Fine-grained attributes mapping of individual buildings from street-view and satellite imagery

Buildings are essential components of urban land resources and smart city infrastructure. Accurate mapping of their attributes is vital for effective urban management. However, existing methods often suffer from limited attribute diversity, dependence on open-source footprint data, and challenges in robust cross-view alignment. This study presents the Cross-View Building-level Mapping (CVB-Mapping) framework, which maps fine-grained attributes of individual buildings by leveraging spatially aligned street-view and satellite image pairs. We first construct a new dataset based on OmniCity, adding three underexplored attributes (i.e. number of floors, year built, and floor area ratio) alongside land use, and establish benchmark experiments. We then propose Segment Anything Model-based Building Footprint Extraction (SAM-BFE), a novel method for extracting large-scale building footprint polygons from satellite images using the Segment Anything Model (SAM), augmented by prompt generation and post-processing method. Additionally, we introduce a cross-view alignment approach to align street-view-predicted attributes with satellite-derived footprints under uncertainty. Experimental results demonstrate that CVB-Mapping surpasses state-of-the-art satellite-based and cross-view methods. It improves accuracy by 4% to 22% over the best-performing cross-view baselines in predicting land use and capacity-related attributes. This work establishes CVB-Mapping as an effective framework for detailed building attribute mapping across cities. The dataset, code, and trained models have been made publicly available at: https://github.com/Daniel-Chender/CVB-Mapping.