Rapid Assessment ofthe Coastal Ocean Environment
209
phases, a descriptive phase, a dynamical phase, and a predictive phase (Robinson et
al., 1996). In the descriptive phase the relevant circulation structures, their time
and space scales and their variabilities need to be identified. Concomitant1y, the
forecast models must be validated, i.e., shown to be applicable to the relevant
structures (e.g., a barotropic model is not appropriate for baroclinic structures) and
observational requirements must be established. The extension to interdisciplinary
phenomena is direct. In the dynamical phase, the specific dynamical processes
responsible for i) the evolution of the circulation and its variabilities, and ii) for
synoptical dynamical events must be determined. In this phase the models are calibrated, i.e., physical, domain and computational parameters tuned to the region
and its phenomena by sensitivity analyses (Lermusiaux et al., 2000) and measurement requirements refined. Although forecasts are carried out in every phase, the
predictive phase is devoted to forecast model and assimilation scheme verification.
Real time forecasts with high quality data sets for initialization, assimilation and
verification, obtained by oversampling, are required (e.g., Robinson et al., 1996).
These data sets can then be used to design an optimally efficient observational network component for the regional forecast system.
The predictive capability ofthe regional system needs to be evaluated both qua1itatively and quantitatively. Regional dominant variabilities need to be defined and
used to characterize the forecast. Examples include: the number of branches
present in a coastal current that may bifurcate or trifurcate; the existence or not of
transient vortices; the location and shape of a permanent meander. Additionally,
appropriate quantitative skill metrics need to be defined such as root mean square
errors and pattern correlations (Miller et al., 1995; Robinson et al., 1996). The
application of such metrics must generally allow for phase errors, i.e., errors in the
location or timing of events. Predictive capability without data assimi1ation is ultimately limited by loss of predictability (Ehrendorfer, 1997; Goswami and Shukla,
1991; Houghton, 1991; Latif et al., 1998). Small differences in initial conditions
after some time lead to reasonable but completely different synoptic states. In geophysical fluid dynamical systems, small-scale initial errors are non-linearly transferred to sca1es of operational interest. Error attribution (initial or boundary
conditions, atmospheric forcing, model deficiencies) is important. It is desirab1e to
forecast error fie1ds as weB as state variables, as is done, for examp1e, in the ESSE
method introduced in section 11.6.2 (Lermusiaux and Robinson, 1999). Automated objective adaptive sampling error minimization metrics can then be related
to the quantitative forecast skill metrics (Robinson and Glenn, 1999). The topics
discussed in this paragraph require extensive research efforts.
11.5.2 REA System Validation Issues
The sum of differences between validation measurements and model predictions
has only limited value as a quantitative measure for predictive capability for operational applications. Capability assessment should comply with the requirements of
the end user of the prediction, who is affected by certain deviations much more
than by others. The end user cannot be convinced by a small rms error if erroneous
209
phases, a descriptive phase, a dynamical phase, and a predictive phase (Robinson et
al., 1996). In the descriptive phase the relevant circulation structures, their time
and space scales and their variabilities need to be identified. Concomitant1y, the
forecast models must be validated, i.e., shown to be applicable to the relevant
structures (e.g., a barotropic model is not appropriate for baroclinic structures) and
observational requirements must be established. The extension to interdisciplinary
phenomena is direct. In the dynamical phase, the specific dynamical processes
responsible for i) the evolution of the circulation and its variabilities, and ii) for
synoptical dynamical events must be determined. In this phase the models are calibrated, i.e., physical, domain and computational parameters tuned to the region
and its phenomena by sensitivity analyses (Lermusiaux et al., 2000) and measurement requirements refined. Although forecasts are carried out in every phase, the
predictive phase is devoted to forecast model and assimilation scheme verification.
Real time forecasts with high quality data sets for initialization, assimilation and
verification, obtained by oversampling, are required (e.g., Robinson et al., 1996).
These data sets can then be used to design an optimally efficient observational network component for the regional forecast system.
The predictive capability ofthe regional system needs to be evaluated both qua1itatively and quantitatively. Regional dominant variabilities need to be defined and
used to characterize the forecast. Examples include: the number of branches
present in a coastal current that may bifurcate or trifurcate; the existence or not of
transient vortices; the location and shape of a permanent meander. Additionally,
appropriate quantitative skill metrics need to be defined such as root mean square
errors and pattern correlations (Miller et al., 1995; Robinson et al., 1996). The
application of such metrics must generally allow for phase errors, i.e., errors in the
location or timing of events. Predictive capability without data assimi1ation is ultimately limited by loss of predictability (Ehrendorfer, 1997; Goswami and Shukla,
1991; Houghton, 1991; Latif et al., 1998). Small differences in initial conditions
after some time lead to reasonable but completely different synoptic states. In geophysical fluid dynamical systems, small-scale initial errors are non-linearly transferred to sca1es of operational interest. Error attribution (initial or boundary
conditions, atmospheric forcing, model deficiencies) is important. It is desirab1e to
forecast error fie1ds as weB as state variables, as is done, for examp1e, in the ESSE
method introduced in section 11.6.2 (Lermusiaux and Robinson, 1999). Automated objective adaptive sampling error minimization metrics can then be related
to the quantitative forecast skill metrics (Robinson and Glenn, 1999). The topics
discussed in this paragraph require extensive research efforts.
11.5.2 REA System Validation Issues
The sum of differences between validation measurements and model predictions
has only limited value as a quantitative measure for predictive capability for operational applications. Capability assessment should comply with the requirements of
the end user of the prediction, who is affected by certain deviations much more
than by others. The end user cannot be convinced by a small rms error if erroneous