Image Euclidean distance-based manifold dimensionality reduction algorithm for hyperspectral imagery

Abstract

提出两种基于图像欧氏距离的非线性降维方法.该方法利用高光谱图像物理特性, 将图像欧氏距离引入到传统的流形降维算法中.与其它应用于高光谱图像的降维算法相比, 该算法具有诸多优点.图像欧氏距离的引入, 在考虑高光谱图像本身的空间关系的同时, 很好地保持了数据点之间的局部特性, 可以实现有效地去除原始数据集光谱维和空间维的冗余信息.实际高光谱数据的实验结果表明, 该算法应用于高光谱图像分类时, 与其它常见的方法相比具有更高的分类精度.;Two nonlinear dimensionality reduction methods were proposed based on image Euclidean distance. Considering the physical characters of hyperspectral imagery, the methods introduced image Euclidean distance into traditional manifold dimensionality reduction. Compared with other methods, our methods have several advantages. The introduction of image Euclidean distance not only considers hyperspectral image’s spatial relationship, but also preserves the local feature of datasets well. Thus the proposed methods can discard efficiently the redundant information from both the spectral and spatial dimensions. The experiment results demonstrated that the proposed methods have higher classification accuracy than other methods when applied to hyperspectral image classification.