Sam-Powered Building Footprint Updating for Global Cities: Sparse Labels Meet Historical Data Repurposing in Urban Monitoring

Abstract

Accurate building footprint databases are fundamental for sustainable urbanization yet face persistent updating challenges due to the rapid pace of urban change. Traditional methods rely on multi-temporal imagery comparisons for change detection, requiring extensive new labels to retrain models, incurring prohibitive costs. We propose a source-free updating paradigm that eliminates historical imagery dependency and reduces labeling costs by more than 95%, leveraging the Segment Anything Model (SAM). SAM is a versatile segmentation model trained on a vast dataset and has demonstrated exceptional generalization across diverse scenarios. Our method adopts a lightweight adaptation strategy and only requires a small number of samples, e.g., only 0.4% of the buildings samples. Besides, the temporal fusion module in our method can integrate historical building footprints with feature of latest imagery, which also alleviate the pressure of small sample training. The training process is conducted in a semi-supervised manner, enabling the model to learn from both labeled and unlabeled regions. The proposed methods is evaluated through: (1) Longitudinal case studies in Christchurch (postearthquake reconstruction), Beijing-Shanghai (megacity expansion), and Rio de Janeiro (Olympic-driven urban renewal ; (2) Global-scale validation across 60 cities spanning six continents. This work advances urban building updates by overcoming reliance on paired historical imagery for change detection, and large amounts of up-to-date labeling. This approach provides a scalable solution for Sustainable Development Goal 11 (Sustainable Cities and Communities) monitoring, enabling less developed countries to use free and publicly available product data to track urban expansion patterns with only a small number of labels.

Keywords: Building footprint, Semi-supervised learning, Large-scale mapping, change detection, Deep Learning