Towards resource-frugal deep convolutional neural networks for hyperspectral image segmentation