Enhancing Geo-Transferability in Building Function Classification Through Similar Place Knowledge Retrieval

Understanding the use of buildings (i.e., building functions) is crucial for effective urban management and policy making. Recently, many deep learning-based models have been developed to identify building functions. However, their high performance depends on a large labeled dataset, which is a significant challenge for cities with limited labeled data. To close this gap, we propose a similar-place knowledge-retrieval-based framework for building function identification that enables a trained model to be directly applied/transferred to other cities without additional training needed. Our approach combines remote sensing imagery, POI data, and building footprints to capture the macro and micro built environment. Specifically, this framework includes two different modules: 1) a geo-aware transformer-based model to capture fine-grained spatial interactions between buildings; 2) a hierarchical node clustering algorithm with a similar-place knowledge-retrieval layer for portraying the coarse-grained features. This combination clusters all neighbors of a target building into a ’place’ and then conducts a coarse-grained knowledge transfer based on the similarity match. The evaluation conducted across nine cities demonstrates that the proposed model outperforms all state-of-the-art baselines, achieving an average improvement of 21.98% in cross-city predictions, as measured by Cohen’s kappa coefficient. More importantly, the ablation study shows that integrating our knowledge-retrieval layer into other deep-learning baselines significantly improves their cross-city prediction performance. Additionally, the transferability performance analysis indicates that greater similarity between the sample distributions in the source and target cities contributes to the proposed model’s higher transferability. In summary, the proposed knowledge-retrieval layer is an effective and reliable component for enhancing cross-city prediction and has strong potentials for application in cities without labeled data to support city management and sustainable urban planning.