OpenEarthMap consists of 2.2 million segments from 5,000 aerial and satellite images, covering 97 regions in 44 countries across 6 continents, with manually annotated land cover labels for 8 categories, with a ground sampling distance of 0.25-0.5 meters. Semantic segmentation models trained on OpenEarthMap generalize globally and can be used as off-the-shelf models in a variety of applications. OpenEarthMap promotes research including but not limited to semantic segmentation and domain adaptation.

Related papers and resultsOpenEarthMap: A Benchmark Dataset for Global High-Resolution Land Cover Mapping" has been included in WACV 2023.

Classes and Annotations

The dataset provides annotations for eight categories: bare land, pasture, developed space, roads, trees, water, agricultural land, and buildings. All annotations were done manually, taking an average of 2.5 hours per image.