IBM and the European Space Agency (ESA) on Tuesday unveiled TerraMind, a cutting-edge generative artificial intelligence model for Earth observation that they say outperforms all existing alternatives on community-standard benchmarks.

Developed with partners including KP Labs, Jülich Supercomputing Center, and the German Space Agency, the model has been open-sourced on the AI repository Hugging Face.

The new foundation model, trained on TerraMesh — the largest geospatial dataset ever created — integrates insights from nine different types of Earth observation data.

It is designed to help researchers and governments better understand and respond to environmental challenges, including land degradation, climate change and natural disasters.

“TerraMind is not just processing data — it understands it,” said Juan Bernabé-Moreno, director of IBM Research UK and Ireland. “It outperforms other Earth observation foundation models by over 8 percent on real-world tasks like land cover classification and change detection.”

Based on a symmetric transformer-based encoder-decoder architecture, TerraMind can simultaneously ingest pixel-based, token-based, and sequence-based inputs.

Despite being trained on 500 billion tokens, the model is described as lightweight, requiring significantly less computational power than traditional approaches.

Breakthroughs in Model Architecture and Data Efficiency

In benchmark testing conducted by ESA using the PANGAEA suite, TerraMind topped a field of 12 Earth observation AI models in accuracy and flexibility.

The model was trained using spatiotemporally aligned data samples drawn from all global biomes and land types, with contributions from various ESA missions and scientific institutions.

Simonetta Cheli, ESA’s director of Earth Observation Programmes, said TerraMind’s ability to intuitively integrate contextual information sets it apart: “This is a critical step in unlocking the value of ESA data. TerraMind enables businesses and researchers to uncover a much deeper understanding of Earth systems.”

Among its most innovative features, TerraMind includes a self-tuning capability called “Thinking-in-Modalities,” or TiM, which allows the model to generate new training data across modalities.

IBM scientists liken the approach to the “chain-of-thought” prompting in large language models.

According to IBM Research scientist Johannes Jakubik, TiM tuning enhances the model’s accuracy when specialized for use cases such as water scarcity prediction or wildfire risk assessment.

A Global Collaboration for Earth Science Innovation

While IBM and NASA have previously released geospatial models through the Granite and Prithvi initiatives, TerraMind represents a leap forward in precision, scalability and data integration. It was developed with the support of the Jülich Supercomputing Center and validated in part through NASA’s Open Science initiative.

In the coming weeks, IBM plans to release fine-tuned TerraMind variants for disaster response and other high-impact use cases via its Granite Geospatial repository.

“With Earth observation science, technology, and international collaboration, we are unlocking the full potential of space-based data to protect our planet,” said Nicolas Longepe, Earth Observation Data Scientist at ESA. “This project is a perfect example of what’s possible when the scientific and technology communities join forces.”