Learning Discriminative Embedding for Hyperspectral Image Clustering Based on Set-to-Set and Sample-to-Sample Distances.