DOFA

DOFA (Dynamic One-For-All) is a unified vision foundation model for Earth observation from the Technical University of Munich. Instead of training one model per sensor, it uses a wavelength-conditioned dynamic hypernetwork so a single Transformer backbone adapts to optical, radar, multispectral, and hyperspectral inputs with flexible channel configurations, pretrained continually across five satellite sensor families.

The model reaches state-of-the-art transfer performance across EO downstream tasks and generalizes to unseen modalities, with pretrained weights available through TorchGeo and an open codebase. A DOFA-CLIP extension adds vision-language alignment over six EO modalities. Caveats: it is a research release, so production licensing terms deserve a check, and like all EO foundation models it still needs task-specific fine-tuning to beat hand-tuned pipelines on narrow problems.