基于聚类和联合偏度与峰度指数的高光谱数据波段选择算法

Hyperspectral Data Band Selection Based on Clustering Joint Skewness-Kurtosis Index

高光谱数据的光谱分辨率高,数据量大,在提供丰富、详细的地物或目标信息的同时,不同光谱波段特别是相邻波段间有较强的相关性,导致光谱波段之间有大量冗余信息。针对这一问题,面向高光谱异常检测,提出了基于聚类和联合偏度与峰度指数的波段选择方法。首先利用虚拟维度进行估计,确定高光谱数据的本征维度,再结合最大-最小距离的思想进行聚类中心的更新,避免了随机选取的初始值可能导致距离太近的问题。然后,考虑到异常目标常常表现为不满足背景高斯分布的特点,使用联合偏度与峰度指数作为准则函数进行波段选择,有效选择出了重要波段。在三组代表性高光谱数据集上进行了实验,结果表明本文所提出的算法有效提升了高光谱异常检测的效果并降低了虚警率。

Hyperspectral data has hundreds of bands with a huge amount of data. While providing rich and detailed information, there is strong correlation between different bands, especially between adjacent bands, leading to a large amount of information redundancy. To solve this problem, a joint Skewness-Kurtosis index band selection method based on cluster was proposed for hyperspectral anomaly detection. Firstly, the virtual dimension is used to estimate the intrinsic dimension of hyperspectral data, and then the clustering center is updated with the idea of maximum and minimum distance to avoid the problem that the initial values are randomly selected too close to each other. Then, considering that the abnormal target does not satisfy the background Gaussian distribution, the joint Skewness-Kurtosis index is used as the criterion function to select the important bands effectively. Experiments on three representative hyperspectral data sets show that the proposed algorithm can effectively improve the performance of hyperspectral anomaly detection and reduce the false alarm rate.