太阳成8722(中国)有限公司-GREEN NO.1

标题: A graph-based neural network approach to integrate multi-source data for urban building function classification

作者: Kong, B (Kong, Bo); Ai, TH (Ai, Tinghua); Zou, XY (Zou, Xinyan); Yan, XF (Yan, Xiongfeng); Yang, M (Yang, Min)

来源出版物: COMPUTERS ENVIRONMENT AND URBAN SYSTEMS  卷: 110  文献号: 102094  DOI: 10.1016/j.compenvurbsys.2024.102094  Early Access Date: MAR 2024  Published Date: 2024 JUN  

摘要: Accurately understanding the functions of buildings is crucial for urban monitoring, analysis of urban economic structures, and effectively allocating resources. Previous studies have investigated building function classification using single or dual data sources. However, the complexity of building functions cannot be fully reflected by a limited number of data sources. In addition, the functions of adjacent buildings often exhibit correlations, and contextual information between buildings has been ignored in previous studies. To address these problems, we propose a graph-based neural network (GNN) approach for building function classification that integrates multisource data and mines contextual information between buildings. This approach initially extracts four types of features related to building functions: morphological features from vector-buildings, visual features from streetview images, spectral features from satellite images, and socio-economic features from points of interest. The buildings are then modeled as a graph, where the nodes and edges represent the buildings and their proximity. Descriptive features of the nodes were obtained by concatenating the aforementioned features. Finally, the constructed graph was fed into the GraphSAmple and aggreGatE (GraphSAGE) model, which is a typical GNN model for building function classification. The experimental results showed that our approach achieved an F1score of 91.0%, which was 10.3-12.6% higher than that of the three comparison approaches. In addition, ablation experiments using different data sources revealed that the four data sources were complementary to each other and contributed to improving the building function classification. Our strategy provides an alternative and efficient solution for building function classification.

作者关键词: Building function; Classification; Multi-source data integration; Graph-based neural network

地址: [Kong, Bo; Ai, Tinghua; Zou, Xinyan; Yang, Min] Wuhan Univ, Sch Resource & Environm Sci, Wuhan 430079, Peoples R China.

[Yan, Xiongfeng] Tongji Univ, Coll Surveying & Geoinformat, Shanghai 200092, Peoples R China.

通讯作者地址: Yang, M (通讯作者)，Wuhan Univ, Sch Resource & Environm Sci, Wuhan 430079, Peoples R China.

电子邮件地址: bokong@whu.edu.cn; tinghuaai@whu.edu.cn; 2022202050037@whu.edu.cn; xiongfengyan@tongji.edu.cn; yangmin2003@whu.edu.cn

影响因子：6.8