端到端分布式联合优化的空谱自编码密度估计模型  

End-to-End Spectral-Spatial Cooperative Autoencoding Density Estimation Model

作　　者：马纪涛 谢卫莹 雷杰[1] 方乐缘 李云松[1] MA Ji-tao;XIE Wei-ying;LEI Jie;FANG Le-yuan;LI Yun-song(State Key Laboratory of Integrated Services Networks,Xidian University,Xi’an,Shaanxi 710071,China;College of Electrical and Information Engineering,Hunan University,Changsha,Hunan 410082,China)

机构地区：[1]西安电子科技大学综合业务网理论及关键技术国家重点实验室,陕西西安710071 [2]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《电子学报》2023年第4期1006-1020,共15页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62121001,No.U22B2014);中国科协青年人才托举工程(No.2020QNRC001)。

摘　　要：高光谱图像(Hyperspectral Image,HSI)由于其丰富的光谱信息和空间信息,被广泛应用于异常检测任务,在对地观测和深空探测中起到了重要作用.然而,现有的基于密度估计的高光谱异常检测(Hyperspectral Anomaly Detection,HAD)方法存在如下问题:一是没有联合优化概率密度估计和特征表示这两个不同的目标函数,导致深度神经网络无法学习到更加准确的概率密度函数和包含HSI固有信息的低维表示;二是缺乏高层次空间语义信息与低维流行中光谱信息的自适应融合.此外,随着光谱成像技术的发展,卫星或无人机所获取的HSI的体积越来越大,在遥感大数据的背景下,传统框架处理HSI变得十分困难,给HAD带来了极大的挑战.本文分别从以上问题出发,提出了端到端联合优化的空谱协同自编码密度估计(End-to-End Spectral-Spatial Cooperative Autoencoding Density Esti-mation,E2E-SSCADE)模型.基于二维卷积提取HSI空间特征,融合高光谱图像光谱特征和空间特征的低维表示以及重构误差表示,联合密度估计网络进行端到端的优化,并利用分布式学习实现了大体积高光谱图像的异常检测.实验表明,所提出的E2E-SSCADE可以从光谱向量、空间维度以及重构空间三个角度更深层次地挖掘HSI固有信息的低维表示,构建更加准确的背景模型,在有效分离背景和异常目标的同时,结合分布式学习实现了快速、准确的大体积高光谱图像的异常检测,在6个经典HAD数据集上达到了99.07%的精度和3.41倍的检测速度.实验代码见https://github.com/majitao-xd/E2E-SSCADE.git.Hyperspectral image(HSI)is widely used in anomaly detection because of its rich spectral and spatial in-formation,and plays an important role in the earth observation and deep space exploration.However,the existing hyper-spectral anomaly detection(HAD)methods based on density estimation have the following problems.First,there is no joint optimization of the two different objective functions of probability density estimation and feature representation,which results in the deep neural network being unable to learn more accurate probability density function and low-dimen-sional representation containing inherent information of HSI;the other is the lack of adaptive fusion of high-level spatial semantic information and low-dimensional epidemic spectral information.In addition,with the development of spectral imaging technology,the volume of HSI acquired by satellites or unmanned aerial vehicles is increasing.In the context of remote sensing big data,it becomes very difficult for traditional frameworks to process HSI,posing a great challenge to HAD.In this paper,based on the above problems,an end-to-end spectral-spatial cooperative autoencoding density estima-tion(E2E-SSCADE)model is proposed.The HSI spatial features are extracted based on two-dimensional convolution,and the spectral features and spatial features of hyperspectral images are combined with the low-dimensional representation and reconstruction error representation.The end-to-end optimization is carried out by combining the density estimation net-work,and the anomaly detection of large hyperspectral images is realized by distributed learning.Experiments show that the proposed E2E-SSCADE can excavate the low-dimensional representation of HSI intrinsic information from three per-spectives of spectral vector,spatial dimension and reconstructed space,and construct a more accurate background model.With distributed training,fast and accurate anomaly detection of hyperspectral images is realized.The proposed method achieves 99.07%accuracy and 3.41 times faster de

关 键 词：高光谱异常检测 空谱协同 估计网络 端到端优化 分布式学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP751.1[自动化与计算机技术—控制科学与工程]