Sparsity/accuracy trade-off for vector machine based hyperspectral classification

Sparsity/accuracy trade-off for hyperspectral image classification based on support vector machines (SVMs) and relevance vector machines (RVMs) is proposed in this paper. In the proposed approach K-means or phase correlation based unsupervised segmentation and RANSAC (random sample consencus) with cross-validation is used to provide a compressed hyperspectral data set before RVM and SVM training. These approaches are used to compress the training data by combining similar hyperspectral data samples, as a result of which the number of training samples is reduced, resulting in an overall smaller support vector amount for SVM classification or a smaller relevance vector amount for RVM classification after training. It is possible to trade of accuracy against sparsity with the proposed approach and also provide faster training as well as classification times.