ORBITaL-Net: A labeled training library for large-scale building feature extraction

Over the course of several years, nearly 1.5 million building outlines have been created from approximately 128,000 training tiles covering roughly 7,000 km² of very high-resolution multispectral overhead imagery, primarily dated between 2010 and 2020. This dataset, dubbed the Oak Ridge Building Image and TrAining Label Net (ORBITaL-Net), is designed for machine learning applications and is global in scope, with samples drawn from 72 countries across North America, South America, Africa, Europe, and Asia. ORBITaL-Net captures a great diversity in geographic setting, structural characteristics, land use (urban and rural), terrain, and imagery conditions. While the labeled building outlines are themselves valuable, the dataset's true strength lies in the pairing of these labels with corresponding reference imagery, which is being released for open source use. Similar to SpaceNet and Replicable AI For Microplanning (ramp), this building outline dataset will allow the larger computer vision community from academia, government, and industry the opportunity to develop robust, scalable, and generalizable geospatial machine learning techniques. Unlike SpaceNet and ramp, which offer high resolution labels and imagery primarily for large urban cities, ORBITaL-Net is not focused on training samples from heavily populated areas but instead aims to capture the innate variability of conditions present in both the physical environment and imagery collections.