地图综合图卷积神经网络点群简化方法  

A point cluster simplification approach of graph convolutional neural network for map generalization

作　　者：肖天元 艾廷华[1,2] 余华飞 杨敏[1,2] 刘鹏程[3] XIAO Tianyuan;AI Tinghua;YU Huafei;YANG Min;LIU Pengcheng(School of Resource and Environmental Sciences,Wuhan University,Wuhan 430079,China;Key Laboratory of Geographic Information System,Ministry of Education,Wuhan 430079,China;School of Urban and Environmental Sciences,Central China Normal University,Wuhan 430079,China)

机构地区：[1]武汉大学资源与环境科学学院,湖北武汉430079 [2]地理信息系统教育部重点实验室,湖北武汉430079 [3]华中师范大学城市与环境科学学院,湖北武汉430079

出　　处：《测绘学报》2024年第1期158-172,共15页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42071455,42071450)。

摘　　要：地图综合是一个多因素影响的复杂决策过程,针对不同场景判断下的综合算子优化选择常规上通过基于规则的方法实施。顾及不同特殊条件影响,这些地图综合规则需要“打补丁”,导致地图综合规则体系越来越复杂,从而失去普适性。人工智能技术下的数据驱动综合方案通过机器学习提取典型案例中隐含的综合规则,并迁移到新的数据场景,为处理特殊规则下的地图综合提供了一种思路。本文引入深度学习技术,采用领域知识与数据驱动相结合的策略,提出了一种基于图卷积神经网络的点群自动综合方法。本文方法通过样本训练与深度学习获取不同数据场景下的地图综合知识,同时融入既有规则进行引导,可以更有效地向人工地图综合结果的目标迈进。首先,构建Delaunay三角网,在点群之间建立空间邻域关系,并根据地理空间上下文关联、空间异质性等领域知识计算各个点的特征信息,构造点群的特征向量。其次,引入拓扑自适应图卷积神经网络,构建点群数据自动综合网络模型。试验表明本文方法在局部区域与整体地图上均可以保持原始点群的各项特征,体现在相对数量保持、上下文特征继承、属性特征一致方面均有良好的表达效果。Map generalization is a complex decision-making process with multiple factors,and the optimal selection of generalization operators for different scenarios is routinely implemented through a rule-based approach.These map generalization rules need to be"patched"to take into account the influence of different special conditions,resulting in an increasingly complex system of map generalization rules that lose their universality.The data-driven simplification scheme with artificial intelligence technology provide a new way of thinking about map generalization under special rules by extracting the generalization rules implied in typical cases through machine learning and migrating them to new data scenarios.In this paper,deep learning techniques are introduced and a strategy combining domain knowledge and data-driven is used to propose an automatic generalization method for point clusters based on graph convolutional neural networks.This method obtains the knowledge of map generalization in different data scenarios through sample training and deep learning,while incorporating established rules for guidance,which can move more effectively toward the goal of artificial map generalization results.Firstly,a Delaunay triangulation network is constructed to establish spatial neighbourhood relationships between points and to calculate the feature information of each point based on domain knowledge such as geospatial contextual associations and spatial heterogeneity to construct the feature vectors of the point cluster.Secondly,the topological adaptive graph convolutional neural network is introduced to construct an automatic generalization network model of the point cluster data.The experiments show that the algorithm can maintain the features of the original point cluster in both the local area and the overall map,which is reflected in the relative quantity maintenance,contextual feature inheritance and attribute feature consistency.

关 键 词：地图综合 点群 DELAUNAY三角网 图卷积神经网络 数据驱动 

分 类 号：P208[天文地球—地图制图学与地理信息工程]