基于双图正则的半监督NMF混合像元解混算法

Hyperspectral Unmixing Algorithm Based on Dual Graph-regularized Semi-supervised NMF

在高光谱图像中混合像元普遍存在,这极大地阻碍了高光谱遥感技术的发展进程,因此,在利用光谱图像的过程中,如何准确高效地进行混合像元解混是一个关键问题。对于高光谱图像混合像元分解,使用原始的非负矩阵分解(Nonnegative Matrix Factorization,NMF)算法面临一些困难:首先,其目标函数为非凸函数,难以求解得到全局最优解;其次,混合像元中并不存在纯像元。为了解决这些问题,文中提出一种新的算法——基于双图正则的半监督NMF(Dual graph-regularized Constrained Nonnegative Matrix Factorization,DCNMF)混合像元解混算法。该算法采用了梯度下降法和迭代更新法则,既考虑了高光谱数据流形与光谱特征流形的几何结构,又能跳出局部极值,从而求解得到全局最优解。通过真实的高光谱图像数据仿真实验表明,DCNMF算法能够准确高效地进行混合像元分解,改善了解混效果,提高了解混精度,节约了计算时间,加快了收敛速度。

In hyperspectral images,the existence of mixed pixels greatly impedes the development of hyperspectral remote sensing technology.Therefore,how to carry out unmixing accurately and efficiently in the process of using spectral images is a key problem.For hyperspectral unmixing,using original non-negative matrix factorization(NMF)algorithm faces some difficulties,for example,the objective function is non-convex function,so it is difficult to solve the global optimal solution.Besides,the pure pixel like element doesn’t exist in mixed pixel.In order to solve these problems,this paper proposed a mixed pixel unmixing algorithm namely dual graph-regularized constrained semi-supervised NMF(DCNMF).This algorithm adopts gradient descent algorithm and iterative updating rule,considers the geometric structures of hyperspectral data manifold and the spectral feature manifold,and can jump out of the local extremum,thus solving the global optimal solution.Real hyperspectral image data simulation experiments show that DCNMF algorithm can be used to decompose the mixed pixel accurately and efficiently,enhancing the effect of unmixing,improving the accuracy of mixing,saving the computing time and speeding up convergence.