Blind unmixing of remote sensing data with some pure pixels: Extension and comparison of spatial methods exploiting sparsity and nonnegativity properties

Multispectral and hyperspectral imaging systems are among the most powerful tools in the field of remote sensing. In remote sensing imagery, pixel values are often linear mixtures of contributions from pure materials contained in the observed scene. In this paper, we extend our recently developed spatial methods for blindly unmixing each pixel of remote sensing data with some pure pixels and we compare their performance, both for multispectral and hyperspectral images. These extended methods are related to the blind source separation (BSS) problem, and are based on sparse component analysis (SCA) and nonnegativity constraints. Spatial correlation-based or variance-based SCA algorithms (which detect a few pure-pixel zones) are firstly used to identify the mixing matrix by means of two different approaches for selecting the columns of this matrix. Nonnegative least squares (NLS) or nonnegative matrix factorization (NMF) methods are then used to extract spatial sources. Experiments based on realistic synthetic data are performed to compare the accuracies and the computational costs of these extended methods. We show that the tested methods yield high accuracy with low computational cost for the variance-based methods as compared to those based on correlation.