CHARACTERIZING RESIDENTIAL BUILDING PATTERNS IN HIGH-DENSITY CITIES USING GRAPH CONVOLUTIONAL NEURAL NETWORKS

In urban morphology studies, accurately classifying residential building patterns is crucial for informed zoning and urban design guidelines. While machine learning, particularly neural networks, has been widely applied to urban form taxonomy, most studies focus on grid-like data from street-view images or satellite imagery. Our paper provides a novel framework for graph classification by extracting features of clustering buildings at different scales and training a spectral-based GCN model on graph-structured data. Furthermore, from the perspective of urban designers, we put forward corresponding design strategies for different building patterns through data visualization and scenario analysis. The findings indicate that GCN has a good performance and generalization ability in identifying residential building patterns, and this framework can aid urban designers or planners in decision-making for diverse urban environments in Asia.