Ocean data assimilation as a component of a climate forecast system

Abstract

The El Niño Southern Oscillation (ENSO) phenomena is the major source of interannual climatic variability in the Tropics. It results primarily from ocean-atmosphere interactions in the tropical Pacific. Over the past decade as part of the Tropical Oceans Global Atmosphere Experiment (TOGA) considerable progress was made in implementing observing systems to document this variability, developing a hierarchy of models, statistical as well as dynamical, to study its physics, and to implement routine experimental forecasts for aspects of ENSO related variability. During the past decade at the National Meteorological Center (NMC), presently the National Centers for Environmental Prediction (NCEP), a unified system for seasonal climate prediction was developed. This consisted of the routine assimilation of the in situ thermal data sets collected by TOGA into an ocean general circulation model to provide analyses for real time climate diagnostics and to provide ocean initial conditions for forecasts and a coupled ocean—atmosphere general circulation forecast model. Conceptually similar systems are currently being implemented at a number of other Centers internationally. A basic requirement for climate diagnostics and prediction is the best definition of the state of the ocean. In the Tropics where the ocean is strongly and directly forced, a model simulation forced with observed stress fields combined with in situ observations through data assimilation, can give a good estimate. These modelbased analyses can provide the basis for diagnostic studies, verification of model simulations and forecasts, and the initial conditions for the forecasts. Comparisons of simulations using existing wind stress products and models to analyses produced using data assimilation show large differences indicating that models and stress fields can still be improved. Without data assimilation model simulations contain significant errors both in their mean spatial structure and also in their low frequency variability. The thermocline topography in the mean is too weak, especially south of the equator where the subtropical gyre is not well defined. Experiments with several different wind products suggest that this is more a result of model rather than forcing field errors. Simulations without data assimilation are also unable to capture the full amplitude and structure of the low frequency variations associated with El Niño. Data assimilation can overcome many of these deficiencies. Even with assimilation, incremental improvements in analysis accuracy are further achieved when better wind forcing is used. However, large corrections can also alter strict dynamical balances. One impact of this is that the near equatorial currents in the western Pacific in NCEP's model-based analyses appear unrealistic. An improved estimation of the low frequency variability of the ocean should lead to higher skill levels in forecasts. This appears to be the case but the results are seasonally dependent. Forecasts initiated from late spring to fall for two versions of the NCEP forecast model show improved skill when data assimilation is used to derive the initial conditions. However, little positive impact is found for forecasts initiated in the winter months. If data assimilation is needed to correct for large errors, then in the forecast mode, where assimilation is not possible, the corrections, especially to the mean field, can lead to large systematic forecast errors. Future skill improvements will result from improvements in the forcing fields and ocean model used in the initialization, and improvements to the coupled forecast model. Indications from experiments at NCEP are that the largest impact on forecast skill is from improvements in the coupled ocean-atmosphere model used in the forecasts. The central role of data assimilation is in producing the best analyses that can be used for improving the ocean models and forcing fields.