METEOROLOGICAL DATA PERSPECTIVE
347
integrated over a layer of the atmosphere. A set of a few of these
measurements thus describes broad vertical structures in temperature and
humidity. Data assimilation in some way or another converts these radiance
measurements in temperature/moisture profiles. Different possibilities exist
to process this information. One can use externally generated retrievals
(profiles deduced from a set of radiances through regression typically),
interactive retrievals using in-house information about short-range forecasts
(e. g. 1D-Var retrievals), or the direct use of radiances (e.g. 3D-Var or 4DVar). In NWP at least, the direct assimilation of satellite raw radiances has
progressively replaced the assimilation of retrievals (Thépaut, 2003). This
has been made possible because 3D and 4D-Var allow for some (weak) non
linearities in the observation operator, and radiances are non-linearly linked
with the atmospheric profiles. Retrievals always need prior background
information, which either comes from independent statistics or from the
short-range forecast. The direct assimilation of radiances has the advantage
to avoid the contamination by such an external background information for
which error characteristics are poorly known. Another advantage of global
variational methods is that increments brought by satellite radiances are
further constrained by many other observations/information. Finally, raw
radiance observations exhibit less spatially correlated errors than processed
retrieved information. In current data assimilation schemes, this allows to
use observations with more spatial density, a subject which will be discussed
further in the next section. Of course this use of raw data comes at the cost
of developing the observation operator and the quality control appropriate
for each observation for each data assimilation system in each NWP centre,
but some of this effort is collaborative through EUMETSAT facilities for
instance.
Zooming now on the period covering the most recent years, Figure 3
shows the number of data used in the ECMWF analysis between 1997 and
2003. This illustrates the tremendous increase in terms of observation
numbers which took place lately, and most of this progression in data
numbers comes from non-conventional asynoptic observations.
Such observation numbers have a significant impact, especially in an
advanced data assimilation scheme such as 4D-Var which has been used
since 1997 (Rabier et al, 2000). 4D-Var stands for Four-Dimensional
Variational Data assimilation and it performs a global optimization of the
model trajectory over a period of 6 to 12 hours typically. It performs an
adjustment of the model trajectory with the observations taken explicitly at
the precise time of the observation, thus allowing for a consistent use of data
spread in time throughout the optimization period, such as satellite
observations. In the linear approximation, 4D-Var is equivalent to a Kalman
smoother: at any time in the assimilation window, information from past and
future observations within this window will be taken into account to provide
the best estimate of the flow (Rabier and Liu, 2003). It can also use the time-
347
integrated over a layer of the atmosphere. A set of a few of these
measurements thus describes broad vertical structures in temperature and
humidity. Data assimilation in some way or another converts these radiance
measurements in temperature/moisture profiles. Different possibilities exist
to process this information. One can use externally generated retrievals
(profiles deduced from a set of radiances through regression typically),
interactive retrievals using in-house information about short-range forecasts
(e. g. 1D-Var retrievals), or the direct use of radiances (e.g. 3D-Var or 4DVar). In NWP at least, the direct assimilation of satellite raw radiances has
progressively replaced the assimilation of retrievals (Thépaut, 2003). This
has been made possible because 3D and 4D-Var allow for some (weak) non
linearities in the observation operator, and radiances are non-linearly linked
with the atmospheric profiles. Retrievals always need prior background
information, which either comes from independent statistics or from the
short-range forecast. The direct assimilation of radiances has the advantage
to avoid the contamination by such an external background information for
which error characteristics are poorly known. Another advantage of global
variational methods is that increments brought by satellite radiances are
further constrained by many other observations/information. Finally, raw
radiance observations exhibit less spatially correlated errors than processed
retrieved information. In current data assimilation schemes, this allows to
use observations with more spatial density, a subject which will be discussed
further in the next section. Of course this use of raw data comes at the cost
of developing the observation operator and the quality control appropriate
for each observation for each data assimilation system in each NWP centre,
but some of this effort is collaborative through EUMETSAT facilities for
instance.
Zooming now on the period covering the most recent years, Figure 3
shows the number of data used in the ECMWF analysis between 1997 and
2003. This illustrates the tremendous increase in terms of observation
numbers which took place lately, and most of this progression in data
numbers comes from non-conventional asynoptic observations.
Such observation numbers have a significant impact, especially in an
advanced data assimilation scheme such as 4D-Var which has been used
since 1997 (Rabier et al, 2000). 4D-Var stands for Four-Dimensional
Variational Data assimilation and it performs a global optimization of the
model trajectory over a period of 6 to 12 hours typically. It performs an
adjustment of the model trajectory with the observations taken explicitly at
the precise time of the observation, thus allowing for a consistent use of data
spread in time throughout the optimization period, such as satellite
observations. In the linear approximation, 4D-Var is equivalent to a Kalman
smoother: at any time in the assimilation window, information from past and
future observations within this window will be taken into account to provide
the best estimate of the flow (Rabier and Liu, 2003). It can also use the time-