Increasing hyperspectral image classification accuracy for data sets with limited training samples by sample interpolation