In a major milestone for geospatial technology, IBM and the European Space Agency (ESA) have open-sourced TerraMind, the most advanced generative AI foundation model for Earth observation to date. Now available on Hugging Face, TerraMind significantly outperforms existing models on key benchmarks while using a fraction of the compute power, making it both high-performing and highly energy efficient.
Trained on TerraMesh—the largest geospatial dataset ever assembled—TerraMind surpasses 12 leading Earth observation models by more than 8% on ESA’s PANGAEA benchmark. The model demonstrates superior performance across critical applications such as land cover classification, environmental monitoring, and change detection.
“At present, TerraMind is the best performing AI foundation model for Earth observation according to well-established community benchmarks,” said Juan Bernabé-Moreno, Director of IBM Research UK and Ireland.
Multi-Modal Intelligence for Real-World Impact
What sets TerraMind apart is its unique ability to integrate nine distinct types of Earth observation data—from satellite sensors and vegetation indices to geomorphology and spatial context. This multi-modal approach enables TerraMind to provide deep, contextual insights into complex environmental and planetary conditions.
“TerraMind combines insights from several modalities of training data to increase the accuracy of its outputs,” said Simonetta Cheli, Director of ESA Earth Observation Programmes. “It can uncover a deeper understanding of the Earth for researchers and businesses alike.”
Thinking-in-Modalities: A New AI Paradigm
A hallmark innovation of TerraMind is its “Thinking-in-Modalities” (TiM) tuning, which allows the model to generate its own training data across different types of input—pixel, token, and sequence. This any-to-any multi-modal capability makes TerraMind uniquely adaptive and data-efficient, enhancing its specialization in areas like water scarcity, biodiversity loss, and climate risk assessment.
“TiM tuning boosts data efficiency by self-generating the additional training data relevant to the problem being addressed,” explained Johannes Jakubik, IBM Research Scientist.
Scalable, Cost-Effective, and Built for Strategic Use
Not only does TerraMind deliver industry-leading performance, but it also operates with 10x less compute power than typical models—enabling sustainable, large-scale deployments. Built in collaboration with KP Labs, DLR, and the Jülich Supercomputing Center, TerraMind is poised to support a wide range of high-impact use cases including disaster response, sustainable agriculture, climate modeling, and infrastructure monitoring.
TerraMind complements existing IBM-NASA geospatial AI models such as Prithvi and Granite, and fine-tuned versions for specific applications will soon be accessible via IBM’s Geospatial Studio.
“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, ESA Earth Observation Data Scientist.
A Leap Forward in Earth Intelligence
With its unmatched blend of performance, efficiency, and adaptability, TerraMind represents a transformative leap in Earth observation capabilities. It offers researchers, governments, and businesses the ability to make more informed, timely decisions in addressing pressing global challenges—from climate resilience to resource sustainability.
Bottom Line: TerraMind is not just a better AI model—it’s a paradigm shift in how we see and understand our planet.
