A graph-based multimodal data fusion framework for identifying urban functional zone

Accurately mapping urban functional zone (UFZ) provides crucial foundational geographic information services for urban sustainable development, territorial spatial planning, and public resource allocation. UFZs are blocks within urban environments that serve specific functions, typically comprising physical objects with specific spatial distribution patterns and semantic objects of various types. However, previous studies for identifying UFZs have focused on physical or semantic aspects of UFZs, overlooking the spatial relationships and connectivity among objects. Furthermore, few have leveraged the constructed graphs by heterogeneous geospatial data to identify functional zones by street block-based mapping units. To bridge this gap, we developed a graph-based multimodal data fusion framework (G2MF) to identify UFZs. It is a fully graph-based identification framework with a feature-level fusion strategy that integrates very high-resolution remote sensing images and point of interest data. Firstly, physical objects within a UFZ unit are classified using semantic segmentation technology; then, the two independent graph structures are constructed for both physical and semantic objects within the UFZ unit; finally, the graphs are input into the proposed graph-based multimodal fusion network for UFZ identification. Experimental results show that the proposed G2MF achieves an overall identification accuracy of 88.5 % on test data from four Chinese cities and also exhibits good generalization ability on test data with geographic isolation. This study not only promotes the development of automatic UFZ identification technology but also provides new directions and methodologies for future urban big data analysis. Our source codes are released at https://github.com/yuantaogiser/G2MF.