CrossGeoNet: A Framework for Building Footprint Generation of Label-Scarce Geographical Regions

Building footprints are essential for understanding urban dynamics. Planet satellite imagery with daily repetition frequency and high resolution has opened new opportunities building mapping at large scales. However, suitable methods scarce less developed regions, as these regions lack massive annotated samples to provide strong supervisory information. To address this problem, we propose learn cross-geolocation attention maps in a co-segmentation network, which is able improve the discriminability of buildings within target city more general representation different cities. In way, limited information resulting from insufficient training examples cities can be compensated. Our method termed CrossGeoNet, consists three elemental modules: Siamese encoder, module, decoder. More specifically, encoder learns feature pair images two geo-locations. The cross-location module aims learning similarity based on global overview common objects (e.g., buildings) decoder predicts segmentation masks using learned original convolved images. proposed evaluated datasets spatial resolutions, i.e., dataset (3 m/pixel) Inria (0.3 m/pixel), collected various locations around world. Experimental results show that CrossGeoNet well extract sizes alleviate false detections, significantly outperforms other competitors.