The conflict of high dimensionality and the very limited number of available training samples is one of the problems in the classification of hyperspectral images. At the same time, the redundance between different bands brings trouble to the classification. The ensemble learning provides a new way for solving the problem mentioned above. Based on the correlation between different bands, a band grouping is carried out. By selecting different bands from different groups new subsets of spectral bands is formed. The redundance reduces bands in the new spectral band subsets are independent and used to train the maximum likelihood (ML) classifiers which can be used later for ensembling. The combining of classifiers is done by the simple majority voting and the ensemble classifier is formed. Experimental results of the hyperspectral remotely sensing image demonstrate that the method presented here has an excellent classification result and outpeforms many other methods.