面状居民地形状分类的图卷积神经网络方法  

Graph convolution neural network method for shape classification of areal settlements

作　　者：于洋洋 贺康杰[1,2] 武芳 许俊奎[1,2] YU Yangyang;HE Kangjie;WU Fang;XU Junkui(College of Geography and Environmental Science,Henan University,Kaifeng 475004,China;Key Laboratory of Geospatial Technology for the Middle and Lower Yellow River Regions(Henan University),Ministry of Education,Kaifeng 475004,China;College of Surveying and Mapping,Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]河南大学地理与环境学院,河南开封475004 [2]黄河中下游数字地理技术教育部重点实验室(河南大学),河南开封475004 [3]信息工程大学地理空间信息学院,河南郑州450001

出　　处：《测绘学报》2022年第11期2390-2402,共13页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(41471386);自然资源部城市国土资源监测与仿真重点实验室开放基金资助课题(KF-2020-05-037)。

摘　　要：形状识别和分类是地图制图综合的重要内容之一,面状居民地要素作为地理空间矢量数据的重要组成部分,其形状认知是制图综合的基础。本文针对当前几何和统计形状分类方法的不足,借助图卷积神经网络的图数据分类能力,提出了一种基于图卷积神经网络的面状居民地形状分类方法。该方法首先从面状居民地轮廓多边形入手,提取其轮廓的多个特征,获取形状的图表达;然后,利用图卷积神经网络对居民地形状信息进行多轮次提取和聚合,将形状信息嵌入一个高维向量中;最后利用全连接神经网络对高维形状向量进行分类。试验表明,该方法能够有效提取居民地形状信息,克服了传统分类方法人为设置指标的不足,实现了端到端的居民地形状信息提取与分类。Shape recognition and classification is one of the important contents of cartographic generalization.Areal settlement is an important part of geospatial vector data and its shape cognition is a basic technique of cartographic generalization.To solve the shortcomings of traditional geometric and statistical shape classification methods,this paper proposes a novel areal settlements shape classification method based on graph data classification ability of graph convolutional neural network.Firstly,the computation graph is generated according to the contour polygon of areal settlement,and the features of the contour shape are extracted as the attributes of the vertices of computation graph.Secondly,the vertex attributes of the computation graph are aggregated and transmitted for multiple rounds,and the shape information is embedded into a high dimension vector with these vertices attributes.Finally,the graph vectors are input into a fully connected neural network to realize the classification of graphs.The experimental results show that this method can effectively achieve the end-to-end shape information extraction and classification of areal settlements.And it overcomes the deficiency of setting parameters through experience in traditional methods.

关 键 词：面状居民地 图卷积神经网络 形状分类 制图综合 图分类 

分 类 号：P208[天文地球—地图制图学与地理信息工程]