Semisupervised support vector machine classification for hyperspectral imagery

Variety of techniques with the capability of being applied on original data spaces has been developed for hyperspectral (HS) classification. Among them, support vector machine (SVM) presents a high classification accuracy, however, its performance for too small ratios of number of available training samples to number of features is relatively low due to the Hughes effect. This paper proposes a new semisupervised approach through combining appropriate discriminant data transforms such as principal component analysis with SVM to tackle the above mentioned drawback. The experiments on real HS data validate the superiority of the proposed combined approach over the traditional pure SVM technique.