一种用于高光谱图像特征提取的子空间核方法

摘要

特征提取对于实现高光谱遥感图像的有效信息挖掘和利用以及提高后续分类应用有着重要价值。为了改进降维效果，提出一种子空间调制的核主成分分析方法，将高光谱数据分组特性整合到一个统一的核方法框架中，并构造子空间调制核。子空间调制核依靠特征分组实现了在光谱波段上的稀疏调制，它也是一个数据自适应的核，用于度量高光谱数据样本间的非线性相似性。该方法利用AVIRIS真实高光谱图像进行评估，并且与传统的核方法、光谱加权核方法进行了比较。实验结果表明，基于子空间调制的核方法更充分地利用了波段间复杂相关的物理特性，进而在高光谱图像分类方面的结果好于传统的核方法与光谱加权核方法。%Feature extraction is quite valuable for the mining and utilization of valid information in hyperspectral re-mote-sensing imaging and the increase of subsequent classified applications. For improving the dimension reduction effect, a subspace-modulated kernel principal component analysis ( SM-KPCA) method is proposed. With this method, the grouping natures of hyperspectral data are integrated into a uniform kernel method framework and a subspace-modulated kernel is constructed. SMK( subspace-modulated kernel) achieves a sparse modulation on the spectral waveband by means of feature grouping;in addition, it is a data-adaptive kernel for measuring the nonlin-ear similarities among the hyperspectral data specimens. With the proposed method, AVIRIS( airborne visible infra-red imaging spectrometer) real hyperspectral imaging is applied for evaluation. Additionally, this method is com-pared with the conventional kernel method and the spectrally weighted kernel method. The experimental results show that the SM-KPCA method more sufficiently utilizes the complex and relevant physical characteristics between wavebands. Therefore, it outperforms both the conventional kernel methods and the spectrally weighted kernel meth-od regarding the aspect of the classification of hyperspectral images.