Addressing Uncertainties in Global Ocean Ensembles

The relative merits of forecast systems addressing the impact of uncertainties in ocean ensembles is investigated using a set of model experiments. These experiments are designed to account for sources of uncertainty due to errors in initial conditions and surface boundary conditions. Various methods have been proposed for dealing with the uncertainties in the atmosphere and ocean forecasting ensembles. Here, perturbed initial conditions (P_ic), and observations (P_ob) are used for the generation of initial uncertainties, and perturbed surface forcing (P_sf) for the surface boundary uncertainties. The strategy is to run three experiments introducing each uncertainty at a time to understand the relative impact on the forecast performance. The fourth experiment is aligned with the Navy Earth System Prediction Capability, wherein ensemble forecast are performed with a coupled atmosphere-ocean model through perturbed observations and surface forcing techniques (P_cos). A 20-member ensemble is generated. A host of deterministic and ensemble probabilistic forecast verification methods are applied to quantitatively evaluate the performance of the four ensemble systems in terms of accuracy, skill, reliability, resolution, and discrimination. The forecast period is one month in April 2014. There is a systematic growth in the ensemble spread with addition of each uncertainty thereby reducing the discrepancy between the spread and the error of the ensemble mean, the RMSE. This is an indication for improved statistical reliability, although they are underdispersive (spread < RMSE). Perturbed surface forcing (+P_sf, P_cos) has beneficial impact on the surface variables while perturbed observations improve the reliability of subsurface fields. The improved reliability in P_cos is due to the significant bias reduction and a smaller underdispersion. The resolution and discrimination among the ensembles are very similar, but they differ between the variables. SST has greater resolution and discrimination than SSHA, while both variables are skillful relative to the climatology at all lead times. It is evident from the reliability diagram that the forecast tends to overestimate (underestimate) the observed frequency for the high (low) forecast probabilities. Both variables exhibit good sharpness, although SST has nearly perfect forecast distribution.