张一鸣 陈培培 王汉谱 徐将 欧先锋 徐智

摘要：由于光谱变异性现象在高光谱图像数据中所具有的普遍性，经常导致“同物异谱”现象，高光谱目标检测结果受先验目标特征的影响很大，检测性能受到很大影响。因此提出一种基于稀疏表示的方法来生成优化的目标光谱，当缺乏關于感兴趣目标对象的全面信息时，通过稀疏字典重构的方法优化目标，然后由一组选定的字典生成像素，这些像素包含具有不同状态的目标信号，用来优化先验目标签名。最后将先验目标用稀疏字典重构的方式，利用光谱角分别构造目标训练和背景训练样本，使得来自有限目标训练的样本能够减轻光谱变异性对高光谱目标检测的影响。实验结果表明：该方法在不同数据集中具有良好的检测效果和精度，尤其在AVIRIS数据集中的检测精度高达0.997 8，优于其他典型检测算法。

关键词：光谱角;稀疏表示;光谱变异性;高光谱图像，目标检测

中图分类号：TP391 文献标志码：A

文章编号：2095-5383（2020）04-0007-06

Hyperspectral Images Target Detection based on

Spectral Angle Sparse Dictionary Reconstruction

ZHANG Yiming1，2， CHEN Peipei1，2， WANG Hanpu1，2， XU Jiang1，2， OU Xianfeng1，2， XU Zhi 3

（1. School of Information and Communication Engineering， Hunan Institute of Science and Technology， Yueyang 414006， China; 2. Machine Vision & Artificial Intelligence Research Center， Hunan Institute of Science and Technology， Yueyang 414006， China; 3. Guangxi Key Laboratory of Images and Graphics Intelligent Processing， Guilin University of Electronics Technology， Guilin 541004， China）

Abstract： Hyperspectral images have “integration of map and spectrum”， high spectral resolution， and a large number of bands advantages. However， due to the universality of spectral variability in hyperspectral image data， it often leads to the phenomenon of“same object with different spectra” and reduces detection performance， and the detection results of hyperspectral targets are greatly affected by the characteristics of a priori targets.A method based on sparse representation was proposed in this paper to generate optimized target spectra. When there is a lack of comprehensive information about the target object of interest， the target is optimized by sparse dictionary reconstruction， and then pixels are generated from a set of selected dictionaries. These pixels contain target signals with different states to optimize the prior target signature. Finally， the prior target is reconstructed with a sparse dictionary， and the spectral angle is used to construct target training and background training samples respectively， so that the samples from limited target training can reduce the influence of spectral variability on the detection of hyperspectral targets. Experiments show that this algorithm has obvious advantages in different data sets， especially the detection accuracy in the AVIRIS data set is as high as 0.997 8， which is significantly better than other typical algorithms.

Keywords：

spectral angle; sparse representation; spectral variability; hyperspectral images; target detection

高光谱图像（hyperspectral images，HSI）具有很高的光谱分辨率，通常包含数百个连续波段，达到nm级，表现为不同地物在光谱维上的细微差异，能够精细地刻画地物的反射光谱，从而大大提高對地物分类与识别的能力。高光谱图像相比多光谱图像在目标检测方面更具优势，在材料分类、地质特征识别和环境监测等方面都具有重要应用[1-2]。目前，世界各国对高光谱成像遥感技术的发展都十分重视，随着技术的日趋成熟，高光谱成像遥感已经广泛应用于植被生态监测、精细农业食品安全、产品质量检测[3-4]、资源探测[5]等多个领域。……