Data assimilation using an ensemble of models: a hierarchical approach

Abstract

Abstract. One characteristic of biogeochemical models is uncertainty about their formulation. Data assimilation should take this uncertainty into account. A common approach is to use an ensemble of models. We must assign probabilities not only to the parameters of the models but also to the models themselves. The method of hierarchical modelling allows us to calculate these probabilities. This paper describes the approach, develops the algebra for the most common case and then applies it to the Atmospheric Tracer Transport Model Intercomparison Project (TransCom). We see that the discrimination among models is unrealistically strong, due to optimistic assumptions inherent in the underlying inversion. The weighted ensemble means and variances from the hierarchical approach are quite similar to the conventional values because the best model in the ensemble is also quite close to the ensemble mean. The approach can also be used for cross-validation in which some data are held back to test estimates obtained with the rest. We demonstrate this with a test of the TransCom inversions holding back the airborne data. We see a slight decrease in the tropical sink and a notably different preferred order of models.