This article reviews the requirements for a data assimilation system from the perspective of initializing seasonal forecasts. It provides a historical perspective of the developments in ocean data assimilation and ocean observing systems. It also discusses the differences between state estimation and initialization, and presents a brief assessment of different initialization strategies. The value of assimilating ocean data to estimate the ocean state and to initialize seasonal forecasts is demonstrated. However, it is also shown that the assumption of unbiased models in conventional data assimilation methods is not suitable for the production of long temporal records of ocean initial states. This is due to the combined effect of model-forcing error and the changing nature of the observing system. Bias correction algorithms are therefore important in the estimation of long records of ocean states. In the equatorial ocean, the delicate balance between the mass and the velocity fields should be preserved in order to maintain realistic circulations. The most common approach for initializing seasonal forecasts is the so-called full uncoupled initialization, which basically consists of producing an ocean reanalysis by assimilating ocean observations into an ocean model driven by atmospheric fluxes. Alternative approaches are the so-called anomaly initialization, which only attempts to initialize the anomalous state without any attempt of correcting mean; the latter is usually conducted in coupled mode, but coupled and anomaly initialization are not synonymous, and there are approaches where the initialization of the full state is done in coupled mode. The relative value of the approaches is system dependent, but as a long-term strategy the full initialization in coupled mode is more promising.
Balmaseda, Magdalena A.. 2017. "Data assimilation for initialization of seasonal forecasts." Journal of Marine Research 75, (3). https://elischolar.library.yale.edu/journal_of_marine_research/436