TerraMind

TerraMind summarises satellite imagery as numbers, like other Earth observation foundation models, and it can also generate one kind of layer from another: a radar image or a land-use map from an ordinary Sentinel-2 picture, for instance.^1,2 A generated layer is the model's guess at what that layer would show, and the model card warns against reading it as a reconstruction.³ On the project page, a Sentinel-2 scene of Singapore becomes a radar image in which ships and runways show and the clouds are gone.¹

It was built by IBM, the European Space Agency's Φ-lab and the FAST-EO project, and released on Hugging Face on 22 April 2025.^1,4 KP Labs, the Jülich Supercomputing Centre, where it was trained, and the German Aerospace Center (DLR) also took part.^1,4 The paper appeared at ICCV 2025, a computer vision conference.^1,2 It is a separate model from the Prithvi models IBM makes with NASA (AlphaEarth / Prithvi); IBM's announcement names both as part of its work applying AI to the planet, and says NASA experts helped validate TerraMind.⁴

It was trained on TerraMesh, nine million matched samples from around the world in nine kinds of data.⁴ These include Sentinel-1 radar, Sentinel-2 optical imagery, land use and land cover, a vegetation index, elevation, coordinates and short text descriptions.^1,4 As an input it accepts Sentinel-2 at two processing levels, Sentinel-1 in two radar formats, an elevation model or a plain colour image, and it copes when some of these are missing.³

Run as an encoder on your own imagery, the base model gives 768 numbers for each 16 by 16 pixel patch of a 224 by 224 pixel input, averaged across the input layers by default.³ Fine-tuning goes through IBM's TerraTorch toolkit, with example settings for flood mapping, burn scars and crop types, and multi-date data goes through a time-series wrapper in that toolkit.² An option called Thinking-in-Modalities has the model first imagine a missing layer, land cover by default, and then use it to improve its answer.^1,2

Code and weights are under Apache 2.0 and come in tiny, small, base and large sizes.^2,3