Ocean ensemble forecasting. Part I:Ensemble Mediterranean winds from a Bayesian hierarchical model

A Bayesian hierarchical model (BHM) is developed to estimate surface vector wind (SVW) fields and associated uncertainties over the Mediterranean Sea. The BHM–SVW incorporates data-stage inputs from analyses and forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF) and SVW retrievals from the QuikSCAT data record. The process-model stage of the BHM–SVW is based on a Rayleigh friction equation model for surface winds. Dynamical interpretations of posterior distributions of the BHM–SVW parameters are discussed. Ten realizations from the posterior distribution of the BHM–SVW are used to force the data-assimilation step of an experimental ensemble ocean forecast system for the Mediterranean Sea in order to create a set of ensemble initial conditions. The sequential data-assimilation method of the Mediterranean forecast system (MFS) is adapted to the ensemble implementation. Analyses of sample ensemble initial conditions for a single data-assimilation period in MFS are presented to demonstrate the multivariate impact of the BHM–SVW ensemble generation methodology. Ensemble initial-condition spread is quantified by computing standard deviations of ocean state variable fields over the ten ensemble members. The methodological findings in this article are of two kinds. From the perspective of statistical modelling, the process-model development is more closely related to physical balances than in previous work with models for the SVW. From the ocean forecast perspective, the generation of ocean ensemble initial conditions via BHM is shown to be practical for operational implementation in an ensemble ocean forecast system. Phenomenologically, ensemble spread generated via BHM–SVW occurs on ocean mesoscale time- and space-scales, in close association with strong synoptic-scale wind-forcing events. A companion article describes the impacts of the BHM–SVW ensemble method on the ocean forecast in comparisons with more traditional ensemble methods.