Independent component analysis-based band selection techniques for hyperspectral images analysis

Abstract

Extended multiattribute profiles (EMAPs) are morphological profiles built on the extracted features of a hyperspectral image. These profiles proved, when used in a hyperspectral image classification task, their ability to combine the spectral and spatial information offered by this type of data. We propose building EMAPs on the features selected from a hyperspectral image. To do so, three band selection techniques are proposed. The first one is a modified version of the existent independent component analysis (ICA)-based band selection. The other two are based on the initialization-driven ICA. To test the effectiveness of the aforementioned feature selection methods, we used them to build the EMAPs of hyperspectral images; then, the generated profiles served as the input of two hyperspectral image analysis tasks: a hyperspectral image classification task-based on the sparse representation of EMAPs and an EMAP-based change detection technique that we are proposing in this paper.