Q. J. R. Meteorol. Soc. (2006), 132, pp. 1927–1947 doi: 10.1256/qj.
ADM-Aeolus Doppler Wind Lidar
Observing System Simulation Experiment
By A. STOFFELEN1∗, G.J. MARSEILLE1, F. BOUTTIER2,
D. VASILJEVIC3 , S. DE HAAN1 and C. CARDINALI3
1KNMI, De Bilt, The Netherlands
2Me´te´o France, Toulouse, France
3ECMWF, Reading, UK
(Received 13 May 2005; revised 31 March 2006)
SUMMARY
Within the Atmospheric Dynamics Mission Aeolus (ADM-Aeolus) the European Space Agency (ESA) has
approved a Doppler Wind Lidar (DWL) to fly on a dedicated platform orbiting dawn-dusk at 400 km altitude,
planned for launch in 2008. Rigorous design trade-offs have resulted in a lidar concept capable of delivering
high-quality wind component profiles, but with a limited coverage. A companion paper describes the realistic
simulation of this DWL, whereas this paper sets out to assess the impact of such lidar in meteorological analyses
and forecasts. To this end an Observing System Simulation Experiment (OSSE) is run. The superior conventional
observation coverage of 1993 is used to simulate all conventional observations, whereas, on the other hand, a
limited set of satellite observations is simulated. As a consequence, only the northern hemisphere DWL impact in
the OSSE is assumed realistic. Here, in a 15-day period with variable weather, out of 15 daily forecasts, 14 show
beneficial impact of the DWL. Although the experiment is limited, it corroborates other practical and theoretical
evidence that the ADM DWL will demonstrate a beneficial impact in meteorological analyses and forecasts.
KEYWORDS: Atmospheric Dynamics Mission - Aeolus OSSE Data assimilation
1. INTRODUCTION
The quality of state-of-the-art Numerical Weather Prediction (NWP) is among
other factors determined by the availability and quality of meteorological observations.
However, conventional wind profile data lack coverage and a uniform distribution
over the globe. On the other hand, NWP models have improved much over the last
decades, and advanced 4D-var techniques are now being used for the analysis. The
spatial resolution of global circulation models has improved as well, which leads to
a need for more observations to initialise the sub-synoptic scales. On these scales the
wind field, rather than the atmospheric temperature field determines the atmospheric
dynamics. Furthermore, a prime factor determining meteorological instability is vertical
wind shear.
For the study of climate processes extensive re-analysis experiments are being
conducted. These experiments use the technique of data assimilation, as used for NWP,
to establish long time series of the weather in support of climate studies. However, 3D
wind information has been lacking in the tropics for an accurate definition of the Hadley
circulation.
To fill in the gap in the global observing system, ESA has selected ADM-Aeolus
as an Earth Explorer core mission to provide wind profile observations globally. This is
achieved by flying a Doppler wind lidar on a free-flyer platform in a dawn-dusk polar
orbit and measuring in the ultra-violet (UV) part of the electromagnetic spectrum at
355 nm. The instrument is non-scanning with a fixed scan angle perpendicular to the
direction of satellite propagation. Profiles of wind components at about 1 km vertical
resolution, ranging from the earth surface up to 26 km, are retrieved from received
light backscattered from clouds, aerosol and molecules. The ADM wind requirements
have been focussing on quality (error reduction) rather than quantity (coverage), in
∗ KNMI, Postbus 201, 3730 AE De Bilt, The Netherlands, e-mail: Ad.Stoffelen@knmi.nl
c© Royal Meteorological Society, 2006.
1927

1928 A. STOFFELEN et al.
accordance with the WMO requirements. Moreover, past experience in data assimilation
shows that quality can usually not be traded off against quantity without a degrading
effect (e.g., Butterworth and Ingleby 2000; Rohn et al. 2000). To yield wind observation
quality comparable to radiosondes the instrument is operated at 100 Hz during 7
seconds so that measured wind profiles are representative for a 50 km atmospheric track.
Moreover the instrument is operated in burst mode with a 25% duty cycle, meaning that
a wind profile is observed every 200 kilometres (see also ESA 1996; 1999; Stoffelen et
al 2005).
The ADM requirements focus on the spatial representativeness and accuracy of the
wind profiles obtained, rather than on the number of wind profiles. Since good-quality
conventional wind profiles are known to have large analysis impact this choice is based
on practical experience. The potential detrimental effects of poor quality observations
are also well known from Observing System Experiments (OSE). To achieve spatially
representative and accurate observations, the 50-km-size wind profile cells are sampled
by multiple shots. If these shots were spread over a larger domain one would get (i)
fewer shots in a cell and therefore a lower number of photons returned, resulting in a
poorer assessment of the wind conditions in the cell and (ii) a poorer sampling of the
subcell wind variability and therefore an increased representativeness error. Both work
in the same direction and favour accuracy rather than coverage as a wind profile mission
driver. It is this choice that makes the ADM a feasible space-borne DWL demonstration
mission. Furthermore, multiple shots in a cell allow the use of signal and wind variability
measures for quality control purposes.
OSEs by the European Centre for Medium-range Weather Forecasts (ECMWF)
(Kelly 1996) have confirmed the value of tropospheric wind profile data for NWP.
ECMWF tested this in a series of experiments where they excluded conventional
wind profile observations (TEMP/PILOT), or parts thereof in the free troposphere,
and compare to experiments where conventional (TEMP/PILOT), or satellite (TOVS)
temperature or humidity profile data, or single level observations, were excluded. In
more recent experiments (Kelly 2004) satellite soundings play an increasingly important
role and show clear positive impact in the northern hemisphere (NH). There are two
probable causes for this: (i) satellite soundings have improved and are better exploited
in more recent data assimilation systems, as may be inferred from the rapidly improving
forecast skill in the southern hemisphere (SH); (ii) the conventional wind sounding
network has been decaying in important parts of the NH, i.e., from 1993 to 1999 the
number of TEMP/PILOT wind profiles has halved.
Complementary experimentation has been performed at the Deutscher Wetterdi-
enst (DWD) to test the impact of continental North American wind profile observations
(Cress 2001). From these experiments, a few points are noteworthy: (i) these exper-
iments confirm the importance of wind profile data, compared to the importance of
temperature/humidity data (Baker et al. 1995; ESA 1996; 1999); (ii) near-surface wind
observations (Planetary Boundary Layer (PBL) winds) seem less important than winds
in the middle and upper troposphere; (iii) in the OSE experiments, a small number of
(good quality) wind profiles already show a positive impact on the quality of NWP.
The results and conclusions of OSEs give an insight into the effect of a particular
existing observation type in an existing data assimilation system. However, it is difficult
to draw conclusions from this on the added value of supplementary measurements for
future meteorological analyses and forecasts. Such added value may be investigated
through OSSEs. Me´te´o France has made a first step in assessing the value of ADM.
The work involved running OSSE experiments with the French Arpege NWP model, in
order to test the impact of the OSSE database DWL wind profiles from a 10 micron laser

