4
4
3
3
3
2
2
3.7f-12f
4
11-12
5
1
1
0
-1
-1
-1
-2
260
-2
260
300
310
270
280
310
300
310
1
0
290
300
2
0
11
290
RAD7 BTs
5
280
280
12f
5
270
270
12
RAD7 BTs
3.7-11
(A)RC BTs
5
11-12
3.7-11
(A)RC BTs
5
290
300
310
-2
260
270
12
280
290
12f
Figure 3. Comparison of RT simulated BTs and observed distributions from ATSR2 for RFM (top) and older RAD7
model. Contours show probability of observed. BT in K-2 (Values from outside: 10-6 10-5 10-4 10-3 10-2). Red line shows
best fit though simulations. Blue line shows best fit though observations.
input for the DISORT code is a profile of optical depth
and single scattering properties which can be easily
calculated from the RFM clear-sky optical depths
(transmittances) and an aerosol profile (given an
appropriate
model
of
aerosol
scattering
properties).However, full scattering calculations of the
type performed by DISORT are relatively slow and it is
not practical to run the code at the full spectral
resolution used for the line-by-line results. Therefore the
level-to-space line-by-line transmittances produced by
RFM are convolved with the (A)ATSR SRFs to
calculate channel-integrated clear-sky optical depths.
The DISORT code is then executed twice – once using
only the clear-sky optical depths, and once including the
aerosol profile. This allows us to calculate the
perturbation to brightness temperature (“delta-BT”)
from the aerosol, which can be added to the clear-sky
BT calculated directly from the line-by-line radiances.
( $$#
Optical properties of tropospheric aerosols are available
in the Optical Properties of Aerosols and Clouds
(OPAC) software package [11]. The package defines
several different aerosol types (i.e. marine, continental)
as a combination of various aerosol components (i.e.
sea-salt, soluble, soot) following the number density
profile given by Eq. 2
N ( z ) = N ( 0) e
−z
H
(2)
Where z is the altitude in km and H is the scale height.
We define the aerosol number density using the Global
Aerosol Data Set (GADS) [12] which is an aerosol
climatology defined in terms of OPAC aerosol
components and specifies N(0) for each component on a
5 degree grid.
(( $% $$#
After the 1991 eruption of Mt Pinatubo the increase in
stratospheric sulphuric acid aerosol initially caused
biases in the SST retrieval of ~1 K. Reference [3]
describes how updated SST retrieval coefficients were
made robust to the effects of volcanic aerosol.
In [3], aerosol optical properties were taken from
the MODTRAN radiative transfer program, which were
generic properties derived from observations of
previous eruptive events. For the (A)RC project we use
size distribution measurements relevant to Pinatubo
from Laramie, Wyoming [13] and Mie theory to
calculate the aerosol optical properties.
+ For initial validation of the RTM calculations the BTs
are compared with statistical distributions of observed
clear-sky BTs. ATSR2 data from the first ten days of
January, April, July, and October 1997 were
downloaded and all clear-sky ocean BTs were used to
generate 2D histograms comparing various channels and
channel differences.
Fig. 3 shows such histograms for three channel
combinations both for the new RT simulations, and the
previous data used in [3]. There is much better
(A)RC
RAD7
0.05
0.06
-0.01
-0.17
-0.31
-0.17
0.19
0.07
-0.08
-0.21
-0.51
-0.28
0.06
0.20
0.09
0.06
0.01
-0.15
0.2
0.0
-0.2
-0.4
New
Bias : -0.0389
s.dev: 0.2121
-50
Operational
Bias : 0.0159
s.dev: 0.2245
0
Latitude
50
Figure 4. D2-D3 differences for operational (red) and
experimental (black) coefficients. Crosses show
standard deviation in each bin. Solid line shows
frequency distribution of observations.
ATSR2 - 22 July 1996
Assumed relationship
1.75
Forward path length
Channel
difference
Nadir
3.7-11
3.7-12
11-12
Forward
3.7-11
3.7-12
11-12
Nadir-Forward
3.7
11
12
0.4
D3 - D2 / K
agreement between the observations and the (A)RC
simulations than with the RAD7 simulations, especially
at the lower temperatures corresponding to high-latitude
conditions. A very sensitive means of comparing
simulated and observed distributions is to look at the
differences between BTs in different channels or view
directions. Tab. 2 shows the average difference between
the simulations and observations for these colder
temperatures (BT at 11 µm < 275 K). All channel
combinations appear to be better modelled in the (A)RC
simulations except the forward view of the 11 micron
channel, which appears to be 0.1 – 0.2 K colder relative
to the other channels than is found in observations. This
is still being investigated, potential causes being the
assumed HNO3 concentrations (could be too high?) and
the revised emissivity model.
1.70
1.65
Table 2. Average simulations – observation differences
for 11 µm BT < 275 K
. The simulated BTs have been used to generate linear
SST retrieval coefficients similar to the current
operational coefficients. When applied to the Met Office
AATSR-buoy match up database (MDB), the new
coefficients marginally reduce the scatter compared to
the operational coefficients, and show a similar
latitudinal error structure (as expected, since the
technique to deal with systematic geographical errors is
yet to be defined). Fig. 4 shows the latitudinal Dual2Dual3 (D2-D3) biases for the operational and
experimental coefficients. D2-D3 biases are the
difference between the SST retrieval using the daytime
dual-view and night time dual-view coefficients. This is
a useful addition for verification to looking at satellitebuoy differences, as it does not rely on the accuracy of
the buoy measurement and allows us to compare with
spatial distributions of predicted biases.
1.00
1.02
1.04
Nadir path length
1.06
1.08
Figure 5. Actual ATSR2 view angles and assumption
used to generate operational coefficients (solid line).
