Abstract Building structure characteristics are essential for seismic risk assessment and management. However, accurately classifying building structural types at the city scale remains challenging due to limited data availability, spatial heterogeneity, and the difficulty in capturing contextual dependencies. In this study, a novel graph neural network (GNN)-based framework, BSPGNN (Building Structural Type Prediction using Graph Neural Networks), was proposed, integrating spatial relationships and building geometric features to improve structural type classification. A Delaunay triangulation (DT) graph was constructed from building footprint centroids to represent spatial proximity, and node features included footprint area, height, and construction year. Experiments using a real-world building dataset from Tianjin, China demonstrated that BSPGNN significantly outperformed traditional machine learning models, such as random forest (RF) and support vector machine (SVM), particularly in capturing spatially coherent patterns. The proposed model achieved a classification accuracy of 90.25% and 85.51% on the training set and validation set respectively, showing robust performance under missing data conditions. The results highlight the potential of spatial graph-based models in advancing building structural type classification for seismic risk assessment and management.
Ma et al. (Thu,) studied this question.