DOPPLER WIND LIDAR OSSE 1929
on a free-flyer satellite in a polar orbit (Cardinali et al. 1998). This scenario provided
a wind profile density over the oceans comparable to the current conventional wind
profile density over land in the NH. The assimilation experiments were performed with
a low-resolution version of the NWP model (T42 spectral resolution, i.e. ∼500 km
horizontal resolution), and the DWL impact could be well demonstrated, even though
the subsynoptic scales where wind observations become most relevant are not well
resolved at this resolution.
DWL OSSEs performed in the United States indicate an impact even for low
measurement accuracy (Atlas et al. 2003). However, the forecast quality was almost
exclusively based on DWL information from the SH and therefore was bound to show
an improvement against the control analysis which did not contain relevant observations
in this area. More recent OSSE work (so called bracketing OSSEs) with the US National
Center for Environmental Prediction (NCEP) NWP model aims to explore the bounds
of the potential impact of DWL by considering various DWL concepts each focussing
on particular atmospheric regions and based on scanning and non-scanning (ADM-type)
instruments. Significant DWL impact has been demonstrated, e.g. larger than TOVS in
the tropics and SH for all considered DWL scenarios (Masutani 2002; 2004).
For an operational system, the impact on NWP often depends on the capability
of the data assimilation system used. Therefore it is worthwhile to perform an OSSE
with the state-of-the-art ECMWF 4D-var system (Tan and Andersson 2004) in order to
consolidate the requirements for an operational mission. Section 2 discusses the general
OSSE setup and required attributes. Section 3 discusses results of the OSSEs performed
to demonstrate the impact of ADM on atmospheric circulation analyses and NWP. For a
correct interpretation of the results and to verify the realism of the OSSE method results
were validated against the 1999 operational model in section 4. Section 5 provides a
summary of the main conclusions.
2. OBSERVING SYSTEM SIMULATION EXPERIMENTS
OSSEs can be used to assess the potential impact of any new observing system,
provided that the error properties of the system are well understood. The basic elements
of an OSSE are a state-of-the-art data assimilation system, a nature run ”truth” and
a corresponding database of simulated observations (Atlas 1997; 2003). The latter
includes both simulated observations of conventional meteorological systems, covering
a network similar to the operational network, and simulated observations of the new
instrument to be assessed. Generation of the nature run and database for conventional
observation systems has been reported extensively in the past (Stoffelen et al. 1994;
Becker et al. 1996).
To build up a database of simulated observations one needs a description of the
atmosphere over a certain time period. For this purpose, a synthetic ”true” atmosphere
is generated through a long period integration of a forecast model. This is called
the ”nature run”. The nature run that we use in this study was the result of a 30-
day integration, initiated on 5 February 1993 00 UTC and ended on 7 March 1993
00 UTC. Integration was performed with the operational forecast model at ECMWF
in 1993, i.e. T213 (∼100 km) horizontal and 31 levels vertical resolution (Stoffelen et
al. 1994). An extensive evaluation of the nature run cloud cover has been performed at
the National Centre for Environmental Prediction (NCEP) within the National Oceanic
and Atmospheric Administration (NOAA) (Masutani et al. 1999). The nature run
cloud cover was compared to available data sets from space-borne and surface-based
observation systems in February 1993. From this study it was concluded that nature

1930 A. STOFFELEN et al.
N
A
T
U
R
E
R
U
N
ADM DWL
orbit +
shot pattern
ADM
observation
simulation
simulation new ADM instrument
”perfect”
simulated
observations
-
simulated
observations
(BUFR)
6 6
-
simulation operational data
coverage observationerror
? ?
-
-
-
assimilation
”with”
new instr.
assimilation
”without”
new instr.
4D
-V
ar -
-
-
-
forecast
”with”
new instr.
forecast
”without”
new instr.
impact
assessment
6
?
?
6
ff
ECMWF
OSSE
database
-
ff -observation simulation
ff -OSSE preparation ff -OSSE experiment
Figure 1. OSSE setup. Data assimilation ”with” (”without”) includes (excludes) the simulated observations of
the new instrument.
run clouds generally agree well with observations. Main differences are found over
both the North pole and South pole which show much more cloud cover in the nature
run. In addition, the nature run generally overestimates high-level cloud cover and
underestimates low-level cloud cover. On the other hand, high-level nature run cloud
optical depth showed good agreement with observations. For low-level cloud cover it
appeared that the nature run underestimates marine stratocumulus. Later on we show
that DWL data are most important in the free troposphere in regions of atmospheric
activity, so lack of marine stratocumulus is probably not serious in assessing lidar data
impact through the OSSE study. Moreover, the winds in the PBL of the atmosphere
below the stratocumulus clouds are relatively well sampled by the (simulated) ASCAT
scatterometer in our experiments. In summary, it is concluded that nature run cloudiness
is sufficiently representative of the real atmosphere.
Simulated observations for the OSSE are obtained through interpolation of the
nature run fields to observation locations. This results in so-called ”perfect” observa-
tions. For conventional observations, the locations are extracted from a real observa-
tion database to produce a representative sampling. Observation coverage charts can
be found in (Stoffelen et al. 1994). The data coverage for new observation instruments
such as a DWL needs to be simulated based on expected orbit characteristics and shot
pattern. Finally, realistic stochastic observation errors, including instrument error and a
representativeness error that accounts for the variability of the observed parameter on
scales that cannot be resolved by the NWP model, are added to the perfect observations
to simulate real observations, see figure 1. The result is a database with simulated obser-
vations of conventional meteorological observation systems, three earlier infrared lidar
concepts and the ESA selected UV concept, denoted ADM UV (Stoffelen et al. 2005).
To assess the impact of a simulated new observation system, two assimilation runs
need to be performed; one excluding and one including the new system, see figure 1. For
each run, we performed fifteen days of data assimilation, with an interval of six hours,
starting at 5 February 1993, 12 UTC and finishing at 20 February 1993, 12 UTC. On
each of these 15 days, one 10-day forecast was run from the corresponding analysis
at 12 UTC. Differences in the forecasts from the two parallel runs are only due to

DOPPLER WIND LIDAR OSSE 1931
the impact of the simulated new observation system, hence their respective quality is
a measure of observation impact.
(a) Fraternal twin problem
To be a useful tool for impact assessment of new instruments in NWP, OSSEs must
represent meteorological practice as closely as possible. In meteorological practice,
NWP model runs and nature truth diverge with time. In the OSSE the model run should
also diverge from the truth, represented by the nature run. This divergence may not be
completely realistic since error growth can be different in a genuine NWP system on the
one hand, and between the model runs involved in an OSSE. In practice, the models used
for the OSSE nature run and the experiments may be different, but nevertheless exhibit
related flaws. When that happens, the differences between short range forecasts initiated
from the data assimilation, and the nature run, may be smaller than genuine forecast
errors. This problem, called the ’fraternal twin’ problem, will bias the interpretation of
results from the OSSE experimentation.
In this study the nature run was generated using the 1994 ECMWF operational
forecast model, cycle 12r1. The OSSE has been conducted in 1999 with the 1999
operational ECMWF forecast model, cycle 21r1. Potentially this incurs the possibility
of two similar or fraternal twin models, depending on model evolution in the period
1994 to 1999. The most significant model changes in this period may be obtained from
the ECMWF webpages†, but in summary show substantial changes in model dynamics,
radiation and cloud parameterizations, and ancillary codes. Based on these changes, we
expect the nature run atmospheric model and the OSSE model to be as different as any
two other realistic models of the atmosphere. The experimental divergence of the nature
run and OSSE model run is checked in section 3.
3. LIDAR IMPACT ASSESSMENT
This section discusses the use of OSSEs to demonstrate the impact of the DWL
profiles on the atmospheric analyses and forecasts. The analysis step of the data
assimilation cycle combines the knowledge on the atmospheric state from observations
and a short range forecast, called background. The resulting most likely atmospheric
state constitutes a compromise between the observations and the background based on
their respective estimated errors (Lorenc 1986; Courtier et al. 1998). So, if at a particular
location the observation and the background disagree, then the model state is modified,
such that a more likely state results. The amplitude of the modification depends on the
estimated error covariance of the observation relative to the estimated error covariance
of the model (Derber and Bouttier 1999). The lower the estimated observation error is,
the more impact it has. The errors of the observations and the background are assumed
uncorrelated in the analysis.
(a) OSSE database extension
The operational ECMWF 4D-var assimilation system has been extended to enable
the proper assimilation of lidar data. This requires the modification of the existing
observation operator that relates observed variables to the model state vector for the
assimilation of lidar data. The lidar observation operator includes the interpolation of
model state parameters to locations of lidar observation and conversion of horizontal
† http://www.ecmwf.int/products/data/technical/model id/index.html

