An AdaBoost Ensemble Classifier System for Classifying Hyperspectral Data

This paper presents a new multiple classifier system based on AdaBoost to overcome the high dimensionality problem of hyperspectral data. The hyperspectral data are )rst split into a number of band clusters based on the similarities between the contiguous bands, and each band group is considered as an independent data source. The redundant bands in each cluster are then removed using branch and bound technique. Next, a support vector machine (SVM) is applied to each cluster and the outputs are combined using the weights calculated in AdaBoost iterations. Experimental results with AVIRIS and ROSIS datasets clearly demonstrate the superiority of the proposed algorithm in both overall and single class accuracies when compared to other multiple classi)er systems. For AVIRIS data, which contains classes with greater complexity and fewer available training samples, the differences between the overall accuracies of the AdaBoost results are signi)cantly higher compared to those of the other methods, and more pronounced than for the other dataset. In terms of class accuracies, the proposed AdaBoost approach also outperforms other methods in most of the classes.