测绘学报 ›› 2023, Vol. 52 ›› Issue (9): 1574-1583.doi: 10.11947/j.AGCS.2023.20220216

• 地图学与地理信息 • 上一篇    下一篇

卷积神经网络支持下的建筑物选取方法

安晓亚1,2, 朱余德3, 晏雄锋4   

  1. 1. 西安测绘研究所, 陕西 西安 710054;
    2. 地理信息工程国家重点实验室, 陕西 西安 710054;
    3. 广东国地规划科技股份有限公司, 广东 广州 510650;
    4. 同济大学测绘与地理信息学院, 上海 200092
  • 收稿日期:2022-03-24 修回日期:2023-08-09 发布日期:2023-10-12
  • 作者简介:安晓亚(1982-),男,副研究员,研究方向为地图学与地理信息系统。E-mail:chxyaxy2022@163.com
  • 基金资助:
    地理信息工程国家重点实验室基金资助项目(SKLGIE2020-Z-4-1)

A building selection method supported by graph convolutional network

AN Xiaoya1,2, ZHU Yude3, YAN Xiongfeng4   

  1. 1. Xi'an Research Institute of Surveying and Mapping, Xi'an 710054, China;
    2. State Key Laboratory of Geo-information Engineering, Xi'an 710054, China;
    3. Guangdong Guodi Planning Science Technology, Guangzhou 510650, China;
    4. College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China
  • Received:2022-03-24 Revised:2023-08-09 Published:2023-10-12
  • Supported by:
    State Key Laboratory of Geo-Information Engineering (No. SKLGIE2020-Z-4-1)

摘要: 建筑物选取是地图综合的关键步骤,需要考虑目标大小、方向、形状、密度等多种上下文因子进行重要性评价与选取决策。现有方法大多考虑单一或少数几个因子,通过人工设置选取规则与参数,导致选取模型适应性不强。本文构建一种数据驱动的图卷积神经网络选取方法,该方法利用Delaunay三角网将建筑物目标组织为图结构,节点表示建筑物中心点,连接边体现建筑物之间的邻近关系,并计算建筑物的大小、方向、形状、密度特征作为对应节点的描述特征;然后叠加多个图傅里叶卷积运算构建图学习模型,并采用半监督学习方式训练模型,使之具备决策单个建筑物保留与否的能力。试验表明,本文方法能从少量的标注样本中有效地学习建筑物选取知识,在保留重要个体目标的同时也能较好地保持原有空间分布密度关系,克服传统方法在规则定义与参数设置方面的难题且不依赖于大量人工标注,为建筑物综合选取的智能化实施提供潜在的技术途径。

关键词: 地图综合, 建筑物选取, 图卷积神经网络, 半监督学习

Abstract: As a fundamental aspect of map generalization, building selection requires thorough consideration of various contextual factors such as size, orientation, shape, density, and more. However, many existing methods have only focused on single or a few factors, often relying on manual selection rules and parameters, which limits their practicality. In this study, we propose a data-driven building selection method using the graph convolutional network(GCN). Our method organizes buildings into a graph using Delaunay triangulation, with nodes representing building centers and edges denoting adjacent relationships between buildings. The size, orientation, shape, and density of each building are computed as the descriptive features for the associated nodes. Further, a GCN is constructed by stacking multiple graph Fourier convolutions and trained with semi-supervised learning to enable it to decide whether a building is selected or not. Experiments show that our method can effectively learn selection knowledge with few labels and perform well in maintaining the original spatial distribution densities and selecting important individual objects. It overcomes the difficulties in rule definition and parameter setting of traditional methods and does not rely on a large number of manual labels, which provides a promising solution for intelligent generalization.

Key words: map generalization, building selection, graph convolutional network, semi-supervised learning

中图分类号: