To deal with the problems of limited samples, high dimension and non-linear in hyperspectral imagery classification, two new techniques, support vector machine (SVM) and relevance vector machine (RVM), are researched in this paper. Similarities and differences are compared and analyzed between SVM and RVM in hyperspectral imagery classification theoretically and experimentally. By simulations and experiments on classification accuracy, computational cost and sparsity, the results show that RVM model is sparser compared with SVM, which makes its test time much shorter, and thus more suitable for online testing of large amount of hyperspectral data. However, the main drawback of RVM is that its classification accuracy is slightly lower than that of SVM. To improve this performance, Fisher linear discriminant analysis (FLDA) is utilized before classification as a pre-processing to make transformation of hyperspectral data. In this way, not only the dimension of image is reduced, but also the classification accuracy of RVM is improved, especially in small land-cover patches, which makes the application of RVM more widely.