Using NWP model outputs to improve the retrieval of SST from MetopA AVHRR brightness temperatures

nullpitΔίκτυα και Επικοινωνίες

23 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

251 εμφανίσεις

1











Using NWP model outputs to improve the ret
rieval of SST
from Metop
-
A

AVHRR brightness temperatures


Igor Tomažić

AGO
-
GHER, University of Liège, Liège, Belgium


Visiting Scientist Report
OSI_AS12_02


2


Using NWP model outputs to improve the
retrieval of SST
from Metop
-
A

AVHRR brightness temperatures



Abstract

Classical multispectral sea surface temperature (SST) algorithms applied to infrared (IR)
radiometer data exhibit regional biases due to the intrinsic inability of the SST algorithm to
cope with the vast range of atmospheric types, mainly influenced by water vapor and
temperature profiles. Deriving a SST correction from simulated brightness temperatures (BT),
obtained by applying a Radiative Transfer Model (RTM) to Numerical Weather Pred
iction
(NWP) atmospheric profiles and first guess SST, is one of the solutions to reduce regional
biases. This solution is envisaged in the particular case of
Metop
-
A Advanced Very High
resolution (AVHRR) derived SST . Simulated BTs show errors, linked to
RTM
,

atmospheric
profiles or guess field errors. We investigated the conditions of adjusting simulated to
observed BTs in the particular case of the Mediterranean Sea over almost one year. Our study
led to define optimal spatio/temporal averaging parameter
s of the simulation observation
differences, both during day and night, summer and colder season and for two simulation
modes: operational (with reduced vertical resolution


15 levels
-

NWP atmospheric profiles
and two days old analysis used as first gues
s SST) and delayed (full vertical resolution


91
levels
-

and concurrent analysis used as first guess SST). Each BT adjustment has been
evaluated by comparing the corresponding corrected AVHRR SST to the AATSR SST, that we
adopted as validation reference.

We obtained optimized result across all defined conditions
and modes for spatial smoothing of 15 deg and temporal averaging between 3 and 5 days.
Specifically, analyses based on 10 day averages showed that a standard deviation based
criterion favors spati
al smoothing above 10 deg for all temporal averaging, while a bias based
criterion favors shorter temporal averaging during daytime (< 5 days) and higher spatial
smoothing (>10 deg) for nighttime. This study has shown also the impact of diurnal warming
bot
h in deriving BT adjustment and in validation results.



3


Contents


1

INTRODUCTION

................................
................................
................................
................................
..............................

4

2

DATA

................................
................................
................................
................................
................................
..................

8

2.1

D
ATA DESCRIPTION

................................
................................
................................
................................
..........................

9

2.1.1

AATSR

................................
................................
................................
................................
................................

9

2.1.2

Metop
-
A/AVHRR

................................
................................
................................
................................
..............

9

2.1.3

OSTIA

................................
................................
................................
................................
...............................
10

2.2

D
ATA INTERCOMPARISON

................................
................................
................................
................................
...............
11

2.3

D
IURNAL WARMING

................................
................................
................................
................................
......................
13

3

BRIGHTNESS TEMPERATU
RE SIMULATIONS

................................
................................
................................
...........
17

3.1

“O
PERATIONAL


SIMULATIONS
................................
................................
................................
................................
.......
17

3.2

“D
ELAYED MODE


SIMULATIONS

................................
................................
................................
................................
....
17

4

BT ADJUSTMENTS AND S
ST CORRECTION
................................
................................
................................
...............
24

4.1

BT

SIMULATION ADJUSTMEN
T METHOD

................................
................................
................................
..........................
24

4.2

SST

CORRECTION

................................
................................
................................
................................
..........................
28

5

RESULTS

................................
................................
................................
................................
................................
..........
29

5.1

O
PERATIONAL DA
YTIME
-

WHOLE PERIOD
................................
................................
................................
........................
29

5.2

O
PERATIONAL NIGHTTIME



WHOLE PERIOD

................................
................................
................................
...................
31

5.3

O
PERATIONAL TO DELAYE
D MODE


WHOLE PERIOD

................................
................................
................................
........
31

5.4

C
OLDER PERIOD


OPERATIONAL AND DELA
YED MODE

................................
................................
................................
.....
32

5.5

D
ISCUSSION
:

................................
................................
................................
................................
................................
.
32

6

CONCLUSIONS AND FUTU
RE WORK

................................
................................
................................
.........................
40

ACKNOWLEDGEMENTS

................................
................................
................................
................................
..........................
41

APPENDIX A
................................
................................
................................
................................
................................
..............
42

APPENDIX B

................................
................................
................................
................................
................................
..............
44

REFERENCES

................................
................................
................................
................................
................................
.............
46



4


1

Introduction


Real time simulated Brightness Temperatures (BTs) are increasingly used in operational Sea
Surface Temperature (SST) calculations, either in Optimal Estimation (OE) methods
[
Merchant et al.
, 2009;
Merchant et al.
, 2008]

or Bias Correction methods
[
Le Borgne et al.
,
2011;
Petrenko et al.
, 2011]
. BTs are simulated in the adequate Infrared window channels by
applying a Radiative Transfer Model (RTM) to Numerica
l Weather Prediction (NWP)
atmospheric profiles, using a guess SST field (guess SST) as surface temperature. These
simulations will be referred to as simBT
i

and the corresponding BTs, measured by the
radiometer onboard the satellite as obsBT
i
. Bias Correct
ion or OE methods result in
expressing the final SST as:


=
𝑔𝑢𝑒𝑠𝑠 
+


𝑖
(
𝑜𝑠𝐵

𝑖

𝑠𝑖𝑚𝐵

𝑖
)

(
1
)

In the case of bias correction methods envisaged here, a correction to a classical multispectral
algorithm is determ
ined by applying the coefficients of this algorithm to simulations to
produce a simulation derived SST. This “simulated” SST is compared to the “true” guess SST
and the difference is used as a correction term to the operational SST derived with this
algori
thm. In this case then, a
i

in equation
(
1
) are the coefficients of the multispectral
algorithm to which the correction is applied.

Simulations differ from observations
, because:

-

There are differences between the surface guess field and the actual SST field (guess
errors)

-

There are differences between the model atmospheric profiles and the actual profiles
(NWP errors)

-

RTM uses erroneous filter functions (filter errors)

-

RTM may be inaccurate (RTM errors)

-

Profile sampling induces errors (profile sampling errors). For instance the original 91
levels of the ECMWF outputs are sampled onto 15 pressure levels for some
operational applications at Meteo
-
France.


In the case of a bias correction method, guess errors are accounted for by equation
(
1
), but the
other sources of errors should be corrected prior to use simulations
in the SST correction
scheme. In practice adjustments are made by deriving empirical adjustment values from
comparing simulations and observations. Two approaches have been used: analytical
expressions of the simulation errors as a function of water vapor
and satellite zenith angle
[
Merchant et al.
, 2008;
Petrenko et al.
, 2011]

or a dynamic determination of the error
geographical distribution
[
Merchant et al.
, 2009b,
Le Borgne et al.,
2011
]
.

In the framework of the
European Organisation for the Exploitation of Meteorological
satellites (EUMETSAT), the Ocean and Sea Ice Satellite Application Facility (OSI
-
SAF) at
Météo
-
France/Centre de Météorologie Spatiale (CMS) in Lannion
has develo
ped a
geostationary satellite SST chain using simulated BT which became operational in August
2011
[
EUMETSAT
, 2011]
.

5


CMS has been running a prototype chain since November 2011 for testing the same approach
on
Metop
-
A
/Advanced Very High Resolution Radiometer (AVHRR) data. This prototype
aims to test methods that will be applied at full resolution in the Suomi National Pol
ar
-
orbiting partnership (NPP) Visible Infrared Imaging Radiometer Suite (VIIRS) and
Metop
-
A
/AVHRR) future SST processing chains. The prototype processing is done in near
-
real time,
to be as close as possible to operational conditions and uses OSTIA as surf
ace temperature,
ECMWF forecast profiles and the Radiative Transfer for TIROS Operational Vertical Sounder
model version 10
[
Hocking et al.
, 2011]
. Differences between simulated and observed BTs are
routinely monitored at CMS. These routine comparisons are done for pixels of quality level 3
or above in the GHRSST quality referencing system
[
Donlon et al.
, 2007]
, i.e. showing no or
little cloud contamination.
Figure
1

shows the daily g
lobal
Metop
-
A
/AVHRR
comparison
results for
the
year
2012.

