Why Foundation Models Are Transforming Earth Observation

Artificial intelligence has become an essential tool for unlocking the value of this information. Over the past decade, AI has transformed how we analyse satellite imagery, enabling more accurate mapping, environmental monitoring and disaster response. Yet many of today’s models are still designed for a single purpose. A model trained to estimate forest biomass, for example, cannot easily be applied to flood detection or urban monitoring without significant retraining and new labelled datasets.

The next generation of Earth observation is moving beyond this task-specific approach.

Foundation models are changing the way AI learns from Earth observation data. Rather than being built for one application, these models are trained on vast collections of multimodal satellite imagery to learn general representations of the Earth’s surface and its dynamic processes. Once trained, they can be adapted to support a wide range of environmental challenges with far less additional training, making them more scalable, efficient and transferable across regions and applications.

ThinkingEarth is developing the first Copernicus Foundation Models, using data from the Sentinel satellite missions together with self-supervised learning, graph neural networks, causal AI and physics-aware machine learning. Rather than viewing environmental challenges in isolation, the project treats the Earth as a connected system, where interactions between climate, biodiversity, land use, energy and human activity can be understood together.

A key innovation is the concept of Earth as a Graph, where locations, ecosystems and Earth system variables are linked through relationships that evolve over space and time. This allows AI not only to identify patterns but also to understand how changes in one part of the Earth system can influence another, leading to more robust predictions and deeper scientific insights.

These advances have practical benefits across a range of sectors. ThinkingEarth is applying its foundation models to challenges including renewable energy forecasting, urban biodiversity monitoring, forest carbon assessment and food security. By creating AI models that can be adapted to multiple downstream applications, the project reduces the need to develop new models from scratch for every environmental problem, accelerating innovation while making Earth observation more accessible and impactful.

As the demand for timely, reliable environmental intelligence continues to grow, foundation models represent a significant step forward. By combining advanced AI with Europe’s Copernicus programme, ThinkingEarth is helping build a future where Earth observation is not only more powerful, but also more connected, adaptable and capable of supporting better decisions for people and the planet.