图正则重加权稀疏约束的深度非负矩阵分解算法用于高光谱图像解混

Graph regularized reweighted sparsity constrained deep nonnegative matrix factorization for hyperspectral image unmixing

为有效挖掘高光谱图像隐层信息以提升混合像元的分解精度,提出一种图正则的重加权稀疏深度非负矩阵分解算法。算法考虑图像丰度矩阵在局部所具有的浅层图结构信息以及在全局所具有的稀疏特性,通过融合图正则项和基于重加权的稀疏正则项及非负矩阵的多层深度结构来提升对混合像元的分解能力。通过乘法更新规则对深度非负矩阵分解算法进行逐层更新以优化全局框架。基于光谱角度距离和均方根误差评价指标,在模拟数据集和真实数据集上的实验显示所提出的算法相比其他典型算法分别有最大约63%和9.7%的解混精度增益。实验证明所提出的图正则重加权稀疏约束的深度非负矩阵分解算法能有效提升高光谱图像的解混精度,更好地服务于国家重大需求。

In order to effectively exploit the hidden layer information of hyperspectral images and improve the decomposition accuracy of mixed pixels, this paper proposes a graph regularized reweighted sparse deep nonnegative matrix factorization algorithm. The algorithm respectively considers the graph structure information and sparsity properties of the abundance matrix from local and global perspectives. It improves the decomposition ability of mixed pixels by fusing the graph and the reweighted sparse constraints and the multilayer deep structure of nonnegative matrix. The algorithm is solved layer by layer by multiplication update rules for the purpose of optimizing the global framework. Based on spectral angle distance and root mean square error metrics, experiments on simulated and real data sets show that the proposed algorithm has the largest unmixing accuracy gain of about 63% and 9.7%, respectively, compared with other representative algorithms. Experiments indicate that the proposed algorithm can effectively improve the interpretation accuracy of hyperspectral images and serve the major national needs.