Note that a similar monitoring is done at NOAA
[
Liang and Ignatov
, 2011;
Liang et al.
,
2009]
, and their results are d
isplayed through
http://www.star.nesdis.noaa.gov/sod/sst/micros/






(a)





(b)





(c)

Figure
1
: Mean
daily

differences and standard deviation
s

between simulated and observed
Metop
-
A
/AVHRR brightness temperatures

for 2012
, a) 3.7; b) 10.8 and c) 12.0

m
.


The mean daily biases and standard deviation values are quite stable over the
whole year
, the
step observed for
Metop
-
A
/AVHRR 3.7

m simulation
s on the 2
nd

of April 2012 (red circle
on
Figure
1
) is due to a RTTOV version change (from 10.1 to 10.2). On the contrary daily
6


differences are highly spatially variable, with a distribution of positive and negative patches a
priori difficult to interpret, as shown by
Figure
2

presenting an exa
mple of daily simulation


observation differences at 10.8

m on the 1
st

of March 2012 at 00:00 UTC, as produced by
the prototype monitoring system. 10.8

m has been selected as representing an intermediate
case between 3.7 and 12

m. The 1
st

of March has
been chosen randomly.




Figure
2
. Simulated


Observed 10.8

m BT at 00:00 UTC on the 1st of March 2012


To better understand the simulation


observation differences, data filtering was reinforced to
avoid as much as possible co
ntamination by clouds or by residual diurnal warming. This was
done by using SST quality level 5 only, and by excluding data for which the absolute
difference between BT simulations and observations is above 1.5 K and wind below 2 m/s.
These filtered data
have been averaged over one month (March 2012) and the simulation
minus observation difference geographical distribution shows clearer patterns (see
Figure
3
).



7



(a)


(b)


Figure
3
: a) Simulated


Observed 10.8

m BT for nighttime cases averaged over March
2012; b) model integrated water vapor under the same conditions


There is no obvious correlation between the BT differences and potentially influencing
parameters such as SST, water vapor content and latitude. Positive simulation errors are
associated with high water vapor content over the warm pool, but in the Atlantic

they coincide
with dry atmosphere north of the ITCZ.
Figure
4

shows the 3.7 and 10.8

m differences as a
function of the total water vapor content in nighttime condi
tions at 00:00 and 12:00 UTC on
the 1
st

of March 2012. Nighttime at 00:00 UTC corresponds to Atlantic, whereas nighttime at
12:00 corresponds to western Pacific. Over the Atlantic differences decrease sharply with
water vapor content; the opposite is obser
ved over the western Pacific. BT differences at 3.7

m show a much reduced trend in the same conditions, which confirms that it is indeed a
water vapor effect and not a cloud effect.



(a)

(b)





(c)



(d)

Figure
4
: Global nightt
ime differences between simulated and observed BTs at 3.7 and 10.8

m as a function of total water vapor content in nighttime on the 1
st

of March; a)

3.7

m at
00:00 UTC ; b)

10.8

m at 00:00 UTC; c) 3.7

m at 12:00 UTC ; d)10.8

m at 12:00 UTC


8


In consequ
ence we decided to derive adjustment values from the geographical distribution of
the errors averaged over space and time, as previously done in
[
Merchant et al.
, 2008]
. The
temporal and spatial scales of this averaging (how many days, how many pixels) have been
determined up to now through brief preliminary studies (for the
Metop
-
A
/AVHRR prototype
chain we adopted a time

averaging of 10 days and a space averaging of 10°) but were never
studied in details.
In this article we investigate the practical issues of deriving an adjustment
to the simulations by comparing simulations to observations
. A

previous study
[
T
omažić
,
2012]

demonstrated, over the Adriatic region, that using a fine scale regional model in place
of a global one (ECMWF) with a coarser resolution does not impact the simulations results in
the SST retrieval context. Similarly, the same study
[
Tomaži
ć
, 2012]

showed that the use of a
1 km resolution analysis brought no significant improvement compared to using the Met
Office operational analysis
[OSTIA,
Donlon et al.
, 2012]
.

In consequence in the present study
we have used OSTIA associated to
ECMWF profiles. They have been used in the CMS
prototype chai
n (an operational like environment) and in a delayed mode to use analysis
instead of forecast and 91 ECMWF profile levels instead of 15. We selected the
Mediterranean Sea not to redo simulations over the entire domain and because this area offers
a good cl
ear sky coverage with extreme conditions of diurnal warming and a good range of
atmospheric conditions. This choice was also motivated by the first regional application of the
method to VIIRS data acquired in direct readout mode at CMS and covering the Nor
th

East
Atlantic and the Mediterranean Sea. In what follows the two simulation sets will be referred
to as “operational” and “delayed mode” simulations. Various time and space adjustment scales
have been tested for each mode. Each time and space combinatio
n has been evaluated by
validating the SST bias correction deduced from the corresponding simulations by
comparison to the
Advanced Along
-
Track Scanning Radiometer (AATSR)

data, which were
in practice the only validation data available at a large scale ove
r the Mediterranean Sea. The
next section presents the data used (OSTIA, the
Metop
-
A
/AVHRR and AATSR derived SST
fields) and the main validation results by comparison with the AATSR data obtained over the
Mediterranean for selected period. The brightness t
emperature calculations are described in
section
0
. This section presents also the simulations minus observation statistics recorded for
the operational and delayed

mode data. Section
0

describes the method we adopted to test the
various adjustment options. In section 5 we describe the results obtained for averaging time
scal
es from 1 to 30 days and space scales from 0.5 to 50° in latitude
-
longitude degrees.

Section 6 will conclude.

2

Data


All data span the time range between June 2011 and March 2012. Three different sets of data
have been used:
Metop
-
A

12 hourly L3C files, AAT
SR L3C file and OSTIA. We have chosen
a 0.05° regular grid over the Mediterranean domain (see
Figure
5
) as a common grid for all
used datasets.



9



Figure
5
. 10 days (17 March till 26 March 2011) averaged difference between
Metop
-
A

and
AATSR for nighttime over the Mediterranean domain


2.1

Data description

2.1.1

AATSR

The AATSR sensor onboard ENVISAT has an advanced calibration system and dual
-
view
capability which leads to improved atmospheric correction procedure
[
Zavody et al.
, 1995]

and to reduced sensitivity to stratospheri
c aerosol
[
Merchant and Harris
, 1999]
. Accuracy of
t
he AATSR SST retrieval has been confirmed in numerous studies that used in situ
measurements for validation
[
Corlett et al.
, 2006;
O’Car
roll et al.
, 2006, 2008]
, or ship based
radiometry in dedicated validation campaigns to confirm absolute accuracy stability
[
W
immer
et al.
, 2012]
. Due to the lack of in situ data in Mediterranean and due to proven quality of
AATSR data we used AATSR data as a validation reference.

AATSR L2P files obtained from ESA in GHRSST format were remapped (to the nearest
neighbor) to the

common 0.05 deg grid. From all remapped orbits, L3C files were created
centered at 00:00 and 12:00 UTC time, with a ±6 hour time window centered at those times.
Due to the AATSR narrow swath (512 km) there were not overlapping pixels, and the only
criteri
a in deriving L3C file were actual existence of the data in the region of interest. Orbit
times over the area vary from East to West from 07:30 and 10:30 UTC by day and 19:30 to
22:30 UTC by night. This corresponds to about 10.5

0.1 h Local Solar Time (LST
) by day
and 21.5

0.5 h LST by night.

2.1.2

Metop
-
A
/AVHRR

Metop
-
A
/AVHRR SST products are derived operationally at OSI
-
SAF at full resolution over
the whole globe. By day (solar zenith angle smaller than 90°) a non linear split window
equation
(
3
) is used. A triple window equation
(
3
), based on the use of brightness
10


temperatures at 3.7, 10.8 and
12

m is used by night (solar zenith angles larger than 110°). In
the intermediate solar zenith angles (90 to 110°) a weighted mean of daytime and nighttime
algorithm is used to insure a smoothed transition.


NL:


=





+
(



+


)
(




-




)
+

+




E
2
F



T37:


=
(

+


)




+
(

+


)
(




-




)
+

+




E
3
F


q
3.7
, T
10.8
, T
12.0

are the brightness temperatures at 3.7, 10.8 and 12.0

m, respectively; S

=
sec
(

)

1,


is the satellite zenith
angle and T
CLI

is the mean climatological value.


The algorithm coefficients have been derived on a simulated brightness temperature data base
[
Francois et al.
, 2002]
.

