Abstract: Hyperspectral remote sensing images provide almost continuous spectrum for the surface observation of the earth. However, there are more redundancy information within the bands, which pose challenges to storage, analysis and visualization of the hyperspectral remote sensing images. Thus, dimensionality reduction for hyperspectral remote sensing images become a key step of preprocessing. In this paper, the theory of information entropy is adopted to design the dimensionality reduction method. Each band of hyperspectral remote sensing images are abstracted as independent individual, and the decision table matrix of hyperspectral remote sensing images are designed, and then the bands are ordered based on the information gain. According to the accuracy requirement of the analysis of hyperspectral remote sensing images, the different bands are selected based on the ordered bands, which are considered as the results of the dimensionality reduction. Taking the classification accuracy of remote sensing as an example to evaluate the feasibility and superiority of the dimensionality reduction methods in hyperspectral remote sensing, and based on the experimental results, the proposed method has marked advantages and outperformed other methods.

Key words: dimensionality reduction method, information entropy, hyperspectral remote sensing image

摘要： 高光谱遥感影像以其众多的波段数目，为地表观测提供近乎连续的波谱数据；然而海量的高光谱遥感影像存在着大量的信息冗余，为数据的处理带来了挑战。因此在对高光谱遥感影像进行存储、分析及可视化等操作之前，对高光谱遥感影像降维处理成为预处理的关键环节之一。利用信息熵理论，将高光谱遥感影像的各波段抽象为具有相关性的独立个体，设计了高光谱遥感影像的决策表矩阵，进而计算各波段的信息熵，量化各波段的信息量，从而将各波段根据信息增益进行排序。用户可根据高光谱遥感影像应用的精度需求，按排序选择波段组合，从而达到降维目的。以遥感分类结果的精度评价为例，对高光谱遥感降维方法的可行性和优越性进行评价。实验结果表明，该方法相较其他特征选取降维方法，能获得更高的分类精度。

关键词: 降维方法, 信息熵, 高光谱遥感影像