More sparsity in hyperspectral SVM classification using unsupervised pre-segmentation in the training phase