TerraVerse

Multimodal generative models for Earth observation simulation and analysis.

TerraVerse is a generative foundation model that simulates what a satellite records at the top of the atmosphere. Given a surface reflectance image (Level-2A), the acquisition geometry and the atmospheric state, it generates the corresponding top-of-atmosphere observation (Level-1C). Input and output are described by band centre wavelengths and widths, so the same model works for arbitrary optical instruments, from three-band RGB cameras to hyperspectral sensors with more than 200 bands.

L2A surface reflectance  +  atmosphere and geometry  ->  L1C top-of-atmosphere
(any band set)              (AOT, water vapour, angles)    (any requested band set)

Why

New missions need realistic simulated data before launch, and existing missions often have sparse coverage. TerraVerse learns the atmospheric and sensor forward model from real paired observations and can then generate data for instruments with few or no acquisitions. The two mission scenarios of the project are:

• Φsat-2 (MIS1): very-high-resolution multispectral imagery, simulated from Sentinel-2.
• CHIME (MIS2): the upcoming Copernicus hyperspectral mission, simulated before operations.

Approach

• Data: about 2 M paired EnMAP and EMIT L2A / L1C patches with acquisition metadata, based on SpectralEarth-MM.
• Instrument simulation: during training, every sample is shown as the view of an instrument drawn from a pool of real (Sentinel-2, Landsat 8/9, PlanetScope, SPOT-6, WorldView-3, MODIS, VIIRS, CHIME, ...) and randomly generated instruments.
• Model: a FLUX-style diffusion transformer trained as a stochastic bridge from the L2A image to the L1C image, directly in pixel and band space, with a wavelength-conditioned spectral front end.

Repositories

Consortium

Developed under ESA contract 4000150135/25/I-DT by ESA Φ-lab, DLR (German Aerospace Center), IBM Research Europe and VITO.