1932 A. STOFFELEN et al.
wind components to Horizontal Line-Of-Sight, HLOS, wind components (see Marseille
and Stoffelen (2003), in the remainder also denoted M&S).
M&S report results on a pre-OSSE analysis to assess profile quality in clear
air, i.e. without clouds, and on the impact of clouds on atmospheric penetration.
Moreover, wind shear and humidity flux visibility are assessed in relation to clouds. A
recent study (Tan and Andersson 2005) simulates the performance of ADM in aerosol-
rich atmospheres including realistic cloud scenes as measured by the Lidar In-space
Technology Experiment (LITE) in 1994 (Winker et al. 1996). In both studies it was
found that more than 90% of all Aeolus wind observations fulfil the WMO requirements
for wind quality. Here we note that optically thin clouds (such as cirrus) in the upper
troposphere return a strong signal that provide good-quality winds and generally have
very limited (negative) influence on the quality of underlying measured winds. In the
lower atmosphere data coverage is reduced by about 25% due to opaque clouds such as
stratus. In the simulation of lidar observations we assumed uncorrelated errors both in
the horizontal and the vertical. Lidar observation errors are further assumed unbiased
and have a Gaussian probability density function with known but variable standard
deviation (see M&S). In the assimilation of lidar HLOS wind components, we assume
perfect knowledge of observation uncertainty.
(b) Experimental setup
We define two experiments to assess the potential impact of the ADM UV concept
on NWP analysis and forecasts: (i) NoDWL (control) and (ii) DWL (control + DWL).
The NoDWL experiment includes the assimilation of conventional observations as
generated by (Stoffelen et al. 1994), i.e., TEMP, PILOT, AIREP, DRIBU, SYNOP, and
SHIP, and the satellite-inferred data from PAOB, SATOB and ASCAT. Scatterometer
winds are thinned resulting in a message structure containing nodes at 100-km sampling
in both directions of the swath as is normal practise for using ERS scatterometer data.
Cloud motion wind (SATOB) measurements were used at the spatial and temporal
density as available in February 1993. High density winds are available nowadays, but
do not provide substantially larger impacts in the ECMWF data assimilation system and
are thinned prior to use (Rohn et al. 2000). Again, only in the NH the results obtained
with this OSSE are representative of the complete observing system. We assessed the
possibility to assimilate (A)TOVS radiances. Issues of concern were, among others: (i)
incompatibility of the simulated TOVS radiances and the operational weather model,
because of the use of a now obsolete stratospheric extrapolation to simulate radiances,
and because of the OSSE NWP model to include the stratosphere, and (ii) lack of
calibration of a bias correction scheme for OSSE TOVS data.
As mentioned in the introduction, the 1993 OSSE database contains more than
twice as many conventional wind profile data than available in 1999 or nowadays.
On the other hand, the number of SATOB, aircraft winds, and the exploitation of
passive radiometer data, has improved over the last decade. Moreover, sounder data
are missing. Given the demonstrated importance of wind profile data in the ECMWF
data assimilation system, and based on OSE work at ECMWF (Kelly 1996; 2004) and
the more general experience at other meteorological centres we assume over all that
the 1993 OSSE data base is comparable in NWP information content to the current
observing system in the NH. As such we expect a noticeable but limited effect on the
lidar impact assessment in the northern hemisphere. However in the tropics and southern
hemisphere OSSE impact results are not representative of the impact expected in the
presence of todays GOS because of the generally dominating effect of satellite data
here.

DOPPLER WIND LIDAR OSSE 1933
The second experiment contains the same data as the control experiment, but in
addition the simulated lidar measurements of ADM UV. When forecasts and analyses
from this experiment compare better to the nature run than those of the control, then we
have demonstrated positive impact of the ADM in the data assimilation system as used.
The 4D-var incremental analysis is performed at T63 (∼300 km) resolution in the
horizontal and 31 levels in the vertical. The ECMWF 4D-var data assimilation system
contains several switches to include features of the model that are used operationally,
but are not really required to obtain representative results on DWL impact. The features
that are not used include coupling of the ocean wave model (WAM) to the atmosphere
model, and variable land surface fields, such as snow cover; these fields were fixed.
(c) Theoretical Assessment of Lidar Observation Impact
The impact of lidar data on NWP and climate studies is determined by the effec-
tiveness of the 4D-var system to assimilate lidar data. In an idealized situation all data
contain information and have a positive impact on the analysis quality. Meteorological
practice however is more complex as discussed in this section.
In variational assimilation the aim is to minimize a cost function that optimally
combines information from a short-term forecast and observations in a statistical manner
to arrive at a consistent description of the atmosphere. The incremental formulation of
the cost function J is as follows (e.g. Courtier 1997; Courtier et al. 1998)
J(δx) = 12δx
TB−1δx + 12v
TR−1v (1)
with,
v = Hδx − d, and d = y − Hxb (2)
where δx is the increment from the background (background), xb, that is obtained from
a forward model integration initialized with the analysis in the previous time window.
B is the estimated background error covariance matrix, v and d are called innovation
vectors, H is the linearized observation operator, y is the observation vector and R the
observation error covariance matrix. The optimal solution δxa of Eq. (1), also denoted
analysis increment, is added to the background xb to arrive at the analysis xa. The
solution can formally be written
xa = xb + K[y − Hxb], with K = BHT(HBHT + R)−1 (3)
Here, K is the Kalman gain matrix. Assuming that B and R are perfect estimates of
the error covariances it can be shown that for the analysis error covariance matrix A we
then have
A−1 = B−1 + HTR−1H (4)
Since R is positive definite Eq. (4) states that each observation adds information
(for non-zero H) and thus contributes to a reduction of the analysis error covariance,
A. One of the fundamental limitations of variational data assimilation is the lack of
exact knowledge of the background and observational error structures. In operational
practice one uses (imperfect) estimates of the B and R matrices. As a consequence, the
gain matrix K is generally not optimal, meaning that the information content of new
observations is not optimally exploited and might locally even result in mean negative
impact.
The discussion above implies that in an optimal 4D-Var, where the innovation
covariance matrix is correctly specified, all observations are useful and contribute to

