Relevance Vector Machine for Efficient Classification of Scattered Patterns in Hyperspectral Imagery

Analysis of hyperspectral data for defining the land-cover classes through classification techniques, in particular for small patches and scattered land-covers, is not a trivial task. Factors such as high spatial variability of landcover signatures, the “boundary effect” between neighboring land-covers, and the curse of dimensionality make this task more challenging [1]. As the integrity of a land-cover class decreases in an image, i.e. it becomes more scattered and distributed in smaller segments, its heterogeneity increases due to the presence of more mixed pixels (stronger “boarder effect”) and as a consequence the classification accuracy decreases [2]. Variety of techniques has been applied to supervised classification of remotely sensed hyperspectral imagery to deal with the classification accuracy and curse of dimensionality issue in classification of hyperspectral imagery in last decade. Among them feature reduction techniques [3], adaptive statistics estimation by exploitation of classified (semilabeled) samples [4], regularization of the sample covariance matrix [5], analysis of the spectral signatures to model the classes [6], and support vector machines [7] can be referred as the main categories of approaches. Although these approaches have obtained many achievements, they seldom take the real situation of presence of small land-cover patches or insufficiency of available training samples into account. This paper proposes a new efficient classification approach to tackle the problems of complexity and accuracy, in particular for small and scattered land-covers, through relevant vector machine (RVM).