Abstract: The existence of synonyms spectrum phenomenon and the poor sample representativeness and parameter setting in the classification process result in the unstable classification result and poor classification accuracy of hyperspectral images. In order to solve this problem, a novel optimal representative vectors classification algorithm based on sample set optimization was proposed for glacier classification. This algorithm purified original samples in the ROI through an improved density peak clustering method, and selected the optimal representative vectors by clustering the mean vector set of optimized samples as the central vector of each object. Through experimental verification, this algorithm can effectively improve the classification accuracy of the glacier to 90%; it shows that the optimal representative vectors classification method can eliminate the impact of sample differences and improve glacier classification accuracy.