1934 A. STOFFELEN et al.
a reduction of the analysis error. In the operational ECMWF 4D-Var, the usefulness of
an observation will depend on the accuracy of the assumed covariance matrices. Hence,
an observation is likely to be useful if
1. the observation error characteristics are sufficiently well known;
2. its assimilation is not affected by wrong model error assumptions;
3. its errors are uncorrelated with the background and other observations;
4. it is accurately characterized by its forward observation operator;
5. B accurately transforms to observed quantities at observation points, i.e. matrix
HBHT is accurately known.
In the OSSE, all conditions except (2) are fulfilled.
To determine lidar data impact on analyses we define xca and xla that denote the anal-
ysed state vector for the control (NoDWL) and lidar (DWL) experiment respectively, xt
denotes the nature truth. The analysis error covariance matrices of the NoDWL and
DWL experiment, denoted with Ac and Al, respectively are defined by
Ac = cov[xca − xt] = E[(xca − xt)(xca − xt)T]
Al = cov[xla − xt] = E[(xla − xt)(xla − xt)T] (5)
with E denoting the expectation operator. For the first analysis cycle of using DWL data,
it can be shown relatively easily that the inverses of both matrices are related through
A−1l = A−1c + HTl R−1l Hl (6)
with Rl the covariance matrix of lidar observation errors and Hl the lidar observation
operator. Eq. (6) shows that, in theory, all lidar data add information to the analysis
in addition to the conventional data, since Rl is positive definite. As outlined in the
beginning of this section, meteorological practice is less straightforward. This is clearly
illustrated in the next example.
(i) Single case example: 5 February 1993, 18 UTC. The analyses of the NoDWL and
DWL experiment are identical only at the beginning of the experiment at 5 February
1993, 12 UTC. The difference in the analyzed fields of both experiments at 18 UTC
is due to the addition of 6 hours of lidar data in the DWL experiment. To visual-
ize the impact of lidar data on the analysis we plot the differences of the root mean
squared errors (RMSE) of the analyzed fields of the NoDWL and DWL experiments,
both verified against the nature run, i.e. RMS(xca − xt) - RMS(xla − xt). For a single
case this reduces to |xca − xt| − |xla − xt|. Negative/positive values correspond to neg-
ative/positive impact of lidar data on the analysis. Figure 2 displays the lidar impact
for the 5 February 18 UTC analysis. Not surprisingly, the impact of the lidar data on
the wind field is concentrated near the measurement locations, indicated with crosses.
However, adjustment of the wind field is not isolated to lidar locations. The assimilation
system spreads the added information. Beside the positive regions, other places show
negative impact, which is caused mainly by the stochastic properties of the observation
and background errors. Here it is important to note that observations do not everywhere
yield a positive impact as suggested by Eq. (4). First of all, in real life the forecast model
is not perfect and model and observation errors (e.g. spatial representativeness) tend to
depend on the meteorological situation, and can even be systematic (bias). Furthermore,
the response of the model is non-linear, making background error covariance estimates
difficult to assess and inaccurate, and therefore contribute to a wrong relative weight of
observations. However, in line with Eq. (6), a positive impact on the RMS average is
obtained when the random errors are averaged over large areas.

DOPPLER WIND LIDAR OSSE 1935
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
-12
-7
-3
-1.5
-0.5
0.5
1.5
3
7
12
Figure 2. 500 hPa Wind field RMSE difference at 5 February 1993 18 UTC of NoDWL and DWL experiment,
both verified against the nature run. Red denotes negative impact, green denotes positive impact, white denotes
no significant impact. Black crosses indicate the lidar profile locations. The differences are due to 6 hours of lidar
data only.
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
-12
-7
-3
-1.5
-0.5
0.5
1.5
3
7
12
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
-12
-7
-3
-1.5
-0.5
0.5
1.5
3
7
12
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
-12
-7
-3
-1.5
-0.5
0.5
1.5
3
7
12
60°S60°S
30°S 30°S
0°0°
30°N 30°N
60°N60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
-12
-7
-3
-1.5
-0.5
0.5
1.5
3
7
12
Figure 3. Mean lidar observations impact on vector wind analyses over the complete assimilation period at
1000 hPa (upper left), 850 hPa (upper right), 500 hPa (lower left) and 200 hPa (lower right). Impact is visualised
by the difference of the RMSE of the NoDWL and the DWL run i.e. RMS(NoDWL-NR) - RMS(DWL-NR).
The mean is taken over 15 cases. White areas denote a negligible lidar impact, green areas a positive impact, red
denotes a negative impact.

1936 A. STOFFELEN et al.
(d) Results of the 15-day assimilation period
Subsection (i) discusses the impact of DWL on analyses for the complete 15-day
assimilation period. Subsection (ii) discusses data usage in the OSSEs and relates this
to the 1999 operational ECMWF 4D-var system to indicate how well the OSSE obser-
vational network relates to the 1999 operational network, with respect to observation
coverage and quality. Analyses serve as the forecasts initial state. The impact of lidar
data on forecasts is assessed based on the scheme of figure 1 and using some objective
measures to verify forecasts initialized with (DWL) and without (NoDWL) lidar data.
This is discussed in subsection (iii).
(i) Lidar impact on analyses. We compare the analyses of both experiments with the
nature run every day at 12 UTC, starting at 6 February 1993 and finishing at 20 February
1993, i.e. for 15 days. The mean squared errors (MSE) of the analyses wind vector
fields (verified against the nature run) are displayed in table 1 for different pressure
levels and regions of the globe. Here, the MSE is used rather than variances to take into
account possible biases. Table 1 shows a positive impact of lidar data on the analysis
TABLE 1. RMSE of analysis wind fields (m s−1) for the NoDWL and DWL experiments verified against the
nature run, i.e. Ac and Al respectively. The mean is taken over 15 cases, i.e. analyses at 12 UTC from 6 February
1993 until 20 February 1993.
domain boundaries 1000 hPa 850 hPa 500 hPa 200 hPa
N S W E NoDWL DWL NoDWL DWL NoDWL DWL NoDWL DWL
Globe 90 -90 -180 180 2.30 2.18 2.82 2.55 4.58 3.54 5.28 3.97
N.Hemis 90 20 -180 180 2.27 2.24 2.50 2.39 3.33 3.06 2.76 2.54
S.Hemis -20 -90 -180 180 2.57 2.30 3.12 2.70 5.63 3.84 6.37 4.08
Tropics 20 -20 -180 180 2.06 2.01 2.81 2.54 4.49 3.66 5.91 4.88
Europe 75 35 -12 42 1.53 1.52 1.62 1.59 1.77 1.73 1.71 1.68
N.Atlantic 75 20 -75 -5 2.50 2.49 2.63 2.53 3.46 3.15 3.14 2.88
N.America 75 30 -130 -75 1.63 1.57 2.08 1.91 3.04 2.69 2.92 2.45
at all considered pressure levels and regions. The impact increases with decreasing
pressure. Despite the generally high quality lidar data at 1000 hPa, their mean impact is
modest in all regions. This can be understood from the high-quality simulated ASCAT
scatterometer winds that have good coverage over the ocean surface.
Not surprisingly, a large impact is found in the tropics and southern hemisphere,
because of the reduced coverage of satellite data (no AMSU or TOVS). A smaller but
consistently beneficial impact is found over all areas in the northern hemisphere.
Figure 3 shows the mean global impact of lidar data on analyses averaged over the
assimilation period. Again, large impact is seen in the tropics and southern hemisphere
especially over the oceans. Positive impact is also seen over the North Atlantic and
Europe. We note a correlation between regions of negative lidar impact and regions of
low quality lidar data in figure 7 of M&S, in particular in the tropics. This indicates that
low-quality observations can on average deteriorate the analysis.
In particular, the a priori background error covariance estimate, whose evolution
in time is estimated from heuristic relationships, is uncertain in the OSSE. On the
other hand, the observation error structure is perfectly known. Let us elaborate on this.
Equation (4) provides the analysis error covariance as a function of background and
observation error covariances. We identify two cases
1. Observations have relatively low quality, in which case the estimated analysis
quality is entirely determined by the uncertain background error.

