In hyperspectral remote sensing imagery, material usually present two spatial distribution characteristics: one is its dominance in some special areas, another is its consistency on the land surface. By utilizing this two prior information, we propose an algorithm named nonnegative matrix factorization (NMF) with abundance constraint, which introduces both orthogonality and smoothness into abundance. To further improve the algorithm performance, we also propose a new stop criterion and an adjusting method of adapting weight factor to the varying signal-to-noise (SNR) and mixing degree. Experimental results based on synthetic and real hyperspectral data show that our algorithm not only represents material distribution characteristics very well, but also increases the unmixing accuracy. Meanwhile, the algorithm can lead to satisfactory unmixing results under the conditions of low SNR and no pure pixels.