Full resol
ution granules are delivered as operational products; they are also aggregated twice
daily from 18:00 to 06:00 UTC (centered on 00:00 UTC) and from 06:00 to 18:00 UTC
(centered on 12:00 UTC) and delivered on a global grid at 0.05° resolution
[
EUMETSAT
,
2010]
. In this processing,
Metop
-
A
/AVHRR SSTs full resolution pixels allocated to a grid
point are spatially average
d providing they have the same quality level. These gridded
products are stored internally at CMS as “workfiles” with observed BTs and ECMWF outputs
such as wind field and total water vapor content. The Mediterranean sector of these workfiles
constitutes t
he basic material of this study. Orbit times over the area vary from East to West
from 07:00 and 10:00 UTC by day and 19:00 to 22:00 UTC by night. This corresponds to
about 10

1 h LST by day and 21

0.3 h LST by night. Validation results are routinely made
available on the EUMETSAT OSI
-
SAF website (
http://www.osi
-
saf.org
) . They are also
available through the NOAA SST QUAlity Monitor (SQUAM) or published in research
articles
[e.g.
O’Carroll et al.
, 2012]
. The CMS prototype started on the 1
st

Novembe
r 2011
and the AATSR stopped in March 2012. To include a summer season in the studied dataset the
prototype has been run from archives for the June to August 2011 period.


2.1.3

OSTIA

OSTIA SST data
[
Donlon et al.
, 2012]
, obtained from NASA/PO.DAAC, were used as guess
SST field in simulating BTs. They represent “f
oundation” SST, free from diurnal warming
with the spatial resolution of ~6 km (1/20
°
). In a previous study
[
Tomažić
, 2012]

over Adriatic
Sea it was shown that using higher resolution (1 km) guess SST field (CNR UHR,
[
Buongiorno Nardelli et al.
, 2011]
) did not have significant impact on the improvement of the
simulat

ion qualities. A more elaborate global study
[
Saha et al.
,

2012]

showed that OSTIA
fields perform very well for the period when ENVISAT/AATSR were available (8
th

April
2012) and degraded afterwards.

The availability of OSTIA data in NRT applications is an important issue, as it is the case for
current
Metop
-
A

operational processing at Meteo
-
France/CMS. The closest OSTIA field in
time of the full coverage of
Metop
-
A

satellite overpass is available only with two days of
delay, which creates larger guess error compared to using OSTIA from the same day.




11


2.2

Data int
ercomparison


Metop
-
A
/AVHRR and OSTIA SST data sets have been compared to AATSR SST fields over
the Mediterranean for the June 2011 to March 2012 period, as follows:

-

AATSR skin have been converted to subskin values by adding 0.17 K
[
Donlon et al.
,
2002]

-

Saharan dust events are frequent ov
er this area in spring and summer. To eliminate
these cases, we have used the SEVIRI derived Saharan Dust Index
[
Merchant et al.
,
2006]

provided with the OSI
-
SAF SST products in this region. Based on CMS
experience we have discarded cases with SDI larger than 0.1.

-

Metop
-
A
/AVHRR and AATSR quality level 5 only have been used.

-

The maximum absolute tim
e difference allowed between AVHRR and AATSR data is
15 minutes in daytime to minimize diurnal effect (see discussion in section
2.3

below)
and 60 minutes in nightt
ime.

-

Daytime is defined by a Solar zenith angle smaller than 90 and nighttime by a solar
zenith angle larger than 110.

-

Due to cloud cover and AATSR narrow swath, AATSR and AVHRR coincident data
may sample different regions of the Mediterranean area each day. To obtain more
stable results, statistics are produced per 10 day periods (see
Figure
5
).

Figure
6

shows the mean differences per 10 day periods recorded over the domain in the
conditions descri
bed above, as well as the standard deviation and the number of cases.

In nighttime conditions, the agreement between the 3 sources of SST is good, with standard
deviations below 0.4 K (see
Table
1
). In early summer (June, July),
Metop
-
A
/AVHRR SST is
slightly positively and OSTIA slightly negatively biased compared to AATSR, while in winter
time
Metop
-
A
/AVHRR SST bias is slightly negative. This bias seasonality was observed
before over the Adriatic Sea, where comparison were done betwee
n different IR sensors
(NOAA17/AVHRR,
Metop
-
A
/AVHRR; MSG/SEVIRI, Terra/MODIS) and
ENVISAT/AATSR
[
Tomažić and Kuzmić
, 2011]

but also with compa
rison to in situ data for
the year 2003
[
Tomažić et al.
, 2011]
. Th
e main reason of higher bias seasonality corresponds
also to the observed water vapor seasonality
[eg.
Tomažić and Kuzmić
, 2009]

and SST
algorithm not capable to accurately resolve full range of atmospheric profiles.


Table
1
. Total statistics for the validation of
Metop
-
A
-
A SST compared to
OSTIA and AATSR
both for day and night.


DAY

NIGHT


bs
±
std

N

bs
±
std

N

OSTIA


AATSR

-
0.32
±0.57

411877

-
0.13
±0.40

1347258

Metop
-
A



OSTIA

+0.41
±0.57

9727643

0.06
±0.38

8928706

Metop
-
A



AATSR

+0.23
±0.59

343899

-
0.03
±0.29

1171778


In daytime conditions,
Metop
-
A
/AVHRR is positively biased compared to AATSR by about
0.2 K, whatever the season. OSTIA is negatively biased till the end of November. This bias,
which can reach

0.5 K in the midst of summer is due to diurnal warming and was expected
from a compar
ison between foundation (OSTIA) and skin (AATSR) SST (converted to
12


subskin values).
We also performed same intercomparison but with data filtered for the wind
speed, where we used only data with the wind speed above 6 m/s for daytime analysis and
above 2 m
/s for nighttime analysis
[
Donlon et al.
, 2002]
. Results a
re shown in Appendix A
where the main change is expectedly when compared against OSTIA fields (foundation
temperature) in daytime and during summer months.





(a)



(b)

Figure
6
. Time series of
Metop
-
A
, AATSR and OSTIA differences for
(a) daytime

and
(b)
nighttime.


13


2.3

Diurnal warming

Considering that we are using data in the Mediterranean Sea in summer time, we paid special
attention to diurnal warming. We analyze below the
Metop
-
A minus
AATSR error from June
to August 2011 through dependence on time difference, wind and model integrated total water
vapor column to detect a possible diurnal warming effect. First, a trend with time difference
between
Metop
-
A

and ENVISAT overpass was identif
ied for very low winds (< 2 m/s). As
shown by
Figure
7
,
the total biases range between
-
0.5 K for
Metop
-
A

passing 50
-
60 minutes
before ENVISAT and +0.
2 K for
Metop
-
A

passing just after ENVISAT. We also cross
compared this trend with a typical DW cycle over the Mediterranean Sea based on SEVIRI
data (
Figure
8
) averaged over the last 10 days in August. The mean warming rate is about 0.4
K per hour for low wind speed at 10 LST which is similar to the value obtained with direct
measurements between
Metop
-
A
/AVHRR and ENVISAT/AATSR. To minimize the impact o
f
this time difference we used a maximum absolute time difference of 15 minutes as threshold
for all daytime studies. Although in winter time the time difference could have been relaxed,
for consistency and simplicity we used the same threshold value throu
ght the whole period.

After having applied this time filtering, we still observe by day a persisting AVHRR
-
AATSR
difference trend with winds below 4 m/s even for orbits 15 minutes apart (
Figure
9
). Zooming
over low wind regimes shows that 3.7 m/s seems to be more precisely the limit and this marks
also the onset of small waves on sea surface
[
Donlon et al.
, 2002]
.

Below this limit (that we rounded
up to 4 m/s) AATSR and AVHRR in warm summer
Mediterranean conditions show discrepancies mainly due to

i)

a difference in the compared data effective resolutions. AVHRR data have been
indeed averaged over a 0.05 grid mesh, a “SEVIRI like” resolution, whereas
A
ATSR data have been remapped to the nearest neighbor. This may have a
smoothing effect for large amplitudes
[
Gentemann et al.
, 2008]

ii)

reduced sensitivity of
Metop
-
A

derived SST to true

SST variations, as apparent on
Figure
10
, showing the split window NL sensitivity deduced from partial
derivatives of simulated BTs as a function of surface SSTs
[
Merchant et al.
, 2009]
.
This is due to low winds coinciding with high water vapor loading as shown on
Figure
9
.


By night (
Figure
9
b), the agreement between AATSR and AVHRR is not

sensitive to wind and
the main wind impact is visible in the standard deviation values. Indeed
at nighttime, the
cooling rate around 21 UTC is at most 0.1 K per hour (
Figure
8
), so we used 60 minutes as
the nighttime time comparison threshold for the whole period.

It should be noted from
Figure
8

that SST at 21 LST are 0.3 K above the nighttime minimum
for low winds. This important information has to be taken in consideration when comparing
OSTIA based simulations and 21 LST
Metop
-
A

brightness temperature observ
ations for
adjustment purposes.
Figure
11

confirms the trend with wind of the guess field used in
operational simulations (OSTIA two days before) minus AATSR SST diffe
rence.