DOPPLER WIND LIDAR OSSE 1937
6
6
6
6 6
6
6
6
6 4
60°S
30°S
0°
30°N
60°N
120°W 60°W 0° 60°E 120°E
4
4 4
66
6
6
6
6
6 6
6
66
9
9 9 9
9
9
9
9
9
9
60°S
30°S
0°
30°N
60°N
120°W
120°W 60°W
60°W 0°
0° 60°E
60°E 120°E
120°E
2
4
6
9
12
15
20
Figure 4. 500 hPa standard deviation of wind vector background error (m s−1) estimated over the period 6
February 1993 00 UTC until 19 February 1993 18 UTC. Left, a priori estimate as used in 4D-Var. Right, from
background minus nature run field statistics.
2. The background is of relatively low quality, in which case the estimated analysis
quality is determined by the well-known observation error.
Case 1 appears the most problematic; in particular, when the estimated background error
is wrong and too high. In this case the poor observations are assigned too much weight
and are overfitted. This obviously could be harmful to the extent that the analysis is
deteriorated with respect to the background. On the other hand, if the background error
is too low, then the poor observations have too little impact, which is not optimal, but
probably only to the extent that the improvement upon the background by the analysis
is too small. So, low quality observations in the presence of overestimated background
errors appear most detrimental.
We checked the global distribution of a priori specified background error standard
deviations with the a posteriori computed error variances to confirm the occurrence of
case 1. Figure 4 shows that 4D-Var overestimates the background error over the tropical
and subtropical continents. Consequently, the relatively poor lidar data in these regions,
see M&S figure 7, are assigned too much weight resulting locally in negative impacts,
see Fig. 3. Also, 4D-Var underestimates the background error in the North Atlantic,
leaving good quality lidar data insufficient weight to correct the analysis. Further note
the difference in spatial detail between the real and estimated error covariances, showing
the uncertainty in the estimate of B as used in 4D-Var.
In reality, we also expect that the observation error structure is generally better
known than the background error structure, and as such an OSSE seems ideal to
test data assimilation systems. A second conclusion from the above is obviously that
observation quality control is critical for data assimilation, since it may reduce the
negative consequences of case 1 above (Tan and Andersson 2004).
(ii) Data Usage. The operational assimilation system at ECMWF archives informa-
tion on the usage of observations by the system. This includes information on whether
the observations are used or rejected in the assimilation cycle. Rejection may be the
result of quality control or data blacklisting. The difference of the used data with the
background and analysis field is stored to check the performance of the system. After
an experiment, fits of the observations to the background and analysis fields are gen-
erated and visualized by standard RMS plots, bias plots, and histograms (not shown).
From these statistics it is concluded that the simulated lidar winds are unbiased and
have overall a slightly lower quality than radiosonde (TEMP) winds (3-5 m s−1vs. 2-4
m s−1RMS background departure) and highest values in the tropics (3-7 m s−1vs. 3-5
m s−1RMS). We note that as expected the DWL data have more heterogeneous quality

1938 A. STOFFELEN et al.
TABLE 2. Global observation coverage and statistics of OSSE related to the operational ECMWF system in
1999 for the same period, i.e. 5 February 18UTC to 16 February 12UTC. (o-b) Denotes the background departure
and (o-a) the analysis departure.
OSSE experiment Operations 1999
number of data RMS number of data RMS
data o - b o - a data o - b o - a
TEMP-wind [m s−1] 830,118 3.2 2.8 307,301 3.7 3.0
TEMP-T [K] 400,290 3.8 3.7 402,404 3.0 2.8
TEMP-q [kg/kg] 264,425 0.16e-2 0.15e-2 215,811 0.24e-2 0.23e-2
PILOT [m s−1] 328,870 3.2 2.8 241,334 3.7 3.0
AIREP-wind [m s−1] 100,060 5.8 5.3 920,698 4.3 4.0
AIREP-T [K] 66,774 2.6 2.5 407,484 2.2 2.2
TOVS [K] 0 - - 2,347,298 6.5 5.0
LIDAR [m s−1] 532,992 4.2 3.4 0 - -
SYNOPship-10U [m s−1] 49,794 3.0 2.8 67,754 3.9 3.8
DRIBU-10U [m s−1] 2,618 5.1 4.9 11,712 3.3 3.0
SCAT-10U [m s−1] 99,685 2.7 2.1 114,756 1.7 1.2
SYNOPland-2RH [%] 24,727 14.0 14.0 282,571 14.0 14.0
SYNOPship-2RH [%] 16,960 15.0 13.0 33,948 16.0 15.0
SATOB-wind [m s−1] 37,576 4.3 4.1 981,886 5.3 5.2
PAOB (Pa) 3,255 229 207 3,353 307 273
RAOB-wind [m s−1] 1,183,830 3.9 3.5 855,975 5.2 4.6
For instruments measuring profiles, the number of data equals the sum of data at all levels. The data RMS is an average
over all levels. For SCAT sea-surface winds only the closest 10m u-wind vector of the two available ambiguities is
considered.
than radiosonde data, but that high-quality DWL winds are given more weight in the
data assimilation than lower quality DWL winds. RMS fits to other data types were
generally improved when DWL data were used. Very few HLOS winds were rejected
by the variational quality control, consistent with the use of Gaussian errors.
Interpreting results from the DWL OSSE in terms of expected impact of DWL
observations in 2008, when ADM will fly, is not trivial since it requires a comparison
of the 1993 and the (yet unknown) 2008 observational network. As a first approach we
compared the observational network as generated in Stoffelen et al. (1994) with the 1999
operational network, i.e. at the time the OSSE was conducted. To this end we compared
the observation statistics of the OSSE with the operational observation statistics in the
February period of 1999. The results are summarized in Table 2 and show that the OSSE
uses more radiosondes (TEMPs), less AIREPs, less SATOBs and less DRIBUs.
Besides their relative abundance in 1993, the simulated data quality of the network
of wind sounders (TEMP, PILOT) is overestimated as compared to reality in 1999.
Since the wind sounding network is the backbone for NWP (e.g., WMO 2004), one
would thus expect that the impact of real lidar data would be more significant in
the operational system than the simulated data in the OSSE. The effect of reduced
wind profile information available to NWP, currently, is however compensated by the
abundance of other data types (see introduction).
(iii) Lidar impact on forecasts. To assess the impact of lidar data on forecasting, 10-
day forecasts are produced, initiated with the 15 analyses at 12 UTC, i.e. from day 6 to
20 February 1993. Several objective statistical measures to verify forecast quality are
proposed in the literature. Most popular among these are the root-mean-square error
(RMSE) and anomaly correlation coefficient of forecasts against the analyses. In the
special case of an OSSE we verify forecasts against the nature run. Figure5 shows the
wind vector RMSE of the forecasts with respect to the nature run at 500 hPa for the

DOPPLER WIND LIDAR OSSE 1939
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
2
4
6
8
10
12
14
16
18
20
22
m
/s
500 hPa WIND RMSE AREA: EUROPE
NoDWL DWL
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
4
6
8
10
12
14
16
18
20
m
/s
500 hPa WIND RMSE AREA: N.ATL
NoDWL DWL
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
2
4
6
8
10
12
14
16
18
m
/s
500 hPa WIND RMSE AREA: N.AMER
NoDWL DWL
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
2
4
6
8
10
12
14
16
18
m
/s
500 hPa WIND RMSE AREA: N.HEM
NoDWL DWL
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
4
6
8
10
12
14
16
18
m
/s
500 hPa WIND RMSE AREA: S.HEM
NoDWL DWL
0 1 2 3 4 5 6 7 8 9 10
Forecast Day
3
4
5
6
7
8
m
/s
500 hPa WIND RMSE AREA: TROPICS
NoDWL DWL
Figure 5. Forecast skill; 500 hPa wind vector RMSE of forecast (w.r.t. nature run) field for NoDWL (dashed)
and DWL (solid) experiment as a function of forecast range and for six global regions. The top row shows the
forecast skill for the northern hemisphere (left), southern hemisphere (center) and tropics (right), the bottom row
for Europe, the North Atlantic and North America. Forecasts are initialised with analyses at 12 UTC in the period
6 until 20 February 1993. The mean is taken over all 15 cases.
NoDWL and DWL experiments for different global regions. The mean is taken over
all 15 cases. Forecast day 0 corresponds with the analysis. The northern hemisphere
regions show forecast improvements up to half a day. Similar results are found at other
pressure levels (not shown). Anomaly correlation computations (see also section 4)
for geopotential height show similar improvements. Values remain above 60% up to
forecast day 7-8 for the northern and up to day 5-7 for the southern hemisphere (not
shown). The small positive impact over Europe after 2 days originates from the positive
impact of lidar data on the analysis over the North Atlantic as depicted in Fig. 6.
The positive impact over Europe and the North Atlantic is clear, evolving into
half a day forecast gain after 6 days. Local negative impact is observed as well in
the northern hemisphere. Negative forecast impacts seem associated with the locally
negative analysis impacts as described in the previous section. This confirms again the
relevance of a well-tuned data assimilation system and a careful observation quality
control that rejects or downweights relatively inaccurate observations in the analysis.
The total northern hemisphere impact is positive after averaging the local realiza-
tions of the impact scores. We note again, that due to the random noise in the observa-
tions and the chaotic behaviour of the atmosphere a stochastic behaviour of the scores
is expected. As such, it is an important result that the forecast scores vary considerably
from one day to the next, see for example the left figure in Fig. 7, nevertheless for the
northern hemisphere 14 out of 15 forecasts have improved, as shown in the right figure
of Fig. 7, providing convincing evidence of the future benefit of ADM to NWP.

