Building Footprint Extraction

Building footprint extraction uses aerial and satellite imagery, and increasingly LiDAR or other sensor data, to locate buildings and recover their boundaries as masks or vector polygons. Common approaches include semantic segmentation, geometry-aware decoders and losses, instance-level offset prediction, multimodal fusion, and semi-supervised learning when labeled data are scarce. Benchmarks test performance across regions, image resolutions, seasons, lighting conditions, and settlement types, while quality-control methods can flag errors in mapped footprints. A central challenge is balancing precise polygon geometry with reliable transfer to new places and conditions, where differences in data, building forms, and evaluation measures can expose weaknesses hidden by standard benchmarks.