14



Figure
7
. Time difference impact on
Metop
-
A

A
/
AVHRR
-
AATSR

difference in summer time for
very low winds. Colors show the average water vapor content in each bin and its standard
deviation



Figure
8
. Typical summer SST cycle in the Mediterranean as a function of LST and wind
speed, averaged over the last 10 days in August, left full cycle, right zoom around midnight
LST
.



15



(a)


(b)

Figure
9
. Summer 2011
(a)
daytime and

(b)
nighttime difference between
Metop
-
A

and
AATSR in dependence of wind (x) and total column water vapor (z). Winds are binned by 0.1
m/s, and total column water vapor is represented by the color (g/cm2). Upper table show
biases, middle table show stand
ard deviations and lower table show the number of points.


16



Figure
10
.

SST daytime algorithm sensitivity to true SST as a function of wind speed.




Figure
11
: Nighttime: OSTIA

AATSR as a function of
wind speed.



17


3

Brightness temperature simulations

3.1

“Operational” simulations


CMS has been running a prototype chain since
November

20
11

for
testing the bias correction
method applied operationally to SEVIRI data
[
Le Borgne et al.
, 2011]

on
Metop
-
A
/AVHRR
data. The prototype processing is done in near
-
real time, to be as close as
possible to
operational conditions
.
Th
is choice has several consequences:



The prototype processing chain uses 3
-
hourly short lead
-
time atmospheric forecasts
(6, 9, 12 and 15 hour forecasts) from ECMWF, initialized from the analyses at 0000 h
and 1200 h UTC every day.



ECWMF fields are gridded t
o L3C grid (0.05 deg)



At each point of the 0.05° grid where a
Metop
-
A

SST value is present, the simulation
is performed using the corresponding
Metop
-
A
/AVHRR satellite zenith angle and the
closest in time ECMWF forecast, interpolated at grid point
location.



The Met Office operational analysis (OSTIA) has been used as surface temperature
field. To mimic operational conditions the OSTIA field used in the prototype is the
one dated two days earlier than the
Metop
-
A

data being processed.



The radiative t
ransfer computations, based on RTTOV version 10.1, use a limited
number of vertical levels from the model (15 levels at 1000, 950, 925, 900, 850, 800,
700, 600, 500, 400, 300, 250, 200, 150 and 100 hPa), due to limitations of the
operational near
-
real time

dissemination of ECMWF products to Météo
-
France.


3.2

“Delayed mode” simulations


To evaluate the impact of operational constraints, simulations have been also done in a
delayed mode environment with the following differences:

-

ECMWF analysis fields at 00, 06,

12 and 18 UTC were used instead of forecasts

-

simulations are calculated on ECMWF grid (0.1125 deg) and then interpolated to the
Metop
-
A

L3C working grid (0.05 deg)

-

fields are interpolated in time of the mean
Metop
-
A

overpass over Mediterranean for
each av
ailable orbit

-

OSTIA corresponding to the same day as the observations

-


RTTOV version 10.1 (same version) has been applied to 91 profile values (instead of
15 levels)

Simulation minus observation difference time series are displayed in
Figure
12

(operational
simulations) and
13

(delayed mode simulations). As already observed at the global scale
(
Figure
1
) simulations at 3.7 and 10.8

m show practically no biases and simulations
at 12

m are positively biased by about 0.2 K. Channel 12.0
µ
m simulations are more positively
biased for the delayed mode (~0.22 K) compared to operational simulations (~0.15 K).
Standard deviation values are smaller in all channels for delayed mode than
for operational
simulations. The best simulations (without corrections) across both types of simulations are
18


obtained for the 10.8

m channel (bias within 0.05 K and SD ~0.5 K), while delayed mode
give the best simulations for the 3.7

m channel simulation
s (
Table
2
).

Table
2
. Total statistics for the
Metop
-
A

simulations derived in delayed or operational mode
for 3 channels.


DELAYED

DAY

NIGHT


bs
±
std

N

bs
±
std

N

(Simbt


Obs )CH3

-

-

-
0.03
±
0.37

9579784

(Simbt


Obs )CH4

+0.03
±0.51

11855997

+0.04
±
0.48

9579784

(Simbt


Obs )CH5

+0.23
±0.55

11855997

+0.23
±0.55

9579784


DAY

NIGHT

OPERATIONAL

bs
±
std

N

bs
±
std

N

(Simbt


Obs )CH3

-

-

-
0.04
±0.43

10564572

(Simbt


Obs )CH4

-
0.03±0.59

13223490

-
0.00
±0.54

10564572

(Simbt


Obs )CH5

+0.15±0.62

13223490

+0.17
±0.60

10564572


Results shown by Figures
12

and
13

look quite stable with time and rather similar for the
operational and delayed mode. This apparent stability hides significant short scale time and
space variability.
Figure
14

shows 10 day averaged simulations


observations for channels
10.8 µm in October. Channel 10.8 µm simulations have a zero mean bias and October has
been chosen as a mean month. This

figure also shows that operational simulations may differ
from observations by ±1 K in coherent patches and that these errors are significantly
attenuated in delayed mode. Analyzing the simulation to observation daytime differences in
summer time with res
pect to wind speed (
Figure
15
) shows that there is a high dependence
with the wind speed with amplitudes between
-
0.5 K to 1 K for operational mode, and
-
0.4 K
to 0.4
K for the delayed mode simulations. At high wind speeds (above 10 m/s), simulation to
observation differences are much higher for operational compared to delayed mode
simulations. Positive trend for higher winds could be attributed to the different represe
ntation
of the atmosphere in dry conditions (associated with high winds) due to the reduced number
of vertical levels. At low wind speed (below 4 m/s), some wind dependent negative simulation
minus observation differences are expected, due to the fact that

simulations are derived with
OSTIA field not affected by diurnal warming.

In operational mode, guess SST field is OSTIA from two days ago, and in summer time this
means that simulations should be always lower compared to observations, due to the
combined
effect of seasonal increase trend and diurnal warming. The difference between
operational and delayed mode is a combined effect of different OSTIA and the different
vertical resolutions of the profiles. This is visible in the lower wind speed region, where

the
slope is higher in the operational simulations compared to delayed mode simulations.

Nighttime simulations in summer (
Figure
16
) show a positive trend of biases
and negative
trend of water vapor load with the increase of wind speed, which is more pronounced for
operational mode simulations then for delayed mode. The OSTIA timing problem and a still
evident residual diurnal warming effect in the late evening (see p
revious section), led to select
4 m/s as the nighttime threshold in deriving BT adjustment fields (see next section) as a
compromise between having enough number of available points and excluding residual
diurnal warming affected measurements . In winter t
ime (not shown), all this effects are not
visible, but for the sake of simplicity we maintain the same filtering throughout the whole
period.

19





(a)





(b)

Figure
12
.
Metop
-
A
/AVHRR operational mode simulations minus observations for channel
3.7

m (black curve), channel 10.8

m (red) and channel 12.0

m (blue) for (a) daytime

and
(b)
nighttime.

20





(a)


(b)

Figure
13
.
Metop
-
A
/AVHRR delayed mode simula
tions minus observa
tions for channel

3.7

m (black curve), channel 10.8

m (red) and channel 12.0

m (blue) for
(a)
daytime and
(b)
nighttime
.

21


.

(a)




(b)

Figure
14
.

10 days averaged simulations minus observation differences at 10.8

m for
(
a)
operational and
(
b) delayed mode simulations.



22



(a)


(b)

Figure
15
.
Metop
-
A
/AVHRR summer daytime simulation minus observations for channel 10.8

m fo
r
(
a) operational and
(
b) delayed mode.

23



(a)


(b)

Figure
16
.
Metop
-
A
/AVHRR summer nighttime simulation minus observations for channel
10.8 µm for
(
a) operational and
(
b) delayed mode




24


4

BT adjustments and SST correction

4.1

BT
simulation adjustment method


Comparison between BT simulations and observations showed differences due to the
combined effect of guess errors, NWP errors, filter errors, RTM errors or profile sampling
errors, as introduced in section 1. The bias correcti
on method we use is able to account for
differences between guess field and true SST, through equation (1), but not for the other
sources of errors. An adjustment of simulations is needed to solve this issue. Our simulated
BT adjustments are based on tempo
ral averaging and spatial smoothing of the simulation
minus observation differences. We assume a zero mean difference between the guess SST
(used in simulations) and the true SST (corresponding to observations), we assume also that
there is no systematic p
rofile or RTM error differences between the adjustment conditions,
and the conditions when we apply these adjustment. These assumptions are discussed below.
Finding optimal combination of space/time criteria is part of the research, therefore we tested
a r
ange of averaging sizes. For temporal averaging, we used averaging up to 30 days (1, 3, 5,
10, 15, 20, 30), where 1 day means no averaging. For spatial smoothing we used different
window sizes (
Table
3
) from 0.1 deg to 50 deg and over whole domain (Inf). In the prototype
a single temporal (10 days) and spatial smoothing (10 deg) was used, while in the study
phase, both delayed and operational mode simulations were analy
zed over the full range of
space/time parameters.

