IBM, the European Space Agency and TerraMind are launching TerraMind today. It is a new open source AI model that has an “intuitive understanding” of Earth. According to the research group, the system is the most effective AI model for Earth observations. TerraMind, in an ESA-led assessment, beat 12 other AI models using the PANGAEA benchmark – a community standard for earth observation. The model excelled in a variety of real-world tasks including land cover classification and change detection. It outperformed other models on average by 8% or even more.
According to Juan Bernabe Moreno, director of IBM Research UK & Ireland, “What sets TerraMind Apart is its ability go beyond simply processing Earth observations with computer vision algorithm,” “It has an intuitive understanding” of geospatial information and our planet.
TerraMind, a generative AI, can recognize connections between different types of information, such as images, texts, and time-based patterns (like climate patterns). This is especially useful when dealing with a system as complex as Earth.
This model was trained using 9 million samples from nine different data types including satellite images, terrain features, climate records, and vegetation maps. The large dataset covered every biome and region on Earth. Researchers said that the model was designed to minimize bias and ensure it can be used reliably around the world.
The of EU tech
The latest rumblings on the EU tech scene
A story from our wise old founder Boris and some questionable AI artwork. Every week, it’s in your inbox for free. Sign up today!
ESA, IBM expand AI’s push in climate modelling
TerraMind was built on Prithvi – an open-source family foundational climate models launched jointly by IBM and NASA 2023. The Prithvi climate models are more environmentally friendly because they require less computing power than traditional software.
A standout feature of TerraMind is its “Thinking-in-Modalities” (TiM) tuning. TiM is similar to chain-of-thought reasoning used in language models. Johannes Jakubik is an IBM research scientist in Zurich. He said: “TiM tuning increases data efficiency by self generating the additional training datasets relevant to the problem at hand — for example, telling the model to “think” about land cover when mapping waters bodies.” TerraMind, a collaboration between the German Space Agency, Julich Supercomputing Centre, and Polish spacetech company KP Labs was developed in 19659014. The model is now open-source at Hugging Face. In the coming months, fine-tuned versions of the model will be released.
ESA NASA and IBM are not the only organisations that are experimenting with AI models to forecast climate. Google DeepMind recently released an AI weather forecaster that is faster and more accurate than the best system currently available. The EU has also been experimenting with the technology. Last year, the EU unveiled a comprehensive digital twin of Earth that uses vast amounts of data to improve climate forecasts.
