THE ECMWF PERSPECTIVE
369
measure salinity. The diamonds are ARGO float measurements from buoys
which drift at ~1000m, but every 10 days pop up to the surface, measuring
temperature and often salinity, and relay the information via satellite to a
ground station where it is put on the GTS. The mooring and XBT data are
also available in real-time. If there are many data points in close proximity
in space and time, they can be ‘super-obbed’, a process by which they are
combined into one ‘super-observation’ and given an increased weight. See
Smith et al 1991 for discussion on super-obbing and quality control issues.
3.3 The ocean analysis system
The scheme currently in operational use is OI (Optimal Interpolation)
using a time window of 10 days. FGAT has not been implemented in this
scheme: rather all observations in the 10 day window are taken to apply at
the centre of the window. It differs from a standard OI, however, in that the
correction to the FG following an OI is not applied instantaneously; rather
the increment is divided by the number of timesteps in a 10 day window and
then that fraction of the increment is applied every timestep. The idea
behind this was to allow the model to generate its own circulation field
following an OI in which only thermal data were assimilated. Although this
worked to some degree, a better circulation field can be produced. In
principle, one can update the velocity field and salinity field even though
only temperature data are being assimilated. This can be done through
having multi-variate covariances. However, at the time of implementation
we did not have these well tuned. Rather, we sought to improve the
univariate assimilation of T by building in some corrections, done in
physical space rather than through multi-variate covariances. So a
geostrophic velocity field is calculated following the T-assimilation and then
the velocity increments are also spread over 10 days. See Burgers et al 2002
and Balmaseda 2003 for more details.
Salinity is also corrected by applying an S correction such that the T-S
relationship is preserved. In the ocean when T changes, so does S in such a
way as to preserve the T-S relationship. There are regions in which this
approximation doesn t work- for example in the surface layers where heat
and fresh-water fluxes will change T-S. The true T-S relationship is not
known at any time, so we use the model T-S. Unfortunately the model T-S
can drift from truth; for example, if the model has too much or too little
mixing or if the precipitation is wrong and the surface layer starts to interact
with the layers beneath. Nonetheless, although not a perfect solution, it does
give beneficial impact, especially in the equatorial Atlantic and Pacific. See
Troccoli et al 2002 for further details on the strategy and its impacts.
A novel feature of the ocean analysis system is that not just a single
analysis but multiple analyses are performed. The purpose of the analysis is
to provide initial conditions for monthly and seasonal forecasts. Such
forecasts must be probabilistic. This implies that an ensemble of forecasts
’
369
measure salinity. The diamonds are ARGO float measurements from buoys
which drift at ~1000m, but every 10 days pop up to the surface, measuring
temperature and often salinity, and relay the information via satellite to a
ground station where it is put on the GTS. The mooring and XBT data are
also available in real-time. If there are many data points in close proximity
in space and time, they can be ‘super-obbed’, a process by which they are
combined into one ‘super-observation’ and given an increased weight. See
Smith et al 1991 for discussion on super-obbing and quality control issues.
3.3 The ocean analysis system
The scheme currently in operational use is OI (Optimal Interpolation)
using a time window of 10 days. FGAT has not been implemented in this
scheme: rather all observations in the 10 day window are taken to apply at
the centre of the window. It differs from a standard OI, however, in that the
correction to the FG following an OI is not applied instantaneously; rather
the increment is divided by the number of timesteps in a 10 day window and
then that fraction of the increment is applied every timestep. The idea
behind this was to allow the model to generate its own circulation field
following an OI in which only thermal data were assimilated. Although this
worked to some degree, a better circulation field can be produced. In
principle, one can update the velocity field and salinity field even though
only temperature data are being assimilated. This can be done through
having multi-variate covariances. However, at the time of implementation
we did not have these well tuned. Rather, we sought to improve the
univariate assimilation of T by building in some corrections, done in
physical space rather than through multi-variate covariances. So a
geostrophic velocity field is calculated following the T-assimilation and then
the velocity increments are also spread over 10 days. See Burgers et al 2002
and Balmaseda 2003 for more details.
Salinity is also corrected by applying an S correction such that the T-S
relationship is preserved. In the ocean when T changes, so does S in such a
way as to preserve the T-S relationship. There are regions in which this
approximation doesn t work- for example in the surface layers where heat
and fresh-water fluxes will change T-S. The true T-S relationship is not
known at any time, so we use the model T-S. Unfortunately the model T-S
can drift from truth; for example, if the model has too much or too little
mixing or if the precipitation is wrong and the surface layer starts to interact
with the layers beneath. Nonetheless, although not a perfect solution, it does
give beneficial impact, especially in the equatorial Atlantic and Pacific. See
Troccoli et al 2002 for further details on the strategy and its impacts.
A novel feature of the ocean analysis system is that not just a single
analysis but multiple analyses are performed. The purpose of the analysis is
to provide initial conditions for monthly and seasonal forecasts. Such
forecasts must be probabilistic. This implies that an ensemble of forecasts
’