Ensemble Strategies for Classifying Hyperspectral Remote Sensing Data Xavier Ceamanos 1 , Bj¨ orn Waske 2 , J´ on Atli Benediktsson 2 , Jocelyn Chanussot 1 , and Johannes R. Sveinsson 2 1 GIPSA-LAB, Signal & Image Department, Grenoble Institute of Technology, INPG BP 46 - 38402 Saint Martin d’H` eres, France 2 University of Iceland, Faculty of Electrical and Computer Engineering, Hajararhagi 2-6, 107 Reykjavik, Iceland Abstract. The classification of hyperspectral imagery, using multiple classifier systems is discussed and an SVM-based ensemble is introduced. The data set is separated into separate feature subsets using the correla- tion between the different spectral bands as a criterion. Afterwards, each source is classified separately by an SVM classifier. Finally, the different outputs are used as inputs for final decision fusion that is based on an additional SVM classifier. The results using the proposed strategy are compared to classification results achieved by a single SVM and other well known classifier ensembles, such as random forests, boosting and bagging. Keywords: hyperspectral, land cover classification, support vector ma- chines, multiple classifier systems, classifier ensmeble. 1 Introduction Hyperspectral data provide detailed spectral information from land cover, rang- ing from the visible to the short-wave infrared region of the electromagnetic spectrum. Nevertheless the classification of hyperspectral imaging is challeng- ing, due to the high-dimension of the data sets. Particularly with a limited number of training samples the classification accuracy (of conventional statis- tical classifiers) can be limited. Hughes [1] showed that with a limited number of training samples the classification accuracy decreases after a maximum is achieved. Thus, it requires sophisticated classification algorithms to use detailed hyperspectral information comprehensively. In several remote sensing studies it was demonstrated that Support Vector Machines (SVM) perform better than or at least comparable to other classifiers in terms of accuracy, even when applied to hyperspectral data sets [2],[3]. One reason for this success might be the un- derlying concept of SVM classifiers. Their aim is to discriminate two classes by constructing an optimal separating hyperplane to the training samples within a multi-dimensional feature space, by using only the closest training samples of each class [4]. Consequently, the approach only considers training data close to the class boundary and performs well with small training sets. J.A. Benediktsson, J. Kittler, and F. Roli (Eds.): MCS 2009, LNCS 5519, pp. 62–71, 2009. c  Springer-Verlag Berlin Heidelberg 2009