高光谱图像稀疏约束与自编码器特征提取相结合的异常检测方法  

Hyperspectral anomaly detection combining sparse constraint and feature extraction via stacked autoencoder

作　　者：宋尚真 杨怡欣 王会峰[1] 王晓艳[1] 荣生辉 周慧鑫 SONG Shangzhen;YANG Yixin;WANG Huifeng;WANG Xiaoyan;RONG Shenghui;ZHOU Huixin(School of Electronics and Control Engineering,Chang an University,Xi'an 710064,China;School of Communications and Information Engineering,Xi'an Uiversity of Posts&Telecommunications,Xi'an 710121,China;School of Electronic Engineering,Ocean University of China,Qingdao 266100,China;School of Physics and Radio and Television Engineering,Xi'an University of Electronic Science and Technology,Xi'an 710068,China)

机构地区：[1]长安大学电子与控制工程学院,陕西西安710061 [2]西安邮电大学通信与信息工程学院,陕西西安710121 [3]中国海洋大学电子工程学院,山东青岛266100 [4]西安电子科技大学物理与光电工程学院,陕西西安710071

出　　处：《测绘学报》2023年第6期932-943,共12页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(52172324);陕西省重点研发计划(2021GY-285,2021SF-483)。

摘　　要：高光谱图像的异常检测在军事、农业、勘探、防火等领域具有重要的应用价值。传统的高光谱图像异常检测算法未能有效地挖掘图像光谱的深层特征,而深度学习方法具有良好的提取深层特征信息的能力。由于异常检测问题一般无法获取地物先验信息,因此无监督网络相比于监督网络要更为适用。而现有的基于自编码器的异常检测算法没有对局部信息进行有效利用,导致检测效果受限。针对这一问题,本文提出一种基于稀疏表示约束的自编码器深度特征提取方法。首先通过栈式自编码器得到深层次语义信息;然后利用稀疏表示作为约束与编码器进行有效结合,挖掘了潜在隐藏空间中的特征元素的局部表示特性;最后采用分数傅里叶变换,通过空间-频率表示获得原始光谱与其傅里叶变换的中间域中的特征,进一步增强了背景和异常的光谱区分度,且能有效去除噪声的影响。在Hymap、AVIRIS、ROSIS、HYDICE这4种光谱仪采集的5幅高光谱遥感影像上进行了性能验证,得到的曲线下覆盖面积(area under curve,AUC)分别为0.9905、0.9983、0.9990、0.9928和0.9110,相比于对比算法都有了不同程度的效果提升。结果表明本文方法具有更好的检测精度。Anomaly detection of hyperspectral images has important application value in military,agriculture,exploration,fire protection and other fields.Traditional algorithms of hyperspectral image(HSI)anomaly detection(AD)do not effectively mine the deep features of the image spectrum,while the deep learning method has good ability to extract deep feature information.Since the AD problem generally cannot obtain the prior information in advance,the unsupervised network is more suitable.Existing AD algorithms based on autoencoder(AE)does not make effective use of the local information,resulting in limited detection effect.To overcome this shortcoming,the paper proposes an AD method based on sparse representation(SR)constraints for stacked autoencoder(SAE).Firstly,the semantic information is obtained by SAE.Secondly,the SR is used as a constraint to effectively combine with the encoder,and the local characteristics of the feature elements in the potential hidden space are mined.Finally,the fractional Fourier transform is utilized,and the characteristics of the original spectrum and its intermediate domain of Fourier transform are obtained by spatial-frequency representation.Consequently,the spectral discrimination between background and anomalies is further enhanced,and the effect of noise is also removed.The experiment performs verification on 5 HSIs collected by 4 spectrometers including Hymap,AVIRIS,ROSIS,and HYDICE.The area under curve(AUC)values are 0.9905,0.9983,0.9990,0.9928 and 0.9110,respectively.Compared with compared algorithms,the effect of the proposed algorithm can be improved.

关 键 词：高光谱影像 异常检测 深度学习 自编码器 稀疏表示 傅里叶变换 

分 类 号：P227[天文地球—大地测量学与测量工程]