Applying Multi-Model Superensemble Methods to Global Ocean Operational Systems

A number of international organizations are currently running global Ocean Operational Systems (OOS) in near-real-time modes. The daily nowcast and forecast data sets for a variety of oceanographic prognostic parameters are now available to the public through large data servers. Ongoing studies conducted as part of the GODAE OceanView Class-4 intercomparison (Ryan, et al 2015) are demonstrating that these OOS offer complimentary predictive skills. There is also well-documented literature that shows combining multiple forecasts using simple combinations can help to substantially increase accuracy (or reduce error) of such forecasts (Clemen, 1989, Galmarini et al., 2004). Similar multi-model ensemble methods have been employed with success in atmospheric models for some time, but the sheer size of the global data sets coupled with limited public availability of the data and computational resource limits has been a challenge to applying these techniques to ocean models.

A previous study (Spindler, Mehra, Tolman 2013) employed simple and weighted means and k-means clustering algorithms (Hartigan, 1975; Arthur and Vassilivitski, 2006) to improve nowcast error and bias in SST by processing a month of nowcast fields from five global OOS. This study is an extension of that work into the feasibility of applying simple numerical techniques as well as more sophisticated resampling methods to four global OOS (FOAM, HYCOM, MERCATOR, and RTOFS) that offer near-real-time nowcast and forecast data to assess the potential for reducing error and bias in both the current ocean state and forecasts of a number of oceanographic parameters (SST, SSS, SSH, surface current).