Inferring building function: A novel geo-aware neural network supporting building-level function classification

Buildings are fundamental components of urban areas and they play a vital role in supporting human activities in daily life. Understanding the actual building functions is essential for many urban applications, such as city management, urban planning, and optimization of transportation systems. Existing studies for inferring building functions are mainly based on a building's own features, and ignore its "geographic context" (e.g., the influences of nearby buildings). This paper introduces a novel geo-aware neural network to infer the functions of individual buildings. To this end, the proposed model integrates information about the built environment and human activity of a target building and its "geographic context". The model further includes a geo-aware position embedding generator and transformer encoders to better capture the complex relationships between buildings. The evaluation results demonstrate that the proposed model outperforms all baselines and achieves a classification accuracy of 90.8%. Meanwhile, the proposed model works well even with a small amount of training dataset and has a good transferability to another urban area. In summary, the proposed model is an effective and reliable approach for inferring the functions of individual buildings and has high potential for city management and sustainable urban planning.