New approaches on dimensionality reduction in hyperspectral images for classification purposes

Abstract

This paper presents a quasi-unsupervised methodology to detect endmembers within an hyperspectral scene and to derive a pixel-wise classification on its basis. The endmember detection step takes as input an overcomplete spectral library, and detects the materials within a scene by analyzing derivative features under the sparsity assumption. The purest pixels for each detected material are then fed to a classifier based on synergetics theory, which is able to produce accurate classification maps on the basis of a restricted training dataset. As the classifier projects the image onto a subspace composed by the classes of interest found in the first step, a focused dimensionality reduction is performed in which every dimension is semantically meaningful.