面向对象高分影像归纳式图神经网络分类法  

Object-oriented high-resolution image classification using inductive graph neural networks

作　　者：谢志伟 翟帅智 张丰源 陈旻[2,3,4] 孙立双[1] XIE Zhiwei;ZHAI Shuaizhi;ZHANG Fengyuan;CHEN Min;SUN Lishuang(School of Transportation and Geomatics Engineering,Shenyang Jianzhu University,Shenyang 110168,China;Key Laboratory of Virtual Geographic Environment(Ministry of Education of PRC),Nanjing Normal University,Nanjing 210097,China;State Key Laboratory Cultivation Base of Geographical Environment Evolution(Jiangsu Province),Nanjing Normal University,Nanjing 210097,China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application,Nanjing Normal University,Nanjing 210097,China;School of Environment,Nanjing Normal University,Nanjing 210097,China)

机构地区：[1]沈阳建筑大学交通与测绘工程学院,辽宁沈阳110168 [2]南京师范大学虚拟地理环境教育部重点实验室,江苏南京210097 [3]南京师范大学江苏省地理环境演化模拟国家重点实验室培育建设点,江苏南京210097 [4]南京师范大学江苏省地理信息资源开发与应用协同创新中心,江苏南京210097 [5]南京师范大学环境学院,江苏南京210097

出　　处：《测绘学报》2024年第8期1610-1623,共14页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42101353);教育部人文社会科学研究一般项目(21YJC790129);辽宁省教育厅基本科研项目(LJKMZ20220946,LJKMZ20222128)。

摘　　要：传统面向对象分类多采用遥感影像的光谱特征,忽略了影像对象之间的空间特征。本文提出了一种采用改进归纳式图神经网络的高分遥感影像面向对象分类方法,实现了光谱-空间复合节点相似度的融合系数自适应调节,以及邻域节点采样最佳数量的自动确定。首先,改进KNN图构建方法,采用标准差信息量评价法确定用于构建光谱特征和空间特征的复合节点相似度的融合系数;然后,利用反馈曲线法确定最佳的采样邻域节点数量,使用GraphSAGE节点嵌入完成特征表达;最后,依托Softmax函数预测节点类别。以GID和BDCI2017数据集为试验数据,本文的构图方法相较于改进前的构图方法在分类精度上有所提升。本文分类方法的平均Kappa系数和总体精度分别优于CART分类树算法、GCN算法、GAT算法、LANet算法、CCTNet算法和SLCNet算法0.31、0.14、0.13、0.12、0.08、0.02和42.31%、7.4%、6.73%、8.69%、6.03%、1.52%,并且在植被和建设用地提取上具有较好的稳健性。本文方法为高分遥感影像土地覆盖分类提供了有效的工具。Traditional object-oriented classification methods mostly use spectral features of image objects and ignore the spatial features among image objects.In this paper,an object-oriented classification method for high-resolution remote sensing images using improved inductive graph neural network is proposed.The method is able to adaptively adjust the fusion coefficient of spectral-spatial composite node similarity and automatically determine the optimal sampling number of neighboring nodes.First,we improved the K-nearest neighbor(KNN)graph construction method.The standard deviation informativeness evaluation method was used to determine the fusion coefficients for constructing the composite node similarity of spectral and spatial features.Then,the optimal sampling number of neighboring nodes was determined using the feedback curve method,and feature representation was accomplished using GraphSAGE node embedding.Finally,the classifications of the nodes were predicted by Softmax function.We used GID-15 and BDCI2017 datasets as experimental data.The proposed graph construction method has improved the classification accuracy.The average Kappa coefficient of the proposed method was better than CART,GCN,GAT,LANet,CCTNet,and SLCNet by 0.31,0.14,0.13,0.12,0.08,and 0.02.The average overall accuracy,on the other hand,was better than 42.31%,7.4%,6.73%,8.69%,6.03%,and 1.52%.Meanwhile,our method had good robustness in vegetation and built-up land extraction.The method proposed in this paper provides an effective tool for land cover classification of high-resolution remote sensing images.

关 键 词：高分遥感影像 GraphSAGE 节点连接权重 聚合节点 土地覆盖分类 

分 类 号：P237[天文地球—摄影测量与遥感]