Abstract: Because spatial resolution of hyper-spectral images is low, a large number of the mixed pixels were in hyper-spectral images. The presence of the mixed pixels is one of the main reasons of the low accuracy of target classification in hyper-spectral images. Hyper spectral pixel unmixing is of great importance in hyper-spectral remote sensing image processing. Hyper-spectral unmixing is divided into two methods:linear and nonlinear spectral unmixing. Linear spectral unmixing has been studied most widely. Two steps of linear spectral mixing are summed up:firstly, spectral signals of ground objects in the pure pixels are extracted, that is, end-members are extracted. It is the key step. The weighted linear combination of end-members is used to unmix the spectral image of the mixed pixels, that is, the abundance inversion. Main progress of end-member extraction and abundances inversion is briefly introduced, and several typical algorithms for end-member extraction are introduced. Through summing-up, contrasting and analyzing, the characteristics of different endmember extraction methods are summarized. The prospect of hyperspectral unmixing is prospected.