This work presents the application of a novel technique on dimensionality reduction to deal with multispectral images. A distance based on mutual information is used to construct a hierarchical clustering structure. Experimental results show that the method provides a very suitable subset of multispectral bands for pixel classification purposes.