SpectralEarth is a large-scale multi-temporal dataset designed to address the current lack of comprehensive, globally representative hyperspectral datasets, which has limited the development of foundation models in hyperspectral remote sensing. It is intended for users from the AI and data mining domain to develop self-supervised and unsupervised learning algorithms on hyperspectral imagery.

The dataset leverages data from the Environmental Mapping and Analysis Program (EnMAP), a German hyperspectral satellite mission that monitors and characterizes Earth's environment on a global scale. EnMAP delivers accurate data that provides information on the status and evolution of terrestrial and aquatic ecosystems, supporting environmental monitoring, management, and decision-making.

SpectralEarth comprises 538,974 non-georeferenced image patches covering 415,153 unique locations from 11,636 globally distributed EnMAP L2A scenes collected over two years. The patches are non-overlapping, 128x128 pixels in size, with 202 spectral bands each. The dataset has less than 10% cloud coverage. Additionally, 17.5% of the locations include multiple timestamps, enabling multi-temporal HSI analysis. The dataset includes labels for three land cover and crop-type mapping downstream tasks. These are based on the CORINE land cover database, the Cropland Data Layer (CDL), and the National Land Cover Database (NLCD). The labeled subsets can serve as benchmarks for model evaluation.