370
DAVID ANDERSON ET AL
must be made. In the case of the monthly forecast system, the ensemble size
is 51, while for the seasonal forecast system, it is 40. The ensemble is there
to sample uncertainty arising from the chaotic nature of the atmosphere.
However, it should also take into account uncertainty in the ocean initial
conditions. One method of representing this uncertainty is through running
an ensemble of ocean analyses. In our case the ensemble size is 5. This
ensemble is not to be confused with the EnKF in which the size might be
~100. Experiments are underway to assess the EnKF strategy as part of the
EU project ENACT but results will not be presented here.
In the case of the atmosphere, almost all the information on which an
analysis is based comes from observations of the atmosphere. In the case of
the ocean, a substantial amount of information on the ocean state can be
obtained not through ocean observations but through atmospheric
observations - in fact all the observations that are involved directly or
indirectly in defining the surface wind, heat and fresh water fluxes. For
seasonal forecasting the most important of the surface forcings is the wind.
The wind field that is used to force the ocean has uncertainty. We estimate
that uncertainty and then force 5 ocean analysis streams with wind fields that
are perturbed commensurate with the estimated uncertainty in the wind. In
addition, the SST field is not known sufficiently accurately either. So
strategies.
Although the data coverage might look reasonable in fig 3 where the
symbols are quite large, it is probably barely adequate even for today. Ten
or twenty years ago the coverage was much worse. It is thanks in large
measure to a major international programme called TOGA (Tropical Ocean
Global Atmosphere) that the real-time coverage is as good as it is. Starting
in 1985, this programme steered oceanography towards a free exchange of
data in near real-time.
3.4 The value of data assimilation
There are surprisingly few clean sets of experiments to show that data
assimilation improves the skill of seasonal forecasts. Any results probably
apply only to the system being tested, as improvements in either the ocean
model or the forcing fields through improved atmospheric analyses or
reanalyses, could change the results. One clean set of experiments was
performed by Alves et al 2004. Four sets of analyses were performed and
four ensembles of forecasts were run from these analyses. Two different
wind products were used and for each, experiments with and without data
assimilation were performed. Fig 4 shows the growth of error (upper panel)
and the anomaly correlation (lower panel) from these forecasts in the
tropical Pacific, a key region for seasonal forecasts. The forecasts based on
analyses with data assimilation are clearly better than those without: the rms
perturbations to it are also applied. This is discussed more fully in Vialard
et al 2005, who show the spread generated by different ensemble generation
DAVID ANDERSON ET AL
must be made. In the case of the monthly forecast system, the ensemble size
is 51, while for the seasonal forecast system, it is 40. The ensemble is there
to sample uncertainty arising from the chaotic nature of the atmosphere.
However, it should also take into account uncertainty in the ocean initial
conditions. One method of representing this uncertainty is through running
an ensemble of ocean analyses. In our case the ensemble size is 5. This
ensemble is not to be confused with the EnKF in which the size might be
~100. Experiments are underway to assess the EnKF strategy as part of the
EU project ENACT but results will not be presented here.
In the case of the atmosphere, almost all the information on which an
analysis is based comes from observations of the atmosphere. In the case of
the ocean, a substantial amount of information on the ocean state can be
obtained not through ocean observations but through atmospheric
observations - in fact all the observations that are involved directly or
indirectly in defining the surface wind, heat and fresh water fluxes. For
seasonal forecasting the most important of the surface forcings is the wind.
The wind field that is used to force the ocean has uncertainty. We estimate
that uncertainty and then force 5 ocean analysis streams with wind fields that
are perturbed commensurate with the estimated uncertainty in the wind. In
addition, the SST field is not known sufficiently accurately either. So
strategies.
Although the data coverage might look reasonable in fig 3 where the
symbols are quite large, it is probably barely adequate even for today. Ten
or twenty years ago the coverage was much worse. It is thanks in large
measure to a major international programme called TOGA (Tropical Ocean
Global Atmosphere) that the real-time coverage is as good as it is. Starting
in 1985, this programme steered oceanography towards a free exchange of
data in near real-time.
3.4 The value of data assimilation
There are surprisingly few clean sets of experiments to show that data
assimilation improves the skill of seasonal forecasts. Any results probably
apply only to the system being tested, as improvements in either the ocean
model or the forcing fields through improved atmospheric analyses or
reanalyses, could change the results. One clean set of experiments was
performed by Alves et al 2004. Four sets of analyses were performed and
four ensembles of forecasts were run from these analyses. Two different
wind products were used and for each, experiments with and without data
assimilation were performed. Fig 4 shows the growth of error (upper panel)
and the anomaly correlation (lower panel) from these forecasts in the
tropical Pacific, a key region for seasonal forecasts. The forecasts based on
analyses with data assimilation are clearly better than those without: the rms
perturbations to it are also applied. This is discussed more fully in Vialard
et al 2005, who show the spread generated by different ensemble generation