Anomaly Detection Dataset (ADD)

We construct an Anomaly Detection Dataset (ADD) on diverse earth surface phenomena. The benchmark consists of single-date and bi-temporal optical image chips, each cropped to 512×512 pixels, with ground sampling distances (GSD) (spatial resolutions) ranging from approximately 10 m to 0.1 m. All scenes are drawn from diverse regions worldwide, covering different continents, environments, and land-use types, so that anomalies appear under a wide variety of backgrounds and imaging conditions. Pixel-level anomaly masks are manually delineated by visual interpretation and are used only for evaluation.Specifically, the benchmark comprises three subsets aligned with the spectral, structural, and behavioral anomaly categories. The spectral anomaly subset contains 100 single-date RGB images from Google Earth, centered on objects or patches with strong radiometric contrast relative to their surroundings. The structural anomaly subset contains 80 single-date Google Earth RGB images in which targets are primarily characterized by textural or configurational irregularities against otherwise regular patterns. The behavioral anomaly subset includes 20 pre-post image pairs over disaster-affected areas released by the Maxar Open Data Program (https://www.maxar.com/open-data): for each pair, co-registered pre-event and post-event optical scenes are provided, and anomalies are defined as damage or impact footprints that emerge between the two time points.