A Comparison of Two Ensemble Generation Methods Using Oceanic Singular Vectors and Atmospheric Lagged Initialization for Decadal Climate Prediction

The sensitivity of ensemble spread and forecast skill scores of decadal predictions to details of the ensemble generation is investigated by incorporating uncertainties of ocean initial conditions using ocean singular-vector-based (OSV) perturbations. Results are compared to a traditional atmospheric lagged initialization (ALI) method. Both sets of experiments are performed using the coupled MPI-ESM model initialized from the GECCO2 ocean synthesis. The OSVs are calculated from a linear inverse model based on a historical MPI-ESM run. During the first three lead years, the sea surface temperature spread from ALI hindcasts appears to be strongly underestimated, while OSV hindcasts show a more realistic spread. However, for later lead times (the second pentad of hindcasts), the spread becomes overestimated for large areas of the ocean in both ensembles. Yet, for integrated measures such as the North Atlantic SST and Atlantic meridional overturning circulation, the spread of OSV hindcasts is overestimated at initial time and reduces over time. The spread reliability measures are shown to be sensitive to the choice of the verification dataset. In this context, it is found that HadISST tends to underestimate the variability of SST as compared to Reynolds SST and satellite observations. In terms of forecast skill for surface air temperature, SST, and ocean heat content, OSV hindcasts show improvement over ALI hindcasts over the North Atlantic Ocean up to lead year 5.