In order to improve the accuracy of hyperspectral pixel un-mixing, a Kernel based pixel un-mixing method was proposed in this paper. By kernelizing orthogonal subspace projection (OSP) operator, least squares OSP (LSOSP) operator, nonnegative constrained least squares (NCLS) operator and fully constrained least squares (FCLS) operator respectively, the authors established Kernel OSP (KOSP), Kernel LSOSP (KLSOSP), Kernel NCLS (KNCLS) and Kernel FCLS (KFCLS) to hyperspectral imagery pixel un-mixing. The comparison experiments of abundance inversion by using KLSOSP, KNCLS, KFCLS and LSOSP, NCLS, FCLS to CUPRITE AVIRIS data were carried out. The results show that for heavily mixed hyperspectral images, the pixel un-mixing accuracy of Kernels based KLSOSP, KNCLS and KFCLS is higher than that of LSOSP, NCLS and FCLS. Meanwhile, the constraint conditions can improve the accuracy of abundance estimates.