地球信息科学学报 ›› 2016, Vol. 18 ›› Issue (9): 1153-1159.doi: 10.3724/SP.J.1047.2016.01153

• 地球信息科学理论与方法 •    下一篇

基于人工神经网络的多特征因子路网匹配算法

郭宁宁, 盛业华*(), 黄宝群, 吕海洋, 张思阳   

  1. 1. 南京师范大学 虚拟地理环境教育部重点实验室,南京 210023
    2. 江苏省地理信息资源开发与利用协同创新中心,南京 210023
  • 收稿日期:2015-12-28 修回日期:2016-03-07 出版日期:2016-09-27 发布日期:2016-09-27
  • 通讯作者: 盛业华 E-mail:shengyehua@njnu.edu.cn
  • 作者简介:

    作者简介:郭宁宁(1989-),男,硕士生,研究方向为数据集成与地图更新。E-mail: gnn0707@163.com

  • 基金资助:
    国家自然科学基金项目(41271383、41471102)江苏省自然科学基金项目(BK20151547)

Road Network Matching Considering Multiple Geometric Characteristics Based on the Artificial Neural Network

GUO Ningning, SHENG Yehua*(), HUANG Baoqun, LYU Haiyang, ZHANG Siyang   

  1. 1. Key Laboratory of Virtual Geographic Environment of Ministry of Education, Nanjing Normal University, Nanjing 210023, China
    2. Jiangsu Center for Collaborative Innovation in Geographic Information Resource Development and Application, Nanjing 210023, China
  • Received:2015-12-28 Revised:2016-03-07 Online:2016-09-27 Published:2016-09-27
  • Contact: SHENG Yehua E-mail:shengyehua@njnu.edu.cn

摘要:

在综合考虑多个特征因子的线要素匹配时,根据经验知识确定各特征因子的权值会造成人为误差。针对该问题,本文提出了基于人工神经网络的多特征因子路网匹配算法,根据线要素的几何和拓扑特性选取长度、方向、形状、距离及拓扑5个特征因子的相似度作为路网匹配参考因子。首先,分别在参考图层和待匹配图层中选取样本数据组成样本对,计算样本数据的5个特征因子相似度,用样本数据的5个特征因子相似度和样本的匹配度组成学习模式对;然后,利用BP神经网络的误差反向传播机制自动学习调整各神经层之间的连接权值;最后,输入全部数据,计算参考图层的弧段和待匹配图层的弧段间的匹配度,实现综合多特征因子的路网匹配。实验结果表明,利用人工神经网络进行综合多特征因子的路网匹配可以提高匹配效率和匹配准确度。

关键词: BP神经网络, 路网, 匹配, 多特征因子

Abstract:

The matching of line features is the premise and key technology of map conflation, which also plays an important role in change detection, digital map updating and map registration. To improve the accuracy of line feature matching, multiple geometrical characteristics and topological characteristics should be considered, such as distance, length, shape, orientation, node-degree and so on. It may produce errors when the weights of the characteristic factors are determined merely by experience. Therefore, to avoid this problem, a road network matching method considering multiple geometric characteristics based on the artificial neural network is proposed. Length similarity, orientation similarity, shape similarity, distance similarity and topological similarity (which is the node-degree similarity) are the five feature similarities discussed in this paper. They are the neurons of the input layer in BP (Back Propagation) neural network. To implement this method, first of all, samples of the reference layer and the adjustment layer are selected and the values of the five similarities for these samples are calculated. Secondly, the five feature similarities of these samples incorporating the matching rate are serving as the learning modes and are used to train the BP neural network. And the connection weights between the neural network and the threshold values of neurons are adjusted automatically. Next, buffers of the reference layer's arcs are generated. Arcs within the adjustment layer's buffers are defined as the candidate arcs. At last, the five feature similarities between each arc within the reference layer and its candidate arcs are calculated. Then, these similarities are put into the BP neural network to compute the matching rates of these arcs. If the matching rate is greater than 0.8, the relevant two arcs are regarded as an accurate matching; if the matching rate is less than 0.5, the relevant two arcs are considered to be bad matching; otherwise, they are considered to be less accurate matching which needs the interference of manual assistance to determine whether the two arcs match or not. Result shows that the adoption of BP neural network makes the road networks matching more efficient and accurate, and it avoids the assignment of proper weights to different geometrical characteristics at the same time.

Key words: BP neural network, road network, matching, multiple geometric characteristics