Satellite Image Segmentation for Building Detection using U-net

Automatically detecting buildings from satellite images has a lot of potential applications, from monitoring movements of populations in remote areas to evaluating the available surface to implant solar panels on roofs. We develop a Convolutional Neural Network for the extraction of buildings from satellite images, adapted from a U-net originally developed for biomedical image segmentation. We train our model on satellite images and on ground-truth labels extracted from OpenStreetMap. We show that our model achieves a reasonable level of accuracy, though slightly lower than state-of-the-art, and outline some ideas for further improvements.