1940 A. STOFFELEN et al.
4. OSSE CALIBRATION
An important aspect for the interpretation of OSSE results is to validate the realism
of the experimental setup. A common approach is to conduct a series of OSEs for
existing observing systems in the OSSE system (Atlas 1997, Masutani et al. 2004).
Ideally, the OSSE impact of simulated observations should agree with the OSE impact
of the corresponding real observations. Calibration of the OSSE results is required
for a realistic assessment of the expected impact of the new observing system in the
operational NWP system. Any changes in data coverage between OSE and OSSE
would have to be accounted for (see table 2). We adopt an alternative approach, where
we compared the distributions of background and analysis departures for the various
observation types used in the OSSE experiments with those in the ECMWF operational
system. We show that the impact of observations in the OSSE analysis is similar to
that in 1999 operations. In addition we show that the forecast anomaly correlation
coefficient, which is a measure of forecast skill, for the OSSE fits well with operations
30°N
40°N
50°N
60°N
70°N
80°N
80
°W
60
°W
40°W 20°W 0°
20°E
40°E
60°E
80°E
0.25
0.5
1
2
5
-5
-2
-1
-0.5
-0.25
30°N
40°N
50°N
60°N
70°N
80°N
80
°W
60
°W
40°W 20°W 0°
20°E
40°E
60°E
80°E
1
2
4
6
8
-8
-6
-4
-2
-1
Figure 6. 500 hPa lidar observations impact on the wind field (m s−1) over Europe and the North Atlantic for the
analysis (left) and 4-day forecast (right), averaged over 15 cases. The difference of the RMSE of the NoDWL run
and the DWL run is plotted. White areas denote a negligible lidar impact and filled/dashed areas a positive/negative
impact. Lidar data have a positive impact on the 4-day forecast over Europe and the North Atlantic.
M’
6
FEBRUARY 1993
7 8 9 10 11 12 13 14 15 16 17 18 19 2015
20
25
30
35
40
45
50
55
60
65
0 10 20 30 40 50 60 70 80 90
DWL Forecast
0
10
20
30
40
50
60
70
80
90
Co
nt
ro
l F
or
ec
as
t
15 Cases
Mean
Figure 7. Left, forecast verification. RMS error of the forecasts of 500 hPa geopotential height (m) over North
America for three (dashed)- and four (solid)-day DWL forecasts and three (dot)- and four (dash-dot)-day NoDWL
forecasts. Right, 500 hPa geopotential height RMS error (m) of the DWL and NoDWL four-day forecasts north
of 20N. Circles indicate the 15 individual forecasts. The cross represents the mean over the 15 cases.

DOPPLER WIND LIDAR OSSE 1941
over the 1993-1999 time period. Both results confirm the absence of a fraternal twin
problem as was anticipated in section 2(a).
(i) Observation Impact. The divergence of the OSSE forecast model from nature truth
is compensated through the input of observed meteorological data in the assimilation
cycle. Hence, observation impact is related to the extent to which the weather model
diverges from the true atmospheric evolution. For small divergence, the background
fields after six hours of forward integration and the corresponding nature run fields will
be very similar, hence underestimating the impact of additional observations. This so-
called fraternal twin problem was investigated by comparing the observation minus
background differences and observation minus analysis differences of the OSSE in
February 1993 and the operational system at ECMWF in February 1999. Background
(analysis) departure is defined as the departure (y − Hx) between the background
x = xb (respectively the analysis x = xa) and observations y with H the observation
operator that relates model fields to observations. Introducing the ”true” nature run
fields, xt, the departure expression can be written as follows
y − Hx = y − H(xt + x − xt)
= y − Hxt + H(xt − x)
= r + H(xt − x) (7)
with r the observation error. Assuming no correlation between observation and back-
ground field errors, the covariance matrix of the background departures equals Db =
R + HBtHT, with R the observation covariance matrix and Bt the true background
error covariance matrix defined by
R = E[(y − E[y])(y − E[y])T]
Bt = E[(xb − xt)(xb − xt)T] (8)
Using Eq. (3) we may write for the analysis departure y − Hxa = [I − HK](y −Hxb).
Note that the analysis error and observation error are correlated. For the covariance
matrix of the analysis departures, Da, we then have Da = [I − HK]Db[I − HK]T.
For the unlikely event of having perfect knowledge of the background error covariance
matrix the Kalman gain from Eq. (3) is optimal in minimizing the analysis departure
and Da simplifies to Da = [I − HK]R.
We compared the RMS of background and analysis departures for the various
observation types used in the OSSE experiments and in the ECMWF operational system
respectively. The RMS is equal to the square root of the diagonal elements of the
covariance matrices Da and Db. In fraternal twin experiments, the background will
be much closer to the truth than in operations. Then, the true background errors (Bt)
are much smaller than observation errors (R). Additional data thus would have minimal
impact or may even be detrimental for the analysis. The latter is understood by the
fact that the analysis, observation and background weights are pre-determined with
anticipated deviations of the background from truth (based on experience in operations).
The a priori background error covariance (B) is much larger than it truly should be,
resulting in modifications of an accurate background on the basis of relatively inaccurate
observations. In other words, since the true gain matrix Kt is unrealistically smaller than
the estimated K (for B is larger than Bt), observation increments are overestimated
by the assimilation system. Thus fraternal twin experiments generate small and often
negative impact of observations and this results in almost similar background and
analysis departure statistics. However, the RMS differences of background and analysis

