The Hyperspectral image (HSI) contains several hundred bands of the same region called the Ground Truth (GT). The bands are taken in juxtaposed frequencies, but some of them are noisily measured or contain no information. For the classification, the selection of bands, affects significantly the results of classification, in fact, using a subset of relevant bands, these results can be better than those obtained using all bands, from which the need to reduce the dimensionality of the HSI. In this paper, a categorization of dimensionality reduction methods, according to the generation process, is presented. Furthermore, we reproduce an algorithm based on mutual information (MI) to reduce dimensionality by features selection and we introduce an algorithm using mutual information and homogeneity. The two schemas are a filter strategy. Finally, to validate this, we consider the case study AVIRIS HSI 92AV3C. Keywords: Hyperspectrale images; classification; features selection; mutual information; homogeneity