The foundation-model wave reached Earth observation in April: ESA and IBM released TerraMind, an open-source AI model designed to understand the planet — not just classify its pixels.
The distinction matters. A conventional model can label a satellite image "forest" or "flood". TerraMind was trained on a dataset of more than nine million spatiotemporal samples spanning biomes, land use and atmospheric conditions, giving it what its creators call an intuitive sense of geospatial context: why a river looks like a river, wherever on Earth it flows.
Foundation models change the economics of applied AI. Instead of collecting labeled examples for every task — illegal logging here, algal blooms there — one pretrained model is fine-tuned to each problem with a fraction of the data. For a domain where labels are scarce and expensive, that is the unlock.
That it is open-source is the second unlock. Researchers, startups and national agencies can run TerraMind themselves, inspect its weights, and adapt it to their own regions — a deliberate play to build a European ecosystem of EO applications rather than a single product.
The pattern echoes what worked elsewhere in AI: train big on generic data, adapt cheaply to specifics. Earth observation may be the domain where that recipe delivers the most public good — agriculture, disasters, climate — per GPU-hour invested.
TerraMind also landed in a rapidly crowding field — Prithvi at NASA, SatMAE in academia — and the competition is healthy: each open model raises the baseline everyone fine-tunes from, and the ultimate winners are the small teams who never could have trained one from scratch.