Using spatial and spectral information for improving endmember extraction algorithms in hyperspectral remotely sensed images

Mixing constituent elements at the pixel level occurs as the result of low spatial resolution in hyperspectral images. Spectral signature extraction of such constituent elements which are known as endmembers in each mixed pixel and estimating the abundance maps of such pure spectral signatures are two main roles in Spectral Mixture Analysis(SMA). Most of algorithms, which was proposed for endmember extraction algorithms(EEs), are established on only spectral purity information such as OSP, N-Finder, VCA, PPI, IEA, SGA and neglect the spatial context of image pixels. Recently SPP technique that couples spatial and spectral information was proposed prior spectral-based EEs(SBEEs). It is implemented in a distinct module by correcting the hyperspectral image with no modification of next stage. In this paper, we propose a novel Spatial-Spectral Pre Stage(SSPS) algorithm that benefits from the advantages of SPP algorithm without correcting the original hyperspectral image by identifying spatially homogenous and spectrally pure pixels for the next SBEEs when it considers the fact that endmembers probably exist in such areas. With evaluation and comparison of the proposed SSPS algorithm prior the mentioned SBEEs and SPP, in hyperspectral Aviris Cuprite image, we demonstrated that using the proposed module outperforms the other spectral-based EE and SPP algorithms especially in the case of IEA, which is computationally very complex, with noticeable reduction in processing time and RMSE reconstruction of original image.