The most obvious problem induced by using simulations with no adjustment is the
overestimation of channel 12 micron temperatures by about 0.2 K. Since there is near
-
zero
bias for channel T
10.8
, this would r
educe T
10.8


T
12.0

simulated term in equations
(
2
) or
(
3
),
leading to underest
imation of the atmospheric correction term, an underestimation of the
simulated SST and hence to an overestimation of the simulation derived bias correction. We
nevertheless considered this option in the test set.

One assumption of this adjustment is the a
verage zero bias between guess SST and true SST.
For this reason we selected nighttime data. However
Figure
16

shows increase of simulated


observed BT difference wi
th wind speed. Part of the problem is residual DW effect at time of
Metop
-
A

nighttime orbits due to the still evident cooling of the sea surface in the summer
time, while using OSTIA that represents true foundation temperature. To try to avoid any
systematic biases due to residual cooling we discarded all data showing wind below
4 m/s in
the adjustment procedure.

Adjusting operational and delayed mode BT simulations we used slightly different procedures
in the prototype and the study phase:


1)

Sampling and concatenation


Prototype: Observed and simulated brightness temperatures are
sampled every 20 points and
lines (1° longitude/latitude) and concatenated over 10 days, together with latitude, longitude
satellite and solar angles, model water vapor, OSTIA and operational SST. Following filters
were applied in constructing differences:



|OSTIA
-
operational SST| <1K



|simBT
10.8
-
obsT
10.8
| < 1.5K



Confidence flags 3, 4 and 5



Solar zenith angle above 110
°
(only nighttime data)

25


Study phase: Simulations and observation fields are aggregated over T days without sampling
(full resolution 0.05
°
) and

with applied following filtering:



wind speed above 4 m/s



confidence flags 4 and 5



SDI <0.1



Solar zenith angle above 90
°

(to ensure enough data to perform analysis)



|OSTIA
-
SST operational| < 1 K



| simT
10.8
-
obsT
10.8
| < 1.5 K


2)

Satellite zenith angle normaliz
ation.

For a given channel, the simulation minus observation difference is expressed as a function of
the satellite zenith angle secant (operational mode processing) or just satellite zenith angle
(delayed mode processing) by a quadratic formula (
4
)

simB
i
-
o sB
i
=

i
*
+

i
*









(
4
)


Where
S

is either secant of satellite zenith angle or just satellite zenith angle.

The coefficients
in (
4
) are derived by regression on the concatenated file and
i

represent channel number.



s

i
=
*
+
*


+



contains the overall me
an difference and the influence of the
satellite zenith angle, considered over the whole domain. Also, on
Figure
17

we see the
dependence of simBT
4
-
obsBT
4

difference o
n satellite zenith angle for a chosen day before and
after the correction.



Satellite zenith a
n
gle (deg)

Figure
17
: Channel 10.8 µm s
imulation


observation as a function of the satellite zenith
angle for the 3
rd

of August 2011 b
efore and after correction.


Each simulation minus observation difference is normalized by adding a satellite zenith angle
correction to the raw difference as:


26


o
m
i
=
simB

i
-
o sB

i
-
s

i






(
5
)



3)

Building mean
difference fields


Sampled fields are averaged over time after the satellite zenith angle normalization step. In
clear, at a given location deltanorms
i

are averaged over T days (T=10 for prototype
processing). No limitation is imposed on the number of case
s actually contributing to the time
averaging. After time averaging, a smoothing is done over the averaged fields in SxS pixel
boxes, corresponding to S x 0.05° latitude longitude boxes (
Table
3
).

In case of operational processing, smoothing is done over sampled fields in 10x10 pixel boxes
corresponding to 10° latitude longitude boxes, while in case of analysis smoothing is done on
the parameters shown in
Table
3
. Full resolution fields are then reconstructed by expanding
the sampled field to full resolution (re location of the sampled values on the full resolution
grid), then smoo
thing in 20x20 pixel boxes (only for prototype processing). These fields will
be used as a “normalized adjustment fields” (deltanorm_mean). This procedure is equivalent
to choosing S=200.


4)

Applying the adjustment fields


Simulated temperatures are correcte
d by using the satellite zenith angle normalization
equation deltasec and the normalized adjustment local value.


o

simB

i
=
simB

i
-
s

i
-
o m

m

i




(
6
)


In
(
6
), simBT
i
-
deltasec
i

corresponds to the mean global + satellite global BT bias adjustment;
while deltanorm_mean
i

represents the local variation of the global adjustment. Comparison
between Figu
res
12
b,
13
b and
18

as well from the statistics represented in
Table
2

we see that
the adjustment procedure reduced biases

(mainly for channel 12.0 µm) and standard
deviations both for operational and delayed mode simulations. Similar to uncorrected
simulation analysis, delayed mode simulations have lower standard deviations compared to
operational simulations while the wind
dependence is preserved (not shown) after adjustment.


27




(a)



(b)

Figure
18
. Differences between adjusted simulations and observations BT for all channels,
where (a) represents operational simulations after adjustment in the pro
totype and (b) shows
28


delayed mode
simulations
. For the delayed mode simulations 20 days of time averaging and
15
°

spatial smoothing (the T20S301 configuration) were chosen as an example of correction.


Table
3
. Spatial smoothing par
ameters used in analysis. Inf denotes smoothing over the whole
domain (Mediterranean Sea) that produce constant adjustment value for each channel.

Space
(pixels)

1

3

5

9

15

21

31

51

71

101

201

301

351

401

501

601

999

Inf

Space
(deg)

0.05

0.15

0.25

0.45

0.75

1

1.5

2.5

~3.5

~5

~10

~15

~18

~20

~25

~30

~50

Inf


Table
4
. Total statistics for the
Metop
-
A

adjusted simulations derived in delayed and
operational mode in the same configuration for 3 channels.


DELAYED

(T20S301
configuration)

NIGHT


bs
±
std

N

(Simbt


Obs )CH3

+0.03
±0.36

5927782

(Simbt


Obs )CH4

+0.05
±0.47

5927782

(Simbt


Obs )CH5

+0.05
±0.53

5927782


NIGHT

prototype

bs
±
std

N

(Simbt


Obs )CH3

+0.03
±0.41

5264672

(Simbt


Obs )CH4

+0.02
±0.51

5264672

(Simbt


Obs )CH5

+0.03
±0.57

5264672


4.2

SST correction


After deriving nighttime BT adjustment fields, simulated BT fields are corrected by derived
BT adjustment for each combination of space
-
time parameters. We assume that the derived
nighttime BT adjustme
nt could be applied also on daytime simulations. After that, applying
Metop
-
A

SST algorithm to adjusted BT gives a “simulated” SST, and SST algorithm error is
defined as a difference between simulated adjusted SST and the guess SST used in simulation
step
(in our case OSTIA). Corrected SST is then obtained by applying
Metop
-
A

SST
algorithm to BT observations and
subtracting

the derived SST algorithm error for each space
-
time combination.

The main assumption that nighttime BT adjustment is applicable in a
daytime condition is
questionable in conditions of intensive summer DW events because:

1
-

Filtering out low wind conditions concerns large area in the Mediterranean (we see that on
Figure
9
b
-



third subplot) and reduces the representativeness of the nighttime BT adjustment

2
-

DW may induce specific atmospheric profiles, and thus specific DW induced profile errors,
which are not accounted for by nighttime adjustment.

29


M
aybe the optimum solution will be to use different BT adjustments for daytime and
nighttime, but this would require daytime SST analysis be available as surface temperature
field for simulations.


5

Results

Adjusted simulations are used in a bias correction
method as described in the previous
section. Each adjustment time and space combination produces a set of simulations that are
used to derive correction terms to the CMS operational
Metop
-
A

AVHRR algorithm. The
efficiency of the adjustment is measured by t
he validation results of the corrected SST
compared to coincident AATSR data. The comparison conditions are identical to those of the
validation experiment described in section
2
.

