An oblique subspace projection approach for mixed pixel classification in hyperspectral images

Abstract Recently oblique projection has been studied for many applications in signal processing. In this paper, the concept of oblique projection is applied to develop an algorithm for hyperspectral image classification. Compared with the orthogonal subspace projector ( OSP ), it can be found that OSP is a priori classifier but the oblique subspace projection classifier will be referred to a posterior. As a consequence, the oblique subspace projector ( OBP ) can be thought of as a generalized classifier including OSP . Furthermore, the estimation error from the OBP can be evaluated by applying the Neyman–Pearson detection theory to the corresponding receiver operating characteristic (ROC) curve so the accuracy of the classification can be calculated thereafter. Finally, some computer simulations using real airborne visible infrared image spectrometer (AVIRIS) data are accomplished to justify and compare the effectiveness of the above algorithms.