A Markovian generalization of support vector machines for contextual supervised classification of hyperspectral images

Hyperspectral sensors accurately sample the spectral signatures of different land covers, thus allowing an effective discrimination of cover classes or ground materials. However, addressing a supervised classification problem with hundreds of features involves critical small-sample size issues. Moreover, traditional hyperspectral-image classifiers are usually noncontextual. In this paper, a novel method is proposed, that is based on the integration of the support vector machine (SVM) and Markov randomfield (MRF) approachesto classification and is aimed at a rigorous contextual generalization of SVMs. A reformulation of the Markovian minimum-energy rule is introduced and is analytically proven to be equivalent to the application of an SVM in a suitably transformed space. The internal parameters of the method are automatically optimized by extending recently developed techniques based on the Ho-Kashyap and Powell's numerical algorithms and the proposed classifier is also combined with the recently proposed band-extraction approach to feature reduction.