Multiple Instance Bagging Approach for Ensemble Learning Methods on Hyperspectral Images

Abstract

In this work, a novel ensemble learning (EnLe) method is proposed for hyperspectral images by the motivation of bagging method in the multiple instance (MI) learning (MIL) algorithms. Ensemble based bagging is made by using training samples in the hyperspectral scene and multiple instance bags are created by defining local variable windows upon selected instances. A naive classification method used in the multi-instance learning areas is adopted and applied to ROSIS-03 Pavia University hyperspectral image. Obtained classification results are presented along with the results of single classifiers and the results of the state of the art EnLe methods comparatively.