Building clustering method that integrates graph attention networks and spectral clustering

Clustering is an unsupervised learning method; however, it cannot be directly applied to complex building features to reveal the deep-level intrinsic relationships between buildings and assimilate the cartographic experience of experts. Additionally, defining the number of clusters is a problem that remains to be resolved. Therefore, this study proposed a new building clustering method that integrates a graph attention network (GAT) and spectral clustering in two steps. First, supervised graph classification was performed. Under the constraints of road networks, buildings within the region enclosed by the road networks were constructed as multiple graph structures. The GAT was introduced to learn the cartographic experience, thereby completing the graph classification task. Second, the buildings are clustered using spectral clustering by constructing a new graph structure to more scientifically detect the spatial proximity of buildings. The result shows that the method proposed in this study is more in line with the real labels.