Classifying Urban Functional Zones via a Multilevel Graph Neural Network and Multimodal Geospatial Data

Urban functional zone (UFZ) classification can provide important data support for urban management. Existing studies often use multimodal data to improve UFZ classification, but the spatial context information is not sufficiently mined. An end-to-end multilevel graph neural network (MGNN) is proposed for the problem. We optimize UFZ classification using point of interest (POI) data and remote sensing images by fully exploiting the spatial context features within and between zones. We utilize Delaunay triangulation to construct remote sensing image graphs (IGs) and POI graphs (PGs) within zones at the local level, and use graph neural network (GNN) to extract graph embeddings. The embedded of IGs and PGs are interacted and fused by image-POI contrastive learning-based auxiliary (IPCA) and multitask weight-adaptive determination strategy (MWAS). The block graph between zones at the global level is constructed by K-nearest neighbor, and the spatial context features of the zones themselves and their neighbors are aggregated by Graph Sample and Aggregate, and the classification probability is predicted by SoftMax. Taking Shenyang and Chengdu as the study area, the overall accuracy and Kappa of MGNN are at least 5.27% and 4.13% higher than other methods. Our method provides an effective tool for UFZ classification.