Abstract
Automated methods for building function classification are becoming necessary due to restricted access to accurate building use data. Traditional on-site surveys conducted by government agencies are costly and can be influenced by subjective judgment, highlighting the need for more objective and cost-effective approaches. Existing approaches utilize Natural Language Processing (NLP) techniques such as text similarity and topic modeling, which typically struggle with the ambiguity of semantic contexts in textual data representing human activities. This study introduces a method for classifying urban building functions by integrating physical and spatial metrics with contextual embeddings from OpenStreetMap (OSM) tags, employing Large Language Models (LLMs) to improve the precision and relevance of function classifications in urban settings. We employed an XGBoost model trained on 32 features from six city datasets to classify urban building functions, demonstrating varying F1 scores from 67.80% in Madrid to 91.59% in Liberec. Integrating LLM embeddings enhanced the model's performance by an average of 12.5% across all cities compared to models using only physical and spatial metrics, and by 6.2% over models that incorporate direct tags from OSM. This suggests that deep contextual understanding is beneficial for classification. Moving forward, we suggest investigating the discrepancies in classification accuracy across different urban contexts, which is a common occurrence in existing research.