Urban building function classification based on multisource geospatial data: a two-stage method combining unsupervised and supervised algorithms

Urban buildings serve not only as spaces for residence and work but also host diverse functions such as commerce and education. Identifying the functions of urban buildings plays a crucial role in urban planning and management. This study proposes a hierarchical framework for urban building function classification based on Points of Interest (POI) and OpenStreetMap (OSM) road vector data, combining both unsupervised and supervised methods in two stages (BFC-CUS, Building Function Classification by Combining Unsupervised and Supervised Algorithms). First, in the unsupervised classification stage, constrained by OSM road data, buildings are clustered based on their geometric features, with the resulting building clusters serving as classification units. The kernel density values of different categories of POIs are utilized for the preliminary classification of these building clusters. Second, in the supervised classification stage, the average geometric features and POI kernel density values of building clusters are used as attribute features to train a classifier, and the remaining building clusters' functions are identified by the trained classifier. The experimental study was conducted in Beijing and Shanghai, yielding overall accuracy of 83.9% and 81.8% for building function classification in the respective cities. Furthermore, the study analyzes the landscape patterns of different function categories of buildings in Beijing and Shanghai. This research offers promising prospects for the function classification of large-scale urban buildings and holds significant implications for planning urban development based on the analysis results.