In the most applications in remote sensing, there is no need to use all of available data, such as using all of bands in hyperspectral images. In this paper, a new band selection method was proposed to deal with the large number of hyperspectral images bands. We proposed a Continuous Genetic Algorithm (CGA) to achieve the best subset of hyperspectral images bands, without decreasing Overall Accuracy (OA) index in classification. In the proposed CGA, a multi-class SVM was used as a classifier. Comparing results achieved by the CGA with those achieved by the Binary GA (BGA) shows better performances in the proposed CGA method. At the end, 56 bands were selected as the best bands for classification with OA of 78.5 %.