Multi-model fusion and uncertainty estimation for ocean prediction

The dissertation presents three main results: a methodology for integrating multiple ocean models into a single ocean prediction system, a bias correction methodology suited for ocean prediction, and an uncertainty estimation methodology for ocean prediction and multi-model fusion. A characteristic feature of the development is that the proposed methodologies are adaptive and designed to work with a small sample of past validation events. Thus, I address the main difficulty of multi-model regional ocean forecasting: non-stationarity of errors in ocean models. The proposed methodology consists of three general steps: (a) systematic error estimation; (b) uncertainty estimation; and (c) Bayesian model fusion. All the estimation procedures within the methodology are designed to be unbiased and minimum error variance. Each of the outlined three steps constitutes a useful stand-alone result, expected to find applications in regional ocean forecasting, as well as in other numerical forecasting applications. In combination, they provide the methodology for integrating multiple ocean models into a single ocean prediction system.