Urban Building Function Mapping by Integrating Cross-View Geospatial Data

In the face of rapid urbanization, traditional methods for mapping urban functional zones prove inadequate in both efficiency and scale. This study introduces a fully automatic and cost-effective strategy for recognizing and mapping building functions using cross-view geospatial data. Utilizing deep learning object detection and instance segmentation models, we extracted detection boxes and functional prediction results in street view images. Subsequently, we cooperated the Segment Anything Model and a post-processing module to derive high-precision polygons for building footprints. We further improved the cross-view alignment method from our previous work to align cross-view data of building functions predictions and footprint polygons, thereby mapping the building functions. Experimental results demonstrate that our strategy can generate high-precision polygons for building footprints based on satellite images, and accomplish cross-view alignment and mapping. The alignment accuracy achieved 76.60%, and the overall mapping precision for building functions surpasses 60%. This study is expected to contribute to urban planning and management by providing a novel methodology for cost-effective, dynamic analysis of urban spatial patterns.