地球信息科学学报

地球信息科学学报 ›› 2022, Vol. 24 ›› Issue (1): 63-73.doi: 10.12082/dqxxkx.2022.210340

• 第二届中国空间数据智能学术会议(SpatialDI 2021)优秀论文 • 上一篇    下一篇

样本数量不平衡下的建筑群模式识别方法研究

令振飞1,2,3(), 刘涛1,2,3,*(), 杜萍1,2,3, 张耀蓉1,2,3, 杨国林1,2,3, 锁旭宏4   

  1. 1.兰州交通大学测绘与地理信息学院,兰州 730070
    2.地理国情监测技术应用国家地方联合工程研究中心,兰州730070
    3.甘肃省地理国情监测工程实验室,兰州 730070
    4.中交一航局第二工程有限公司, 青岛 266071
  • 收稿日期:2021-06-16 修回日期:2021-07-16 出版日期:2022-01-25 发布日期:2022-03-25
  • 通讯作者: * 刘涛,博士,湖北随州人,教授,主要从事空间关系理论,GIS与遥感一体化,GIS、RS应用与开发等研究。E-mail: ltaochina@foxmail.com
  • 作者简介:令振飞（1996— ）,男,甘肃天水人,硕士生,研究方向为空间模式分析与地理深度学习、空间场景相似性检索。E-mail: 0219757@stu.lzjtu.edu.cn
  • 基金资助:
    国家自然科学基金项目(41761088);国家自然科学基金项目(42061060);兰州交通大学优秀平台支持(201806);兰州交通大学天佑创新团队(TY202001)

Pattern Recognition of Regular Buildings with Unbalanced Sample Size

LING Zhenfei1,2,3(), LIU Tao1,2,3,*(), DU Ping1,2,3, ZHANG Yaorong1,2,3, YANG Guolin1,2,3, SUO Xuhong4   

  1. 1. Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
    2. National-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring, Lanzhou 730070, China
    3. Gansu Provincial Engineering Laboratory for National Geographic State Monitoring, Lanzhou 730070, China
    4. First Navigation Bureau Second Engineering Company Limited, Qingdao 266071, China
  • Received:2021-06-16 Revised:2021-07-16 Online:2022-01-25 Published:2022-03-25
  • Contact: LIU Tao
  • Supported by:
    National Natural Science Foundation of China(41761088);National Natural Science Foundation of China(42061060);Excellent platform support of Lanzhou Jiaotong University(201806);Tianyou innovation team of Lanzhou Jiaotong University(TY202001)

RichHTML

PDF (PC)

赞

摘要：

通过综合空间关系、几何、语义等多种特征对建筑群模式进行智能识别,在地图多尺度表达和数字化制图等领域有着重要的意义。将图卷积神经网络用于建筑群模式智能识别能够克服传统方法依赖人工经验设置参数、制定的规则过于严格等缺点。但是该方法往往存在样本比例失衡的问题,容易导致样本数量较少的类别无法正确识别。本文首先以建筑物质心作为样本进行聚类获得建筑群组。其次,将同一群组内的建筑物质心作为样本进行Delaunay三角剖分来构建建筑群的图结构,其中图节点特征选取能够描述建筑群几何特征的面积、大小、方向等相关指标。再次,通过图结构过采样的方式对样本数量较少的建筑群图结构进行增强,然后将样本数量较少的建筑群图结构增强前后的数据分别输入图卷积神经网络模型进行训练,并结合ROC曲线等多个评价指标对模型的性能进行了评测。实验结果表明,对样本数量较少的建筑群图结构增强之后,模型对于样本数量较少的建筑群识别准确率有了明显的提高。

关键词: 建筑群聚类, 图卷积神经网络, 模式识别, ROC曲线, 图结构过采样, Delaunay三角网, 拉普拉斯矩阵

Abstract:

It is of profound significance in the fields of multi-scale map expression and digital mapping to recognize the architectural pattern intelligently by integrating spatial relations, geometry, semantics, and other features. Applying graph convolution neural network to intelligent recognition of building pattern can overcome the shortcomings of traditional methods, such as relying on artificial experience to set parameters and making rules too strict. However, this method often has the problem of sample proportion imbalance, which easily leads to the category with a small number of samples can not be properly identified. In this paper, the building centriod is used as the sample to cluster to obtain the building groups. Secondly, the building centriod in the same group is used as the node to construct the graph structure of the building group of Delaunay triangulation. The feature selection of the graph node can describe the area, size, direction, and other related indicators of the geometric characteristics of the building group. Thirdly, the building groups graph structure with small number of samples is enhanced by oversampling the graph structure. Then the data before and after the enhancement of the architectural group graph structure with small number of samples are input into the graph convolution neural network model for training, and the performance of the model is evaluated with several evaluation indexes such as ROC curve. The experimental results show that the recognition accuracy of the model for buildings with lesser number of samples is significantly improved after the structure of buildings with small number of samples is enhanced.

Key words: Building cluster clustering, graph convolutional neural network, pattern recognition, ROC curve, graph structure oversampling, delaunay triangulation, laplacian matrix