The tested solutions are referred to by the corresponding T (time) and S (space) values
(defined in previous chapters) where:

-

T00S00: corresponds to no BT adjustment;

-

T01S0.05 to a pixel to pixel BTs adjustment where the corrected SST is
equal to the
SST field used for deriving simulations (OSTIA)

-

SInf corresponds to using one unique value, derived by averaging over T days and
smoothing over the whole domain

The result analysis will be based on two types of figures:


1)

time series showing th
e bias and standard deviation of the operational CMS results
(with or without bias correction) and the results obtained for typical combinations of
simulation adjustment: T (1, 3 and 30 days) and S (0.05, ~15 and ~50 deg).
Figures 19
and 20
represent time
series for the operational and delayed mode simulations,
respectively.
The best compromise solution (
3

days;
15

deg
, see section
5.5

below) is
also shown on these figures.

2)

synthetic figures presenting the three key parameters, bias, bias stability and standard
deviation as a function of all time and space smoothing combina
tions. Operational and
delayed mode simulation results are shown on Figures
21

and
22

for the whole period;
and Figures
24

and
25

for

the winter season only.

These three parameters give almost independent information so they are analyzed separately.
Time series and tables are produced separately for day and night and for operational and
delayed mode simulation, and due to the specific
conditions in summer time in Mediterranean
(intensive DW episodes), we will analyze separately whole time period (with excluded data
with wind below 4 m/s) and only colder season (from 11/2011 to 03/2012) with excluded data
with wind below 2 m/s.



5.1

Operati
onal daytime
-

whole period


The most striking feature in
Figure
19

is the difference between summer and the cold season.
In summer, the correction is overestimated in most cases and leads to negative corrected SST
biases, whereas during the cold season the correction works as expected. The correction is
30


deduced (with oppo
site sign) from the difference between the simulated and the guess SST. If
simulated SST is larger than it should, then the correction is too (negatively) large. A simple
cause of overestimating simulated SST is to overestimate the simulated brightness
tem
peratures. Residual nighttime diurnal warming affected AVHRR observations are larger
than the foundation (OSTIA) SST. The resulting BT adjustment is too large and leads to
overestimated simulations then to overestimated corrections. Different space/time re
alizations
produce prominently different results in summer season, while in colder season all
realizations produce similar results.

Figure
19

introduces the main con
tradiction between bias (
Figure
19
a) and standard deviation
results (
Figure
19
b
). Small time averaging (1 day, blue lines) produces low bias and high
standard deviation results, whereas the opposite is observed for large time averaging.
Similarly small space smoothing leads to large standard deviation. The optimum combination
will

necessarily be a compromise.

If we analyze now to the detail results in
Figure
21
, the best result when using bias as the
selection criterion (
Figure
21
a


1
st

panel), is obtained without time averaging (T
01
) and with
lower spatial smoothing (below
1.5

degree), where the overall bias is
-
0.046

K. The total
range of obtained biases is between
-
0.15

K and
-
0.05K, while the operational not corrected
bias is 0.25 K.

The increase of bias is time averaging dependent with the highest absolute
values for the highest time averaging, but nevertheless all obtained values show improvement
over the not corrected results.

Using stability of the bias (
Figure
21
a


2
nd

panel) as criterion favors lower temporal
averaging (3 days) and again with lower spatial smoothing (below 1 deg). In this case the
minimum values are same as t
he value obtained with operational algorithm without
correction. Other values are increasing about diagonally towards the highest time averaging
and spatial smoothing window with the maximum values about 0.23 K.

Finally, using overall standard deviation (
Figure
21
a


3
rd

panel) as a criterion shows a spatial
smoothing dependence with highest values corresponding to the lowest smoothing and almost
no tim
e averaging dependency. Excluding values of the low smoothing (below ~3.5 deg)
obtained values fall within the small range from 0.29 to 0.31 K, with improvement compared
to standard deviation obtained without correction (0.35 K).

Small time averaging windo
w (1
-
3 days) with higher spatial smoothing (above 5 deg) window
is the best compromise suggested by this study.


31


5.2

Operational nighttime


whole period

During nighttime, operational bias is already close to zero (
-
0.024 K) and it could be hard to
improve res
ult if we rely only on bias as selection criterion. Nevertheless a bias minimum is
obtained when averaging over 3 to 10 days and for higher spatial smoothing (above 10 deg),
where the obtained values are below
-
0.05 K (minimum is
-
0.036 K).

The time series

(
Figure
19
b), show a seasonality in the operational bias signal producing
positive biases in summer time, and negative in wintertime, giving overall bias to near
-
zero
.
Bias stability value is an interesting criterion in these conditions (
Figure
21
b


2
nd

panel), and
the largest improvement is obtained for 15 to 20 days averaging and highest spatial smoothing
(50 deg) with a bias stability value of 0.06 K.

Considering standard deviation (
Figure
21
b


3
rd

panel), the optimum is obtained in the upper
right corner of the table (largest spatial and temporal averaging). As in daytime however, the
variability is small (0.26±0.01 K) for all windows above 17.5
deg.

Comparing day and night results, an overlapping between day and night optimal conditions is
found between 1 and 5 day averaging, and between 10 and 15 degree smoothing, based on
bias and bias stability results. For day and night, optimizing standard d
eviation would lead to
select the largest spatial scales with contradictory options between day (low time averaging)
and night (large time averaging). Altogether a good compromise in nighttime conditions
would be found for time averaging larger than 3 days

and smoothing window above 15 deg.


5.3

Operational to delayed mode


whole period

Operational (
Figure
19
) and delayed (
Figure
20
) mode time series show high similarities, with
over correction of the positive error in the summer daytime (both for the daytime and
nighttime) and error reduction in winter time. No adjustment results (T0
0S00) produce the
best results in summertime, revealing problematic BT simulation adjustment conditions for
this season. In wintertime, the no adjustment options leads to large positive errors, particularly
at night.

It should be stressed again, that in th
e case of operational mode, AATSR data coincident with
Metop
-
A

data are more independent from the guess SST field used in simulations (OSTIA
two days ago), where in the case of delayed mode, the same AATSR data used in validation
was also used in deriving
guess field (OSTIA). Since both types of simulations produce
similar results, this proves that method is robust even when using not coincident guess SST
fields.

Daytime overall bias comparison (
Figure
22
a
-

1
st

panel)

show similar patterns as in the
operational mode results (favoring small time averaging), but the absolute values of biases in
delayed mode were higher (
-
0.093 K) compared to operational mode (
-
0.049 K), due to the
period of very negative bias in Septembe
r and October not included in operational processing
for technical reasons. Bias stability (
Figure
22
a


2
nd

panel) shows the largest differences
between the tw
o modes, favoring higher time (
20

days
) averaging in delay
ed
mode against
lower averaging (
3

days
) in operational mode. Both bias stability minima have values similar
to the results obtained without correction (~0.18 K). Overall standard deviation (
Figure
22
a


3
rd

panel) show consistently the minimum values for high spatial smoothing (above 5 deg)
and small time averaging with similar absolute values between the two modes.

For nighttime, bias result (
Figure
22
b
-

1
st

panel) are

quite similar to the operational ones,
favoring similar time averaging criteria’s (3
-
30 days) and high spatial smoothing (between 10
and 50 deg), with very similar absolute value of biases (
-
0.049 K for delayed mode and
-
0.036
K for operational mode). The

main difference is found for bias stability, where the delayed
32


mode analyses give more stable values throughout the whole period (
Figure
22
a
-

2
nd

panel),
with improvement of almost 50% (from 0.060 to 0.034 K) compare
d to the operational mode.
In the delayed mode, optimum values are found for large time averaging (30 days) compared
to one day in the operational case. The most favorable standard deviation (
Figure
22
b
-

3
rd

panel) co
ndition i
s obtained with no averaging (
1

day
) and smoothing over the medium
window sizes (6


10 deg). Again, all smoothing above 2.5 deg give overall SD values in the
very small range (between 0.248 and 0.266 K).

Delayed mode does not lead to conclusions
significantly different to those of the operational
mode providing the most contradictory time averaging solutions (1 day and 30 days) are
avoided.


5.4

Colder period


operational and delayed mode


Filtering out low wind conditions however is not totally sati
sfying, because in the
Mediterranean Sea, this favors dry atmospheres and may biases results. To check the stability
of our previous conclusions, we have analyzed overall performances only for the colder
period.

For that purpose we created aggregated table
s based on the five months in the colder season
(Figures
24

and
25

in Appendix B
), where we also decreased the wind filtering criteria to 2
m/s, mainly to compensate for the more pronounced cloud mask, since diurnal warming is
much less frequent and of smaller intensity. Results suggest that although the overall bias and

standard deviation has been expectedly improved, the patterns remained however very similar
to the patterns obtained for the total period.

For daytime conditions, both operational and delayed mode analysis give near zero biases
with similar patterns as fo
r the whole period, favoring lower time averaging (1 or 3 days) and
medium to high smoothing window size (10
-
50 deg). It should be noted that the overall
amplitude in the winter time months is also reduced, ranging between 0 and 0.05 K. Bias
stabilities ar
e also expectedly reduced in the winter daytime (from 0.20 K to 0.14 K for
operational and 0.10 K for delayed mode) with the shift to the higher smoothing windows (15
to 18 deg) and smaller time averaging
(1 to
3

days
). Again, total range of values is with
in
0.02

K. Winter time standard deviations are optimal when used without time averaging (T01)
and for the medium smoothing window size (from 10 deg to 15 deg), whereas for the whole
period higher window sizes was preferred in both cases (25 deg for operati
onal and 30 deg for
delayed mode).



5.5

Discussion:


Analyzing both periods (whole and colder season) and both types of simulations (operational
and delayed) leads to the following remarks:


1) The most consistent results throughout the day and night and two simulations types are
obtained for the standard deviation results, where in all cases high spatial smoothing, above
10 deg is always favorable with very small variability (within 1
-
2 hundr
edths of Kelvin)
between different time averaging realizations. The main reason for this is by smoothing over
33


larger area we decrease the noise introduced in simulations through errors in guess field,
atmospheric profiles and RT model.

2) Bias results are
consistent across the different simulation modes, but diverge between day
and night. For the daytime, results are dependent by the time window averaging, improving
from high to low temporal window and the minimum values are obtained between 1 and 5
days av
eraging. During nighttime, results are essentially depending on the spatial smoothing
window size, where the optimum results are obtained for smoothing above 10 deg and higher
time averaging (3
-
30 days). This difference between day and night suggests that
daytime
biases are affected by more transient features (mainly diurnal warming) while nighttime
results are affected by long range weather patterns that change the atmospheric structure.
Again, the difference between day and night is also driven by the dif
ference in the SST
algorithm, where during the daytime split window algorithm is used and during the nighttime
more accurate triple window algorithm is used.

3) Bias stability results show highest variability between both simulations types and between
day
and night, but also smallest variability within each analysis. During daytime, the applied
correction basically shifts constantly positive results (0.25 K) to near zero values, and
therefore bias stability values are similar before and after correction. St
ill, the minimum
values for both types of simulations and both periods are obtained for 3 days averaging and
lower spatial smoothing (below 15 deg). During nighttime, all corrections give improvement
while the overall variability (especially for the delaye
d mode simulations) is lower compared
to the daytime results. Still, the optimum parameters for operational mode simulations favor
both higher temporal averaging (3
-
20 days) and higher spatial smoothing (above 15 deg).

4) Considering all different conditio
ns we suggest using temporal averaging of 3 to 5 days and
spatial smoothing about 15 deg.

For final test purposes, we marked (colored line with black dot) the obtained optimum
com
bination of parameters (3 days; 15 deg
) in time
-
series (
Figures 19 and 20
)
and we see
balanced improvement between bias and standard deviation values, not obtained for other
chosen parameters where for certain period we got either improvement in bias or in standard
deviation.




34





(a)

35






(b)

Figure
19
: Time series of
(
a) daytime and
(
b) nighttime operational simulation correction
results. Each subfigure has three subpanels: time series of 1) biases; 2) standard deviations
and 3) number of points. Colored line with black dots represents the results ob
ta
ined with
optimum parameters (
3

days; 15 deg
).


36








(a)

37





(b)

Figure
20
. Time series of
(
a) daytime and
(
b) nighttime delayed mode simulation correction
results.

Each subfigure has three subpanels: time series of 1) biases; 2) standard deviations
and 3) number of points. Colored line with black dots represents the results obta
ined with
optimum parameters (
3

days;15 deg
).



38



(a)



(b)


Figure
21
: Synthetic operational simulation correction results a) by day; b) by night. Top
panel biases; second panel bias stability; third panel: standard deviation. Title of the each
sub
-
table gives values for the operational algorithm
(CALC), operational correction
(OP_CORR), no adjustment correction (CORR_T0S0), the best possible results (BST)
obtained by averaging the minimum values from each 10 days averaged field and the best
combination of space and time parameters for the defined
period (MIN).

39



(a)


Figure
22
. Synthetic delayed mode results
(
a) by day;
(
b) by night. Top panel biases; second
panel bias stability; third panel: standard deviation.



40


6

Conclusions and future work

We assessed the impact of BT simulation adjustment conditions in correcting
Metop
-
A

SST
over the Mediterranean Sea. We used two different BT simulations: full vertical (91) ECMWF
NWP profiles combined with concurrent OSTIA SST field (delayed mode) and redu
ced
vertical (15) ECWMF NWP profiles combined with two days delayed OSTIA field
(operational). For both types of simulations, separately for day and night we performed BT
adjustments and SST corrections over different temporal and spatial averaging window
sizes.
To validate the method we used AATSR data and three statistical parameters derived from 10
day averaged differences between
Metop
-
A

and AATSR: bias, bias stability and standard
deviation. We took special care in analyzing daytime condition in summer

time during
intensive DW episodes that tend to mask out any atmospheric problems we aimed to correct
for. The final choice results from compromising between different conclusions according to
each of these parameters are used as selection criterion.

Based

on all different conditions we found optimum temporal averaging between 3 to 5 days
and spatial smoothing of 15 deg. Specifically, based on the standard deviations cost function
we consistently obtained that smoothing above 10 deg is the most favorable an
d almost
regardless of temporal averaging. Bias based results diverge between day and night, where
daytime analysis are mostly time averaging dependent and favors short temporal averaging (1
to 5 days) while nighttime results are spatial smoothing dependen
t and favors higher
smoothing (above 10 deg) for any temporal averaging (except no averaging).

The main improvement of using full resolution atmospheric profiles compared to reduced one
is evident in lower standard deviation of the simulation to observatio
n differences and for
nighttime

bias stability

results where almost all space/time realization produces values lower

(between 0.034 and 0.062 K
) than results derived without any correction or with operational
simulations

(between 0.060 and 0.116 K)
.

Conver
ging conclusions derived from operational and for delayed mode simulations,
produced in totally independent ways, is an indication of the consistency of the optimal
solution retained
.

Availability of the guess SST field with even the higher spatial resolut
ion, with possibility to
resolve finer scale mesoscale features, could probably decrease the spatial smoothing window.

Also, high influence of DW in summer daytime conditions in Mediterranean Sea decorrelates
nighttime from daytime, leading ideally to more

appropriate separate BT adjustment for day
and nighttime. In practice this is still not feasible at present since it would require diurnal
warming resolving daytime SST analysis.

This methodology together with obtained optimized spatio/temporal parameters

is foreseen to
be applied at CMS globally for
Metop
-
A

A/AVHRR but also for SUOMI
-
NPP/VIIRS and
Metop
-
A

B/AVHRR SST algorithms.



41


Acknowledgements

This work was realized in the context of EUMETSAT Visiting Scientist program
(OSI_AS12_02) and of the BESST (
Inter
-
sensor Bias Estimation in Sea Surface Temperature)
-

SR/12/158

project funded by the Belgian Science Policy (BELSPO) in the frame of the
Research Program For Earth Observation “STEREO III”.
The data from the EUMETSAT
Satellite Application Facility on Ocean and Sea Ice used in this study are accessible through
the SAF’s homepage:
http://www.osi
-
saf.org
. ENVISAT/AATSR data are provided by the
European Spac
e Agency (ESA) via the project ID 5802. ECMWF data are obtained through
the MARS archive.




42


Appendix

A

Table
5
. Total statistics for the validation of
Metop
-
A
-
A SST compared to OSTIA and AATSR
both for day and night for wind speed
> 6 m/s during day and > 2 m/s during night



DAY

NIGHT



bs
±
std

N

bs
±
std

N

OSTIA
-
AATSR

-
0.13
±0.56

158785

-
0.08
±0.39

1219090

Metop
-
A
-
OSTIA

+0.26
±0.50

3171141

+0.03
±0.39

8669439

Metop
-
A



AATSR

+0.31
±0.37

90818

-
0.01
±0.28

1044023





43





Figure
23
. Time series of
Metop
-
A
, AATSR and OSTIA differences fo
r (a) daytime and (b)
nighttime for data with the wind speed above 6 m/s during daytime and above 2 m/s during
nighttime.



44


Appendix

B


(a)


(b)

Figure
24
. Winter operational mode (a) day and

(b) night.

45



(a)


(b)

Figure
25
. Winter delayed mode (a)
day and (b)
night
.