0 1 ) ! 2 Current (A)ATSR SST retrieval coefficients are
generated for satellite zenith angles nominally relevant
to both the centre and edge of the swath. These are then
interpolated to the required pixel location assuming a
fixed relationship between pixel number and zenith
angles.
Examination of several orbits of (A)ATSR data
showed that the actual zenith angles do not follow the
assumed relationship. Furthermore, the swath is not
symmetric and there is considerable variation during an
orbit as shown in Fig. 5.
In order to assess the significance of this, simulated BTs
were generated using actual nadir/zenith angle pairs and
an SST retrieved using normal interpolated centre/edge
coefficients. Estimated Dual-3 biases are shown in Fig.
6, and there is considerable variation across the scanline
and during the orbit. Dual-2 biases (not shown) follow
the same pattern with approximately twice the
magnitude of the Dual-3 biases. Nadir retrieval
coefficients are not very sensitive – with biases
<0.005K, but following a different pattern to the dualview retrievals.
Bias due to zenith angle assumption
0.3
0.2
Equatorial
D3 bias (ret. - true) / K
Polar North
Polar South
0.1
-0.0
-0.1
-0.2
-0.3
0
100
200
300
Across track pixel number
400
500
Figure 6. D3 SST errors from operation zenith angle
assumptions in SST retrieval, for an example orbit.
For the (A)RC SST retrieval we will generate
coefficients for 30 nadir/forward combinations (5 nadir,
6 forward angles). These will be bi-linearly interpolated
to the angles of the target pixel.
(Note – the SST retrieval in Section 5 used the
standard centre/edge interpolations scheme. This was
necessary as the MDB is based on the 10 arc-minute
spatially averaged product which includes only the
(average) across track pixel number rather than full
nadir/forward zenith angle information.)
Another difference will be that in (A)RC, spatially
averaged SSTs will only be derived by averaging clearsky pixel SSTs, rather than retrieving SSTs from the
spatially averages BTs.
3 New radiative transfer simulations have been performed
for the (A)ATSR series of instruments. These
simulations show closer agreement with actual
(A)ATSR observations than the simulations used to
produce the operational SST retrieval coefficients.
The 11 micron forward view appears to be
underestimated in the new simulations; this is under
investigation. However, initial experimental coefficients
generated from these simulations perform comparably
to (or marginally better than) the operational
coefficients.
The operational retrieval scheme assumes that
nadir and forward view angles are a fixed function of
across-track pixel number. However, the actual view
angles vary during an orbit (especially in the forward
view) and this assumption can lead to across-track
dependent biases ~0.1 K in the D3 retrieval and 0.2 K in
the D2 retrieval.
4 1. Merchant C.J., Llewellyn-Jones D., Saunders R.W.,
Rayner N.A., Kent E.C. (2005). Sea surface
temperature for climate from ATSRs. In Proc.
MERIS-AATSR Workshop, Frascati, Italy
2. Rothman, L.S., Jacquemart, D., Barbe, A., et al.
(2005).
The
HITRAN
2004
molecular
spectroscopic database. Journal of Quantitative
Spectroscopy and Radiative Transfer, 96(2), 139204.
3. Merchant, C.J., Harris, A.R., Murray, M.J., Závody,
A.M. (1999). Toward the elimination of bias in
satellite retrievals of sea surface temperature 1.
Theory, modeling and interalgorithm comparison.
J. Geophys. Res., 104, 23565-23578.
4. Chevallier, F. (2002). Sampled databases of 60-level
atmospheric profiles from the ECMWF analyses.
NWP SAF Tech. Report 4. ECMWF, Reading, UK
5. Matthiesen, S., Merchant, C.J. (2003). Sea Surface
Temperatures in Marginal Ice Zones. Final Report
on Pathfinder Project. University of Edinburgh,
Edinburgh, UK.
6. Waters, J.W., Froidevaux, L., Harwood, R.S., et al.
(2006). The Earth Observing System Microwave
Limb Sounder (EOS MLS) on the Aura Satellite.
IEEE Trans. Geosci. Remore Sens., 44(5), 10751092
7. Watts, P.D., Allen, M.R., Nightingale, T.J. (1996).
Wind Speed Effects on Sea Surface Emission and
Reflection for the Along Track Scanning
Radiometer. J. Atmos. and Ocean. Tech., 13, 126141
8. Pinkley, L.W., Williams, D. (1976). Optical
properties of sea water in the infrared. J. Opt. Soc.
Am., 66(6), 554-558
9. Newman, S.M., Smith, J.A., Glew, M.D., Rogers,
S.M., Taylor, J.P. (2005). Temperature and salinity
dependence of sea surface emissivity. Quart. J.
Royal Met. Soc., 131, 2539-2557
10. Stamnes, K., Tsay, S.C., Wiscombe, W., Laszlo, I.
(2000). A General-Purpose Numerically Stable
Computer Code for Discrete-Ordinate-Method
Radiative Transfer in Scattering and Emitting
Layered Media, DISORT Report v1.1
11. Hess, M., Koepke, P., Schult, I. (1998). Optical
Properties of Aerosols and clouds: The software
package OPAC. Bull. Am. Met. Soc., 79, 831-844
12. Koepke, P., Hess, M., Schult, I., Shettle, E.P. (1997)
Global Aerosol Data Set, Report No. 243, MaxPlanck-Institut für Meteorologie, Hamburg, ISSN
0937-1060.
13. Deshler, T., Hervig, M.E., Hofmann, D.J., Rosen,
J.M., Liley, J.B. (2003). Thirty years of in situ
stratospheric aerosol size distribution measurements
from Laramie, Wyoming (41° N), using balloonborne instruments. J. Geophys. Res., 108(D5)