Skill assessment of different ensemble generation schemes for retrospective predictions of surface freshwater fluxes on inter and multi-annual timescales

Abstract

The long term variability and its predictability of the monthly mean oceanic surface net freshwater fluxes is compared in a set of retrospective predictions. All are using the same model setup, and only differ in the implemented ocean initialisation method and ensemble generation method. The basic aim is to deduce the differences between the initialization/ensemble generation methods in view of the uncertainty of the verifying observational data sets. The analysis will give an approximation of the uncertainties of the net freshwater fluxes, which up to now appear to be one of the most uncertain products in observational data and model outputs. All ensemble generation methods are implemented into the MPI-ESM earth system model in the framework of the ongoing MiKlip project (www.fona-miklip.de). Hindcast experiments are initialised annually between 2000–2004, and from each start year 8 ensemble members are run for 10 years forward. Four different ensemble generation methods are compared: (i) a method based on the Anomaly Transform method in which the initial oceanic perturbations represent orthogonal and balanced anomaly structures in space and time and between the variables taken from a control run, (ii) one-day-lagged ocean states from the MPI-ESM-LR Baseline 1 system, (iii) one-day-lagged ocean and atmospheric states with preceding full-field nudging to re-analysis in the atmospheric and anomaly nudging in the oceanic component of the system – the Baseline MPI-ESM-LR system, (iv) an Ensemble Kalman Filter (EnKF) implemented into oceanic part of MPI-ESM, assimilating monthly subsurface oceanic temperature and salinity using the Parallel Data Assimilation Framework and full-field nudging in the atmosphere. The hindcasts are evaluated probabilistically using freshwater flux data set from NCEP-R2. On the global scale the physically motivated methods (i) and (iv) provide probabilistic hindcasts to some extent higher correlation and reliability than the lagged initialization methods (ii)/(iii) despite the large uncertainties in the verifying observations and in the simulations. We suggest similar approaches for further evaluations of other variables of decadal hindcasts systems.

Keywords

• surface freshwater flux forecast
• mean climate state statistics
• probabilistic evaluation
• ensemble generation methods