引用格式:易瑔,张宇航,宗艳桃,等.基于卷积神经网络的高光谱图像分类算法综述[J].电光与控制,2023,30(3):70-77.YIQ,ZHANGYH,ZONGYT,etal.Asurveyofhyperspectralimageclassificationalgorithmsbasedonconvolutionalneuralnetworks[J].ElectronicsOptics&Control,2023,30(3):70-77.基于卷积神经网络的高光谱图像分类算法综述易瑔,张宇航,宗艳桃,戴颜斌(中国人民解放军陆军装甲兵学院兵器与控制系,北京100000)摘要:高光谱图像拥有光谱分辨率高、图谱合一的优点,已经成为遥感科学的重要研究方向。大多数传统的高光谱图像分类方法是基于浅层人工特征且依赖于专家经验,已经难以满足当下的技术需求。近年来,随着卷积神经网络在人工智能领域的广泛应用,基于卷积神经网络的高光谱图像分类方法已经在分类精度和速度上取得突破性的进展。首先介绍了高光谱图像分类方法,分析了传统分类方法的局限性;然后根据卷积神经网络对高光谱图像特征提取方式的不同,将算法分为基于谱特征、空间特征和空谱特征3大类,并分析了每类算法的优缺点;最后对高光谱图像分类的小样本训练、实际应用和分类结果等问题提出建议。关键词:高光谱图像;深度学习;图像分类;卷积神经网络;特征提取;综述中图分类号:TP391.4文献标志码:Adoi:10.3969/j.issn.1671-637X.2023.03.013ASurveyofHyperspectralImageClassificationAlgorithmsBasedonConvolutionalNeuralNetworksYIQuan,ZHANGYuhang,ZONGYantao,DAIYanbin(DepartmentofWeaponryandControl,ArmyAcademyofArmedForces,Beijing100000,China)Abstract:Hyperspectralimagehasbeenconsideredasoneofthegreatestresearchdirectionsintheremotesensingscienceduetoitsadvantagesofhighspectralresolutionaswellasallowingforthesynchronousacquisitionofbothimagesandspectraofobjects.Mostconventionalhyperspectralimageclassificationmethods,however,arebasedon“shallow”handcraftedfeatures,andhighlyreliesonexpertknowledge,whicharedifficulttomeetthecurrenttechnicalrequirements.Inrecentyears,withthewideapplicationofconvolutionalneuralnetworksinthefieldofartificialintelligence,hyperspectralimageclassificationmethodsbasedonconvolutionalneuralnetworkshaveachievedbreakthroughsinclassificationaccuracyandspeed.Firstly,hyperspectralimageclassificationmethodsareintroduced...