基于加权近邻保持嵌入的高光谱数据降维方法

Dimensionality reduction for hyperspectral data using weighted neighborhood preserving embedding

为降低高光谱数据的信息冗余以提高其分类精度,采用加权距离度量测度来衡量样本间的相似度并进而选择近邻样本,提出一种加权近邻保持嵌入数据降维(WNPE)算法.加权距离的主要思想为根据数据点附近样本点的分布来自适应地决定距离函数,由此可以避免基于标准欧氏距离的近邻选择方法产生的数据冗余现象,从而更好地提取信息量大的光谱波段.CUPRITE矿区高光谱数据上的实验结果表明,与目前具有代表性的稀疏降维和基于流形学习的降维算法对比,WNPE能够有效提高高光谱数据的分类总精度和Kappa系数,分别达到了90.97%和0.878 6.

In order to reduce the information redundancy and thus to improve the classification accuracy of hyperspectral data, a weighted distance metric was used to measure the similarity between samples and thus neighbor samples was selected. Based on the weighted distance met- ric, a dimensionality reduction algorithm called weighted neighborhood preserving embedding （WNPE） was proposed. The main idea of the weighted distance is to adaptively determine the distance function according to the distribution of the near samples. Therefore, the weighted distance-based neighbor-selection can avoid the data redundancy resulted from standard Euclid- ean distance, which is beneficial for the extraction of spectral bands containing large amount of information. Experimental results on CUPRITE hyperspectral data show that WNPE can effec- tively improve the overall classification accuracy and Kappa coefficient of hyperspectral data, which are 90.97% and 0. 8786 respectively, in contrast to present typical dimensionality reduc- tion algorithms based on sparisty representation or manifold learning.