Independent Data Model Selection for Ensemble Dispersion Forecasting

This work aims at introducing an approach to analyze the independence between different data model in a multi-model ensemble context. The models belong to operational long-range transport and dispersion models, but they are also used for the real-time simulation of pollutant dispersion or the accidental release of radioactive nuclides in the atmosphere. In order to compare models, an approach based on the hierarchical agglomeration of distributions of predicted radionuclide concentrations is proposed. We use two different similarity measures: Negentropy information and Kullback-Leibler divergence. These approaches are used to analyze the data obtained during the ETEX-1 exercise, and we show how to exploit these approaches to select subsets of independent models whose performance is comparable to those from the whole ensemble.