46


References

Le Borgne, P., H. Roquet, and C. J. Merchant (2011), Estimation of Sea Surface Temperature
from the Spinning Enhanced Visible and Infrared Imager, improved using numerical
weather prediction,
Remote Sensing of Environment
,
115
(1), 55

65,
doi:10.1016/j.rse.
2010.08.004.

Buongiorno Nardelli, B., C. Tronconi, and R. Santoleri (2011), High and Ultra
-
High
resolution processing of satellite Sea Surface Temperature data over the Southern
European Seas in the framework of MyOcean project, ,
submitted to Remote Sensi
ng
of Environment
.

Corlett, G. K. et al.
(2006), The accuracy of SST retrievals from AATSR: An initial
assessment through geophysical validation against in situ radiometers, buoys and other
SST data sets,
Advances in Space Research
,
37
(4), 764

769,
doi:10.
1016/j.asr.2005.09.037.

Donlon, C. et al. (2007), The Global Ocean Data Assimilation Experiment High
-
Resolution
Sea Surface Temperature Pilot Project,
Bulletin of the American Meteorological
Society
,
88
(8), 1197

1213, doi:10.1175/BAMS
-
88
-
8
-
1197.

Donlon, C.

J., M. Martin, J. Stark, J. Roberts
-
Jones, E. Fiedler, and W. Wimmer (2012), The
Operational Sea Surface Temperature and Sea Ice Analysis (OSTIA) system,
Remote
Sensing of Environment
,
116
(0), 140

158, doi:10.1016/j.rse.2010.10.017.

Donlon, C. J., P. J. M
innett, C. Gentemann, T. J. Nightingale, I. J. Barton, B. Ward, and M. J.
Murray (2002), Toward improved validation of satellite sea surface skin temperature
measurements for climate research,
Journal of Climate
,
15
(4), 353

369.

EUMETSAT (2010),
Low Earth
Orbiter Sea Surface Temperature Product User Manual
,
SAF/OSI/CDOP/M
-
F/TEC/MA/127. [online] Available from: http://www.osi
-
saf.org

EUMETSAT (2011),
Geostationary Sea Surface Temperature Product User Manual
,
SAF/OSI/CDOP/M
-
F/TEC/MA/181. [online] Available
from: http://www.osi
-
saf.org

Francois, C., A. Brisson, P. Le Borgne, and A. Marsouin (2002), Definition of a radiosounding
database for sea surface brightness temperature simulations Applications to sea surface
temperature retrieval algorithm determination
,

Gentemann, C. L., P. J. Minnett, P. Le Borgne, and C. J. Merchant (2008), Multi
-
satellite
measurements of large diurnal warming events,
Geophysical Research Letters
,
35
(22),
L22602, doi:10.1029/2008GL035730.

Hocking, J., P. Rayer, R. Saunders, M. Matrica
rdi, A. Geer, and P. Brunel (2011),
RTTOV v10
Users Guide
, EUMETSAT/NWP SAF, Lannion, France. [online] Available from:
http://research.metoffice.gov.uk/research/interproj/nwpsaf/rtm/docs_rttov10/

47


Liang, X., and A. Ignatov (2011), Monitoring of IR Clear
-
Sky

Radiances over Oceans for SST
(MICROS),
J. Atmos. Oceanic Technol.
,
28
(10), 1228

1242, doi:10.1175/JTECH
-
D
-
10
-
05023.1.

Liang, X.
-
M., A. Ignatov, and Y. Kihai (2009), Implementation of the Community Radiative
Transfer Model in Advanced Clear
-
Sky Processor
for Oceans and validation against
nighttime AVHRR radiances,
J. Geophys. Res.
,
114
(D6), D06112,
doi:10.1029/2008JD010960.

Merchant, C. J., P. Le Borgne, A. Marsouin, and H. Roquet (2008), Optimal estimation of sea
surface temperature from split
-
window obse
rvations,
Remote Sensing of Environment
,
112
(5), 2469

2484, doi:10.1016/j.rse.2007.11.011.

Merchant, C. J., P. Le Borgne, H. Roquet, and A. Marsouin (2009a), Sea surface temperature
from a geostationary satellite by optimal estimation,
Remote Sensing of En
vironment
,
113
(2), 445

457, doi:10.1016/j.rse.2008.10.012.

Merchant, C. J., O. Embury, P. Le Borgne, and B. Bellec (2006), Saharan dust in nighttime
thermal imagery: Detection and reduction of related biases in retrieved sea surface
temperature,
Remote Sen
sing of Environment
,
104
(1), 15

30,
doi:10.1016/j.rse.2006.03.007.

Merchant, C. J., and A. R. Harris (1999), Toward the elimination of bias in satellite retrievals
of sea surface temperature 2. Comparison with in situ measurements,
J. Geophys. Res.
,
104
(C10), 23579

23590, doi:10.1029/1999JC900106.

Merchant, C. J., A. R. Harris, F. Maturi, O. Embury, S. N. MacCullum, J. Mittaz, and C. P.
Old (2009b), Sea surface temperature estimation from the Geostationary Operational
Environmental Satellite
-
12 (GOES
-
12)
,
Journal of Atmospheric and Oceanic
Technology
,
26
(3), 570

581.

Merchant, C. J., A. R. Harris, H. Roquet, and P. Le Borgne (2009c), Retrieval characteristics
of non
-
linear sea surface temperature from the Advanced Very High Resolution
Radiometer,
Geophysi
cal Research Letters
,
36
(17), L17604,
doi:10.1029/2009GL039843.

O’Carroll, A. G., T. August, P. Le Borgne, and A. Marsouin (2012), The accuracy of SST
retrievals from Metop
-
A IASI and AVHRR using the EUMETSAT OSI
-
SAF matchup
dataset,
Remote Sensing of Envi
ronment
,
126
(0), 184

194,
doi:10.1016/j.rse.2012.08.006.

O’Carroll, A. G., J. R. Eyre, and R. W. Saunders (2008), Three
-
Way Error Analysis between
AATSR, AMSR
-
E, and In Situ Sea Surface Temperature Observations,
J. Atmos.
Oceanic Technol.
,
25
(7), 1197

1207
, doi:10.1175/2007JTECHO542.1.

O’Carroll, A. G., J. G. Watts, L. A. Horrocks, R. W. Saunders, and N. A. Rayner (2006),
Validation of the AATSR meteo product sea surface temperature,
Journal of
Atmospheric and Oceanic Technology
,
23
(5), 711

726.

48


Petrenko, B
., A. Ignatov, N. Shabanov, and Y. Kihai (2011), Development and evaluation of
SST algorithms for GOES
-
R ABI using MSG SEVIRI as a proxy,
Remote Sensing of
Environment
,
115
(12), 3647

3658, doi:10.1016/j.rse.2011.09.003.

Saha, K., A. Ignatov, X. M. Liang, a
nd P. Dash (2012), Selecting a first
-
guess sea surface
temperature field as input to forward radiative transfer models,
J. Geophys. Res.
,
117
(C12), C12001, doi:10.1029/2012JC008384.

Tomažić, I. (2012),
Using Fine scale NWP model outputs to improve the retr
ieval of SST from
Metop
-
A

AVHRR brightness temperatures. Part 1: Initial results based on SEVIRI BT
,
Eumetsat Visiting Scientist Report, No. OSI_AS11_P03.

Tomažić, I., and M. Kuzmić (2009), ERA
-
40
-
aided assessment of the atmospheric influence
on satellite
retrieval of Adriatic Sea surface temperature,
Meteorol Atmos Phys
,
104
(1
-
2), 37

51, doi:10.1007/s00703
-
008
-
0015
-
2.

Tomažić, I., and M. Kuzmić (2011), Improving regional AVHRR SST measurements using
AATSR SST data, in
Proceedings of the 2011 EUMETSAT Meteo
rological Satellite
Conference
, vol. P.59, Oslo, Norway.

Tomažić, I., M. Kuzmić, G. Notarstefano, E. Mauri, and P. M. Poulain (2011), A comparative
assessment of satellite
-
derived Adriatic Sea surface temperature,
International Journal
of Remote Sensing
,
3
2
(17), 4871

4892, doi:doi: 10.1080/01431161.2010.492249.

Wimmer, W., I. S. Robinson, and C. J. Donlon (2012), Long
-
term validation of AATSR SST
data products using shipborne radiometry in the Bay of Biscay and English Channel,
Remote Sensing of Environment
,
116
(0), 17

31, doi:10.1016/j.rse.2011.03.022.

Zavody, A. M., C. T. Mutlow, and D. T. Llewellyn
-
Jones (1995), A radiative transfer model
for sea surface temperature retrieval for the along
-
track scanning radiometer,
J.
Geophys.
Res.
,
100
(C1), 937

952.