Low-dimensional enhanced superpixel representation with homogeneity testing for unmixing of hyperspectral imagery

Abstract

Recently superpixel-based approaches have been proposed for dimensionality reduction (DR) in hyperspectral images. The basic assumption of this approach is that the superpixel over-segmentation segments the image into small homogeneous areas. A low-dimensional (LD) image representation is obtained by using the average of the superpixels, which are then used in other image processing tasks up the processing chain. Due to superpixel-segmentation algorithm limitations, the region inside a superpixel may not be homogeneous. Therefore, the average may not be an adequate representation for the superpixel, leading to inaccuracies in the low dimensional representation resulting in errors in the image processing tasks or analysis. Here we present an enhanced superpixel-based dimensionality reduction approach that incorporates homogeneity testing of superpixels. Homogeneous superpixels are represented by their mean but heterogeneous superpixels are represented by multiple representative signatures selected using the SVDSS column subset selection algorithm. The representative signatures for the homogeneous and heterogeneous superpixels provide an improved low-dimensional representation for the hyperspectral image that better captures the image structure. We present experiments applying the proposed enhanced and the conventional superpixel dimensionality reduction approaches to unmixing using the constrained non-negative matrix factorization (cNMF). A subset of data from the Washington DC Mall HYDICE image is utilized. In the experiments, the enhanced superpixel-based dimensionality-reduction approach results in better unmixing results compared to the conventional approach and to unmixing using the full data set.