中文摘要：

由于利用非负矩阵分解方法解决高光谱解混问题时,标准非负矩阵分解目标函数的非凸性影响了最优解的获取.通过对高光谱图像的端元光谱和空间分布特性的分析,提出了以最小估计丰度协方差和单形体各顶点到中心点均方距离总和最小约束的非负矩阵分解（MCMDNMF）算法,其采用投影梯度作为非负矩阵分解的迭代学习规则.MCMDNMF既利用了非负矩阵分解的优点又考虑了高光谱图像的特性,也不需要混合像元中必须有纯像元.仿真实验表明,MCMD-NMF算法能正确地解混出高光谱混合像元中含有的端元光谱,并精确估计出丰度分布.

英文摘要：

The existence of mixed pixels impacts the precision advancement of hyperspectral remote sensing application.It is a new research direction to solve the problem of hyperspectral unmixing by nonnegative matrix factorization（NMF）.The nonconvexity of the objective function causes an error to optimal solution in the classic NMF.In this paper,by analyzing the characteristics of endmember signatures and spatial distribution of hyperspectral images,a new approach called minimum covariance and minimum distances nonnegative matrix factorization（MCMDNMF） was proposed.it is the minimum estimated abundance covariance and minimum the sum of squared distances between all the simplex vertices constrained by the NMF,adopting projected gradient as the iterative learning rule for NMF.MCMDNMF combines the merit of NMF and the characteristics of hyperspectral data,and at the same time,eliminates the pure-pixel assumption.Experimental results demonstrate that the MCMDNMF method can extract the endmember signature and accurately estimate abundance maps.