TerraMind is a dual-scale, any-to-any foundation model for Earth observation, developed by IBM Research, ESA Φ-lab, and Jülich Supercomputing Centre. It learns joint token- and pixel-level representations across nine modalities and outperforms other models on the PANGAEA benchmark.
