One of the prominent problems in hyperspectral remote sensing is the existing of mixed pixel widely. How to effectively interpret mixed pixels is an important problem of hyperspectral remote sensing applications. It is not only a problem of mixed pixels effects identification and classification precision of objects, but also a major barrier for the development of remote sensing technology. Mixed pixel decomposition, which is the most effective method to solve the mixed pixel problem, can break through the limitation of spatial resolution. Aiming to the shortcoming of the traditional algorithm of mixed pixel decomposition, an improved method of mixed pixels was put forward, which can take account of the spatial correlation of spectral information and spectral information, and multi-core parallel processing method to raise its efficiency. The endmembers were automatically extracted, and the abundance charts corresponding to each endmember were obtained at the same time. The performance of the proposed algorithm was verified by using actual hyperspectral image. The experimental results on simulated and real hyperspectral image demonstrated that the proposed algorithm can overcome the shortcomings of traditional method and obtain more accurate endmembers and corresponding abundance, which can provide a strong support for urban object classification.