目标光谱指导下的高光谱图像混合像元分解方法

Target spectra guided spectral unmixing for hyperspectral images

针对现在高光谱图像混合像元分解方法需要对所提取的端元的物理含义进行诠释的问题，提出了一种目标光谱指导下的混合像元分解方法，并给出了其具体算法实现。该方法首先针对若干给定的、具有明确物理含义的目标光谱，将光谱识别步骤引入混合像元分解过程，建立端元光谱与目标光谱间的对应关系，其次在最小距离限制的非负矩阵分解（MDC—NMF）方法基础上，引入光谱特征距离（SFD）作为正则项，以度量和保持存在对应关系的端元光谱与目标光谱间的相似性，并给出求解相应优化问题的优化算法。分别用模拟数据和真实数据对该方法的可行性和实际混合像元分解效果进行了验证，并将其与非监督情况下混合像元分解结果进行了比较分析。实验结果表明，该方法能够在目标光谱指导下较好诠释端元的物理含义，同时解决端元提取中的病态性问题。

In this paper, the problem about how to interpret the physical meanings of the endmembers extracted by spectral unmixing for hyperspectral images is paid special attention, and a new target spectra guided spectral unmixing method is proposed. The method is described below： Firstly, for some given target spectra with specific physical meanings, spectrum recognition is introduced into the procedure of spectral unmixing in order to establish the corre- spondence between endmember spectra and target spectra; Secondly, under the framework of minimum distance constrained nonnegative matrix faetorization （MDC-NMF） method, a new regularization term, namely spectral fea- ture distance （SFD）, is proposed to measure and maintain the similarity between endmember spectra and target spectra which share a correspondence, and meanwhile, to solve the relevant optimization problem, an optimization algorithm is proposed accordingly. The feasibility and the real performance of the new method were examined and compared with that of the unsupervised spectral unmixing method by using a set of synthetic data and real hyper- spectral images. The experimental results demonstrate that, by introducing the guidance of target spectra, the phys- ical meanings of the extracted endmembers can be properly interpreted and the ill-posedness of the solution of spec- tral unmixing can be eliminated.