Improved hyperspectral land-cover analysis using relevance vector machine

In land-cover analysis of hyperspectral remotely sensed images through supervised classification methods, a frequent problem is that the available training samples corresponding to different land-covers are not sufficient. This problem is especially severe for small land-covers and targets which often include the key information of the scene. Furthermore, degrading due to “boundary effect” is more serious in classification of small and scattered patches of land-covers. In this paper, a new supervised hyperspectral classification method through application of a discriminant data transformation in joint with a Bayesian learning-based probabilistic sparse kernel model, i.e., relevance vector machine (RVM), is proposed. The proposed method outperforms other efficient approaches in terms of classification accuracy, robustness to Hughes phenomenon (lack of accuracy due to too small ratio of training sample number to feature number), and computational complexity in particular for small and scattered land-cover classes which are harder to be precisely classified.