自动子空间划分在高光谱影像波段选择中的应用

Study on Auto-Subspace Partition for Band Selection of Hyperspectral Image

针对高光谱遥感影像数据量大、维数高的特点,结合联合熵波段选择算法,提出了一种自动子空间划分的改进方案。该方法充分利用了影像各波段数据之间的局部相关性,根据波段间相关系数矩阵图像的"分块"特点,将整个波段空间自动划分为若干个子空间,然后再进行波段选择。实现了在删减冗余信息的同时选择出含有主要信息的特征波段组合的目的。将此方法得到的结果与用联合熵得到的结果进行了比较分析,结果表明自动子空间划分的联合熵波段选择方法具有较好的效果。

Recent works on spectral band selection include two separate tasks： feature band selection and redundancy reduction. But due to the characteristic of hyperspectral data,it is not sufficient for joint entropy algorithm to select feature bands which aim at dimensionality reduction,for the band combination results it selected are in a series of space. To solve this problem,a new approach based on auto-subspace partition （ASP） was proposed. In this approach the subspace of all bands was dependent on correlation coefficient matrix among all bands,and from that we can get the relations among different bands about its spectral characteristic. In ASP,firstly,all bands were divided into different subspaces according to correlation coefficient matrix,then the optimal bands combination was selected using joint entropy algorithm respectively in different subspaces. The band combination results which were derived from our proposed approach were compared with those from joint entropy algorithm in the experiments. It has shown that the approach we proposed works better than the conventional joint entropy algorithms on hyperspectral data.