1942 A. STOFFELEN et al.
0 2 4 6 8 10
1000
850
700
500
400
300
250
200
150
100
70
50
30
20
10
5
0 2 4 6 8 10
1000
850
700
500
400
300
250
200
150
100
u-wind (ms−1) u-wind (ms−1)
TEMP-Uwind RMS AIREP-Uwind RMS
pr
es
su
re
(h
Pa
)
Figure 8. Statistics of TEMP u-wind component (left) and AIREP u-wind components (right) in OSSE (black)
and the operational ECMWF system (grey) for the period 5 February 1999 12 UTC until 8 February 1999 06 UTC.
Solid/dashed lines denote background/analysis departures, i.e. (o-b)/(o-a).
departures of the OSSE and those from operations are quite similar (Fig. 8). This implies
realistic impact of observations in our OSSE and thus realistic divergence of the OSSE
NWP model from the (nature run) truth.
(ii) Anomaly correlation of OSSE versus operational system. Anomaly correlation
coefficients provide an indication of the forecast skill. Anomaly correlation is defined
as the correlation between the forecast and analyzed deviations (anomalies) from
climatology (Holton 1992; Wilks 1995). The anomaly correlation for a particular
forecast variable is defined as follows
acc(i) =
∑
m
δFm(i)δAm(i)
√
∑
m
δF 2m(i)
∑
m
δA2m(i)
, with δFm(i) = (Fm(i) − Cm) − (Fm(i) − Cm)
(9)
and similar for δAm(i). Here Fm(i) is the i-day forecast field variable at grid point m,
and Am and Cm the corresponding verifying analysis and climatology field variables.
The summation is over an area of interest and the overbar denotes the area mean value.
Values for the anomaly correlation are between -1 and 1, with 1 implying a perfect
forecast. A value larger than 0.6 is generally regarded as an indication of a useful
forecast (e.g. Krishnamurti et al. 2003).
For fraternal twin nature run and operational forecast models one would expect a
much better skill compared to the operational system, since fraternal twin NWP models
would exhibit a more similar time evolution. In fig. 9 we compare the OSSE forecast
skill with the skill of the operational system in the years 1993 until 1999 by computing
anomaly correlations computed from ten 5-day forecasts in the OSSE assimilation
period (i.e. 6 February until 20 February). We concentrated on the northern hemisphere,
more specifically the North Atlantic and Europe, where the OSSE observing system is
representative of the 1999 operational observing system. Forecast skill difference of the
operational system for different years is related to different meteorological situations
and evolution of the forecast model.
The forecast skill of the OSSE is better than for the operational system in 1993 by
roughly half a day on the short-term, but almost similar in the mid-term. Note however
that the OSSE weather is different from the real weather of February 1993, and half
a day of shift in the scores is well within the year-to-year variations and improvement
of the skill of the operational system in the month of February from 1993 to 1999. We
conclude that the forecast skill performance in the OSSE is not significantly better than
in the operational system.

DOPPLER WIND LIDAR OSSE 1943
0 1 2 3 4 5
Forecast Day
65
70
75
80
85
90
95
100
%
Figure 9. 500 hPa geopotential height anomaly correlation coefficients for the OSSE related to the ECMWF
operational system (OPER) for the North Atlantic and European region in the period 1993-1999. The black solid
line corresponds to the OSSE, black-dash (OPER 1993), light-grey-dash (OPER 1994), dark-grey-dot (OPER
1995), light-grey-dot (OPER 1996), black-dot (OPER 1997), dark-grey-solid (OPER 1998) and light-grey-solid
(OPER 1999).
We conclude that assessment of the potential impact of lidar data on NWP through
OSSEs is not significantly affected by the fraternal twin problem in this study.
5. CONCLUSIONS
In this study we realistically simulated the meteorological impact of the UV
Doppler Wind Lidar as proposed for the ESA Core Earth Explorer Atmospheric Dy-
namics Mission, ADM UV. ADM UV has a clear and demonstrable positive impact
on the analyses and forecasts in the northern hemisphere. In the tropics and southern
hemisphere the impact is overwhelmingly positive, but here the OSSE observing system
is not representative of the real-world observing system. In particular in the southern
hemisphere, the satellite temperature soundings could unfortunately not be used. How-
ever, based on current operational experience, this is supposedly of little limitation in the
northern hemisphere in the presence of the extended radiosonde coverage as available
in 1993.
The average benefit of lidar data on medium-range (5-day) 500 (200, not shown) hPa
wind forecast in the OSSE was about 0.25 (0.4) days in the northern hemisphere (above
20N). Local impacts varied and were up to 0.5 (0.8) days, for example for Europe.
To test the significance of our results we verified that time series of forecast impact
showed sufficient day-to-day variability. At the same time, in a clear majority of cases
the DWL forecast was better than the control, indicating that our results are significant,
even though obtained over a limited period of 15 days.
Good quality ADM UV wind observations have a clear and beneficial impact on the
analyses. Some large and beneficial forecast impacts of ADM UV can be traced back
to areas with large analysis impact. Wind profile observations are of key importance to
the GOS, as demonstrated here again. However, the operational wind profile network is
expected to further decrease in the future. As an illustration of this fact we note that the
conventional wind profile network in operations is much smaller than that used in the
OSSE. This will result in a larger impact of satellite data in the future in the northern
hemisphere, both for mass and wind observations. Moreover, the simulated quality in
the OSSE database was somewhat overestimated for the conventional wind profiles.
This reduces the improvements brought by ADM UV in the OSSE. On the other hand,

1944 A. STOFFELEN et al.
more AIREP and wind sounders are available nowadays, mainly resulting in tropopause
flight level observations, but also some profiles over land.
Moreover, a closer look was given to the nature run clouds, but no serious defi-
ciencies were found. The relative lack of PBL clouds over the oceans as compared to
satellite observations may be improved. However, we found that in the PBL over the
ocean, the DWL impact is limited due to the presence of the ASCAT scatterometer. On
the other hand, inaccurate ADM UV data cause negative impacts locally. This occurs
probably because those observations are not properly weighted against the background
model fields in the analysis. The background error estimates are locally poor, probably
frequently resulting in detrimental observation impacts in the analysis. Excessive weight
given to low-quality observations cause detrimental impact. Underweighted high-quality
ones are usually beneficial. In areas with extensive high-level cloud cover negative im-
pacts were most frequent. We may conclude from this that (i) tuning of data assimilation
systems is very important for achieving beneficial observation impact (Tan and Ander-
sson 2004) and OSSE could be used for this, (ii) good accuracy and representativeness
of observations is a prime requirement for their impact and (iii) quality control on real
observations is very important in cloudy regions (Tan and Andersson 2004).
We rigorously tested the presence of a so-called fraternal twin problem, but found
no substantial evidence of such a problem. Although we have verified in this study
that ADM UV is indeed capable of demonstrating the potential value of space-borne
wind profile observations for improving atmospheric analyses and NWP, this study was
of limited extent and more experimentation is desirable as outlined in the following
recommendations:
• OSSEs for other and more periods would reveal more about the significance of
the results that we have found here. A two-week assimilation period is generally
thought of as the minimum to be able to demonstrate impact with an OSE or OSSE.
• OSSEs can be used to tune data assimilation systems.
• Quality control is very important. In the OSSE low-quality ADM UV observations
show often detrimental impact. Observations from LITE (Winker et al. 1996)
have shown their usefulness to investigate the interaction of a lidar with a cloudy
atmosphere and to study quality control issues (Tan and Andersson 2004; 2005).
Also the ICESat mission (Spinhirne 2005) and air and ground measurements may
help to verify processing schemes.
• Where ADM UV is designed to demonstrate the capability of a space-borne DWL,
OSSE-like studies may be used to study scenarios for an operational meteorologi-
cal mission to be implemented when ADM UV has successfully flown. Options
for targeting LOS profiles, multiple LOS or even multiple satellites could be tested
(Marseille et al. 2005).
• To entirely avoid the fraternal twin problem we recommend the use of a foreign
model for the production of the nature run. These fields can then be interpolated
and processed at any location to provide an OSSE database in standard meteoro-
logical format. The ECMWF has a great capability to run OSSEs on such input.
• OSSEs including (A)TOVS and AMSU data would be better capable of assessing
the relative benefit of temperature and wind sounding in the southern hemisphere
and tropics. Simulation of AIRS or IASI or other new observation systems is also
worthwhile. However, we note that for these observations, cloud clearing is a major
issue and consequently error properties are complex and more difficult to simulate
realistically.

