DeepVGI: Deep Learning with Volunteered Geographic Information

Recently, deep learning has been widely studied to recognize ground objects with satellite imageries. However, finding ground truths especially for developing and rural areas is quite hard and manually labeling a large set of training data is costly. In this work, we propose an ongoing research named DeepVGI which aims at deeply learning from satellite imageries with the supervision of Volunteered Geographic Information (VGI). VGI data from OpenStreetMap (OSM) and a crowdsourcing mobile application named MapSwipe which allows volunteers to label images with buildings or roads for humanitarian aids are utilized. Meanwhile, an active learning framework with deep neural networks is developed by incorporating both VGI data with more complete supervision knowledge. Our experiments show that DeepVGI can achieve high building detection performance for humanitarian mapping in rural African areas.