第60卷第2期/2023年1月/激光与光电子学进展0228010-1研究论文基于S2AF-GCN的高光谱遥感图像分类模型宋海林，汪西莉*陕西师范大学计算机科学学院，陕西西安710119摘要在高光谱图像分类任务中，图卷积网络能够建模像素或区域间的结构关系和相似性关系。针对利用像素原始光谱特征计算节点相似度构造邻接矩阵不准确的问题，提出基于空间-光谱聚合特征的图卷积网络（S2AF-GCN），用于特征提取和像素级分类。以像素的空间位置为中心，聚合像素空间邻域内的其他像素特征，利用聚合后的像素特征动态更新与邻域内其他像素的权重，通过多次聚合，实现区域内像素特征平滑，得到像素的有效特征表示。然后利用聚合特征计算相似度并构图，获得更为准确的邻接矩阵，同时利用聚合特征训练网络，获得更好的分类结果。S2AF-GCN在三个常用高光谱数据集IndianPines、PaviaUniversity、KennedySpaceCenter上利用1%的标记样本取得了85.51%、96.95%、94.92%的总体分类精度。关键词图卷积网络；聚合特征；高光谱遥感图像分类；空谱信息中图分类号TP391文献标志码ADOI：10.3788/LOP220612HyperspectralRemoteSensingImageClassificationModelBasedonS2AF-GCNSongHailin,WangXili*SchoolofComputerScience,ShaanxiNormalUniversity,Xi’an710119,Shaanxi,ChinaAbstractForhyperspectralimageclassificationtasks,agraphconvolutionalnetworkcanmodelthestructuralandsimilarityrelationshipsbetweenpixelsorregions.Tosolvetheproblemofinaccurateconstructionofanadjacencymatrixbycalculatingthenodesimilarityusingtheoriginalspectralfeaturesofpixels,agraphconvolutionalnetworkbasedonspatial-spectralaggregationfeatures(S2AF-GCN)isproposedforfeatureextractionandpixel-levelclassification.TheS2AF-GCNconsidersthespatialpositionofthepixelasthecenter,aggregatesotherpixelfeaturesinthespatialneighborhoodofthepixel,andusestheaggregatedpixelfeaturestodynamicallyupdatetheweightsofotherpixelsintheneighborhood.Throughmultipleaggregations,thepixelfeaturesintheregionaresmoothed,andtheeffectivefeaturerepresentationofthepixelsisobtained.Next,theaggregatedfeaturesareusedtocalculatethesimilarityandconstructamoreaccurateadjacencymatrix.Moreover,theaggregatedfeaturesaresimultaneouslyusedtotraintheS2AF-GCNtoobtainbetterclassifica...