DOPPLER WIND LIDAR OSSE 1945
• OSSE are costly and alternative impact simulation experiments may be developed
such as in (Marseille et al. 2005) that include the nowadays operational network
of satellite sounders.
ACKNOWLEDGEMENT
We thank ESA for their support of the project under Contract No. 13018/98/NL/GD.
Siebren de Haan did a great job in interfacing the OSSE observation data base to the
ECMWF system. We acknowledge staff at ECMWF and Me´te´o France for supporting
this study, in particular Drasko Vasiljevic¸ and Carla Cardinali. Special acknowledgement
goes to ECMWF for maintaining the OSSE database and nature run in their archives. We
appreciate the motivating interests of Joachim Fuchs (ESTEC), Paul Ingmann (ESTEC),
and Uwe Kummer (Astrium) in this study.
REFERENCES
Atlas, R. 1997 Observing System Simulation Experiments: Methodology, Exam-
ples and Limitations. In proceedings of CGC/WMO Work-
shop Impact of various observing systems on Numerical
Weather Prediction, Geneva, 7–9 April, World Weather
Watch Technical Report No 18, WMO/TD No. 868, pp 155–
163.
Atlas, R., Emmitt, G. D., Terry, J.,
Brin, E., Ardizzone, J.,
Jusem, J. C., and Bungato, D.
2003 Recent observing system simulation experiments at the NASA
DAO. In proceedings of The Seventh Symposium on Inte-
grated Observing Systems: The Water Cycle
Baker, W., Emmitt, G. D.,
Robertson, F., Atlas, R.,
Molinari, J. E., Bowdle, D. A.,
Peagle, J., Hardesty, R. M.,
Menzies, R. T.,
Krishnamurti, T. N.,
Brown, R. A., Post, M. J.,
Anderson, J. R., Lorenc, A. C.,
McElroy, J.
1995 Lidar-Measured Winds from Space: A Key Component for
weather and Climate Prediction, Bull. Am. Meteorol. Soc.,
76, 869–888
Becker, B. D., Roquet, H,
Stoffelen, A.
1996 A simulated data base including ATOVS, ASCAT and Doppler
Wind Lidar data, Bull. Am. Meteorol. Soc., 77, 2279–2294
Butterworth, P. and Ingleby, N. B. 2000 Recent Developments in the use of Satellite Winds at the UK
Met. Office. In proceedings of the Fifth Int. Winds Work-
shop, 28/2-3/3/, Lorne, Australia, EUMETSAT, Darmstadt,
Germany
Cardinali, C., Pailleux J.,
Th e´paut J. N
1998 Use of simulated Doppler Wind Lidar data in NWP: an impact
study, CNRS, M e´t e´o France.
Courtier, P. 1997 Variational methods, J. Meteorol. Soc. Jpn., 75, 211–218
Courtier, P., Andersson, E.,
Heckley, W., Pailleux, J.,
Vasiljevic, D., Hamrud, M.,
Hollingsworth, A., Rabier, F.,
Fisher, M.
1998 The ECMWF implementation of three dimensional variational
assimilation (3D-Var). Part I: Formulation, Q. J. R. Meteorol.
Soc., 124, 1783–1808
Cress, A. and Wergen, W. 2001 Impact of profile observations on the German Weather Service’s
NWP system. Meteor. Zeitschrift, 10, 91–101
Derber, J., Bouttier, F., 1999 A reformulation of the background error covariance in the
ECMWF global data assimilation system, Tellus, 51A, 195–
221
European Space Agency 1996 The nine-candidate earth-explorer missions-Atmospheric Dy-
namics Mission. ESA Report ESA-SP-1196(4), European
Space Agency, Noordwijk, The Netherlands
European Space Agency 1999 Atmospheric Dynamics Mission. The four-candidate earth-
explorer missions. Report for mission selection. ESA Report
ESA-SP-1233(4), European Space Agency, Noordwijk, the
Netherlands.
Holton, J. R 1992 An introduction to Dynamic Meteorology. third edition, Academic
Press

1946 A. STOFFELEN et al.
Kelly, G. 1996 Influence of observations on the operational ECMWF system,
WMO Bulletin Vol. 46, No. 4, October, pp. 336–342 (in the
English version).
Kelly, G. 2004 OSEs of all main data types in the ECMWF operational system.
In proceedings of the Third WMO Workshop on the Impact
of Various Observing Systems on NWP, Alpbach.
Krishnamurti, T. N, Rajendran, K.,
Vijaya Kumar, T. S. V.,
Lord, S., Toth, Z., Zou, X.,
Cocke, S., Ahlquist, J. E.,
Navon, I. M.
2003 Improved Skill for the Anomaly Correlation of Geopotential
Heights at 500 hPa, Mon. Weather Rev., 131, 1082–1102
Lorenc, A. C. 1986 Analysis methods for numerical weather prediction. Q. J. R.
Meteorol. Soc., 112, 1177–1194
Marseille, G. J., Stoffelen. A. 2003 Simulation of wind profiles from a space-borne Doppler wind
lidar. Q. J. R. Meteorol. Soc., 129, 3079–3098
Marseille, G. J., Stoffelen. A.,
Barkmeijer, J.
2005 PIEW - Prediction Improvement of Extreme Weather, study report
available from KNMI
Masutani, M., Woollen, J. S.,
Lord, S. J., Derber, J. C.,
Emmitt, G. D., Wood, S. A.,
Greco, S., Atlas, R., Terry, J.,
Kleespies, T. J., Sun, H.
2002 Impact assessment of a doppler wind lidar for NPESS/OSSE.
AMS preprint volume for 15th Conference on Numerical
Weather Prediction 12–16 August 2002 in San Antonio, TX.
346-349.
Masutani, M., Campana K. A.,
Lord, S., Yang, S. K.
1999 Note on Cloud Cover of the ECMWF Nature run used for
OSSE/NPOESS project, NCEP Office note No. 427.
Masutani, M., Lord, S. J.,
Woollen, J. S., Yang, W,
Sun, H., Kleespies, T. J.,
Emmitt, G. D., Wood, S. A.,
Katz, B., Treadon, R,
Derber, J. C., Greco, S.,
Terry, J.
2004 Global OSSE at NCEP. In proceedings of the Eighth Sympo-
sium on Integrated Observing and Assimilation Systems for
Atmosphere, Oceans, and Land Surface
Rohn, M., Kelly G. 2000 The use of the MPEF quality indicator, In Proceedings of the
ECMWF seminar: Exploitation of the New Generation of
Satellite Instruments for Numerical Weather Prediction, 4–
8 September, available at ECMWF, Reading, UK.
Spinhirne, J. D. 2005 Aerosol and Cloud Observations and Data Products by The GLAS
Polar Orbiting Lidar Instrument, Invited Paper in proceed-
ings AMS Symposium on Lidar Atmospheric Applications,
San Diego.
Stoffelen, A., Becker, B., Eyre, J.,
Roguet, H.
1994 Theoretical Studies of the Impact of Doppler Wind Lidar Data -
Preparation of a database, ESA-CR(P)-3943.
Stoffelen, A., Pailleux, J.,
Ka¨ll e´n, E., Vaughan, J. M.,
Isaksen, I., Flamant, P.,
Wergen, W., Andersson, E.,
Schyberg, H., Culoma, A.,
Meynart, R., Endemann, M.,
Ingmann, P
2005 The Atmospheric Dynamics Mission for Global Wind Field Mea-
surement, Bull. Am. Meteorol. Soc., 86, 73–87
Tan, D., Andersson, E. 2004 Expected Benefit of Wind Profiles from the ADM-Aeolus
in a Data Assimilation System, ESA contract report
15342/01/NL/MM.
Tan, D., Andersson, E. 2005 Simulation of the yield and accuracy of wind profile measure-
ments from the Atmospheric Dynamics Mission (ADM-
Aeolus) Q. J. R. Meteorol. Soc., 131, 1737–1757
Wilks, D. S. 1995 Statistical Methods in the Atmospheric Sciences, Academic Press
Winker, D. M., Couch, R. H.,
McCormick, M. P.
1996 An Overview of LITE: NASA’s Lidar In-Space Technology Ex-
periment, Proc. of the IEEE, vol. 84, no. 2
WMO 2004 Current Statement of Guidance Regarding How Well Satellite Ca-
pabilities Meet WMO User Requirements in Ten Application
Areas. WMO satellite reports.