A feature extraction method based on spectral segmentation and integration of hyperspectral images

Abstract

In response to the curse of dimensionality in hyperspectral images (HSIs), to date, numerous dimensionality reduction methods have been proposed among which the feature extraction (FE) methods are of particular interest. This paper introduces a new supervised pixel-based FE called spectral segmentation and integration (SSI). In SSI, the spectral signature curve (SSC) of the pixels are identically divided into some non-overlapping segments, called channels. The existing bands in each channel are then integrated using a mean-weighted operator, leading to some new features in a very lower number than the original bands. SSI applies a particle swarm optimization (PSO) algorithm to globally search and locate the optimum positions and widths of the channels. For the sake of evaluation and comparison, the features provided by the proposed SSI method were applied to the well-known SVM classifier. The results were compared to not only a most recent pixel-based FE method, namely, spectral region splitting but also six conventional FE methods, including nonparametric weighted feature extraction, decision boundaries feature extraction, clustering-based feature extraction, semi-supervised local discriminant analysis, band correlation clustering and principal component analysis. Experimental results, obtained on two HSIs, proved the superiority of the proposed SSI.