Ensemble-based global ocean data assimilation

•Conducted an ensemble-based DA study using POP OGCM & DART.•Proposed a new inflation scheme and studied it in a low-order model.•A variety of diagnostics demonstrates improved DA performance.•Assimilated state shows decreased SST error and subsurface temperature errors.•Hindcast skill demonstrated on initialization with assimilated state.

We present results of experiments performing global, ensemble-based, ocean-only data assimilation and assess the utility of such data assimilation in improving model predictions. The POP (Parallel Ocean Program) Ocean General Circulation Model (OGCM) is forced by interannually varying atmospheric fields of version 2 of the Coordinated Ocean Reference Experiment (CORE) data set, and temperature and salinity observations from the World Ocean Database 2009 (WOD09) are assimilated. The assimilation experiments are conducted over a period of about two years starting January 1, 1990 using the framework of the Data Assimilation Research Testbed (DART).We find that an inflation scheme that blends the ensemble-based sample error covariance with a static estimate of ensemble spread is necessary for the assimilations to be effective in the ocean model. We call this Climatology-based Spread Inflation or CSI for short. The effectiveness of the proposed inflation scheme is investigated in a low-order model; a series of experiments in this context demonstrates its effectiveness.Using a number of diagnostics, we show that the resulting assimilated state of ocean circulation is more realistic: In particular, the sea surface temperature (SST) shows reduced errors with respect to an unassimilated SST data set, and the subsurface temperature shows reduced errors with respect to observations. Finally, towards assessing the utility of assimilations for predictions, we show that the use of an assimilated state as initial condition leads to improved hindcast skill over a significant period of time; that is when the OGCM is initialized with an assimilated state and run forward, it is better able to predict unassimilated observations of the WOD09 than a control non-assimilating run (≈ 20% reduction in error) over a period of about three months. The loss of skill beyond this period is conjectured to be due, in part, to model error and prevents an improvement in the representation of variability on longer time-scales.