Abstract:

In order to reduce the redundant information of hyperspectral image data， optimize the computational efficiency and improve the effectiveness of subsequent applications of image data， a hyperspectral band selection algorithm based on Neighborhood Entropy （NE） was proposed. Firstly， in order to efficiently calculate the neighborhood subset of samples， the Local Sensitive Hashing （LSH） was used as the nearest neighbor search strategy. Then， the NE theory was introduced to measure the Mutual Information （MI） between bands and classes， and minimization of the conditional entropy between feature sets and class variables was used as a method to select effective bands. Finally， two datasets were used to carry out classification experiments through Support Vector Machine （SVM） and Random Forest （RM）. Experimental results show that， compared with four MI based feature selection algorithms， from the perspectives of overall accuracy and Kappa coefficient， the proposed algorithm can select effective band subset within 30 bands faster and achieve local optimization. Some experimental results of the proposed algorithm reach 92.99% and 0.860 8 at the global optimum on overall accuracy and Kappa coefficient respectively， verifying that the proposed algorithm can effectively deal with hyperspectral band selection problem.

Key words: band selection, hyperspectral image, Mutual Information (MI), Neighborhood Entropy (NE), approximate nearest neighbor search

摘要：

为了减少高光谱图像数据中的冗余信息，优化计算效率，并提升图像数据后续应用的有效性，提出一种基于邻域熵（NE）的高光谱波段选择算法。首先，为了高效计算样本的邻域子集，采用了局部敏感哈希（LSH）作为近似最近邻的搜索策略；然后，引入了NE理论来度量波段和类之间的互信息（MI），并把最小化特征集合与类变量之间的条件熵作为选取有效波段的方法；最后，采用两个数据集，通过支持向量机（SVM）和随机森林（RM）进行分类实验。实验结果表明，相较于四种基于MI的特征选择算法，从总体精度以及Kappa系数上看，所提算法能够在30个波段内较快地选取有效波段子集，并达到局部最优。该算法的部分实验结果的总体精度以及Kappa系数分别达到全局最优的92.99%以及0.860 8，表明所提算法能有效地处理高光谱波段选择问题。

关键词: 波段选择, 高光谱图像, 互信息, 邻域熵, 近似最近邻搜索