IBM and ESA unveil an AI model that possesses an 'intuitive' comprehension of Earth.

      IBM and the European Space Agency (ESA) launched TerraMind today, a new open-source AI model with an “intuitive” grasp of Earth. The research team claims it is the top-performing AI model for Earth observation.

      In an evaluation led by ESA, TerraMind surpassed 12 leading AI models on the PANGAEA benchmark, which is a community standard for Earth observation. The model demonstrated exceptional performance in various practical tasks, including land cover classification, change detection, and multi-sensor analysis, consistently outperforming other models by 8% or more on average.

      Juan Bernabé-Moreno, director of IBM Research UK and Ireland, stated, “What distinguishes TerraMind is its capacity to transcend mere processing of Earth observations through computer vision algorithms. It possesses an intuitive understanding of geospatial data and our planet.”

      TerraMind is a generative AI model capable of comprehending various data types—such as images, text, and time-based sequences (like climate patterns)—and recognizing connections among these different forms of information. This capability is especially valuable for managing the intricacies of a complex system like Earth.

      The model was trained on 9 million samples derived from nine diverse data types, including satellite imagery, climate records, terrain characteristics, and vegetation maps. This comprehensive dataset encompassed every region and biome on Earth and was designed to minimize bias, ensuring reliable application of the model worldwide, according to the researchers.

      ESA and IBM are furthering the application of AI in climate modeling with TerraMind, which is based on Prithvi, an open-source collection of foundational climate models introduced by IBM and NASA in 2023. The Prithvi models require significantly less computational power than conventional climate modeling software, potentially making them more environmentally sustainable.

      One of TerraMind's notable features is its “Thinking-in-Modalities” (TiM) tuning. Similar to chain-of-thought reasoning in language models, TiM enables TerraMind to self-generate additional data to enhance its performance.

      Johannes Jakubik, an IBM research scientist in Zurich, explained, “TiM tuning improves data efficiency by self-generating additional training data pertinent to the issue at hand — for instance, instructing the model to ‘think’ about land cover while mapping water bodies.”

      TerraMind was developed in collaboration with Polish spacetech company KP Labs, the Jülich Supercomputing Centre in Germany, and the German Space Agency (DLR). The model is now available as open-source on Hugging Face, with fine-tuned versions set to be released in the upcoming months.

      ESA, NASA, and IBM are not the only ones exploring AI models for climate forecasting. Google DeepMind recently introduced an AI weather forecaster that generates faster and more accurate predictions than the best existing systems.

      The EU has also engaged with this technology, unveiling a comprehensive digital twin of the Earth last year that leverages extensive data to enhance climate predictions.