An initialization method of non-negative matrix factorization for hyperspectral data unmixing based on spectral shape and information dissimilarity
-
摘要: 在高光谱像元解混应用中,好的端元光谱矩阵初始化方法对于提高盲信号分解精度具有重要意义.针对空间分辨率较高的高光谱数据,提出了一种新的面向非负矩阵分解(non-negative matrix factorization,NMF)的初始化方法.该方法通过计算像元在谱线形状和信息量差异等方面的参数,利用像元谱线峭度、KL散度和光谱角等参量,从众多混合像元中识别出纯像元;并分辨出不同类型纯像元(或类纯像元)之间的差别,从中选择最适合代表每一类型端元的纯像元(或类纯像元)作为算法的初值像元,完成端元矩阵的初始化.将此方法分别用于模拟数据和真实数据的实验结果表明,该方法能够明显提高高光谱混合数据的NMF精度,相比其他常用初始化方法具有更好的效果.
-
关键词:
- 初始化 /
- 盲信号分解 /
- 非负矩阵分解(NMF) /
- 谱线形状 /
- 信息量差异
Abstract: When blind signal separation technique is applied to unmixing hyperspectral data, a good initialization is vital for improving separating precision. Aimed at the hyperspectral data with relatively high spatial resolution and simple surface features, the authors put forward a reasonable hypothesis that the data contain pure pixel or approximate pure pixel corresponding to the each type of end-members, and proposed a new initialization method of non - negative matrix factorization ( NMF ) , which has great potential in pixel unmixing. By calculating parameters to quantify the spectral shape and information difference among candidate pixels, this method extracts pure pixels from mixed pixels, recognizes the information dissimilarity among different types of pure pixels and choose the existing pixels that are most suitable for representing each type of end -members as NMF ' s initial values. The experimental results show that the method proposed in this paper can improve NMF' s decomposition accuracy of hyperspectral data significantly, and its performance is better than that of other NMF initialization meth-ods. -
-
计量
- 文章访问数: 419
- PDF下载数: 24
- 施引文献: 0
下载: