摘要： 传统的非平滑约束的非负矩阵分解算法(nsNMF)在处理高光谱数据时，存在对初始值敏感、容易陷入局部最优值等缺陷。为此，提出一种基于粒子群优化(PSO)的nsNMF算法。采用传统nsNMF算法迭代的结果作为初始值，以避免PSO的盲目搜索。通过PSO搜索端元光谱矩阵，利用nsNMF算法更新端元光谱矩阵和丰度矩阵，以缩小搜索空间，降低计算复杂度，避免陷入局部最优。在合成数据集和真实数据集上的实验结果表明，与传统nsNMF算法相比，该算法能获得更好的全局最优解，端元光谱和丰度值更接近真实值。

关键词: 非负矩阵分解, 粒子群优化算法, 高光谱, 线性光谱模型, 全局最小值, 稀疏性

Abstract: The traditional nonsmooth Nonnegative Matrix Factorization(nsNMF) is an effective algorithm to deal with hyperspectral data, but the weaknesses of being sensitive to initial values and easily falling to local minimum limit its applying. To solve this problem, the paper proposes a nonsmooth nonnegative matrix factorization algorithm based on Particle Swarm Optimization(PSO) which can find global minimum. It uses the output of nsNMF as the initial values to avoid blind search. It searches endmembers fraction by PSO, then updates endmembers and abundance matrix by nsNMF algorithm. Experimental results based on synthetic data set and truthful data set shows that this algorithm has better global optimal solution. Its endmember and abundance are closer to true value.

Key words: Nonnegative Matrix Factorization(NMF), Particle Swarm Optimization(PSO) algorithm, hyperspectral, Liner Spectral Mixture Model(LSMM), global minimum, sparseness

中图分类号: 

• TP18