A Kernel—based Nonlinear Subspace Projection Method for Dimensionality Reduction of Hyperspectral Image Data  

作　　者：GUYanfeng ZHANGYe QUANTaifan 

机构地区：[1]DepartmentofElectronicsandCommunicationEngineering,HarbinInstituteofTechnology,Harbin150001,China

出　　处：《Chinese Journal of Electronics》2003年第2期203-207,共5页电子学报（英文版）

摘　　要：A challenging problem in using hyper-spectral data is to eliminate redundancy and preserve useful spectral information for applications. In this pa-per, a kernel-based nonlinear subspace projection (KNSP)method is proposed for feature extraction and dimension-ality reduction in hyperspectral images. The proposed method includes three key steps: subspace partition of hyperspectral data, feature extraction using kernel-based principal component analysis (KPCA) and feature selec-tion based on class separability in the subspaces. Accord-ing to the strong correlation between neighboring bands,the whole data space is partitioned to requested subspaces.In each subspace, the KPCA method is used to effectively extract spectral feature and eliminate redundancies. A criterion function based on class discrimination and sepa-rability is used for the transformed feature selection. For the purpose of testifying its effectiveness, the proposed new method is compared with the classical principal component analysis (PCA) and segmented principal component trans-formation (SPCT). A hyperspectral image classification is performed on AVIRIS data. which have 224 svectral bands.Experimental results show that KNSP is very effective for feature extraction and dimensionality reduction of hyper-spectral data and provides significant improvement over classical PCA and current SPCT technique.

关 键 词：特征特取 Kernel函数 非线性子空间 图形数据 KPCA 

分 类 号：TN911.73[电子电信—通信与信息系统]