Fusion trees for fast and accurate classification of hyperspectral data with ensembles of γ-divergence-based RBF networks

Ensembles of RBF networks trained with $$\gamma$$ ? -divergence-based similarity measures can improve classification accuracy of hyperspectral imaging data significantly compared to any single RBF network as well as to RBF ensembles based on the Euclidian distance. So far, the drawback of using classifier ensembles is the need to compute the results of a typically large number of RBF networks prior to combination. In this paper, a modified approach to the fusion of classifier outputs is proposed which is based on decision trees. It is shown for several real-world datasets that a small subset of the RBF networks contributes to the decisions in the average case. Hence, for any decision, a conditional computation of required RBF network outputs yields a significant decrease in the computational costs. Additionally, a selection scheme for subsets of RBF classifiers based on their relevance in the fusion process is proposed. This alternative approach can be used, if analysis requires fixed settings, e.g., to meet time constraints.