What information would an AI model need to really understand our planet? That is the question to which IBM researchers have been proposed to answer this year, that, KP Labs, Julich Supercomputing Center (JSC) and the German Space Agency (DLR) Within the framework of a initiative led by ESA for Improve access to foundational models within the Earth Observation Community.
Therefore, IBM and ESA have presented today Terraminda new Earth observation model that the group has published in open source in Huggingface. It is the set of Greatest geospacial data to date.
Despite having been trained with 500,000 million tokens (basic text unit that an AI model uses to understand and process information.), Terramind is a small and light model, which uses 10 times less computer resources than the use of standard models for each modality. This means that users can implement it at a scale with a smaller cost, while reducing the total consumption of energy in the inference stage.
“For me, what really distinguishes Terramind is its ability to Go beyond satellite image processing with artificial vision algorithms. It has an intuitive understanding of geospatial data and our planet – confirms Juan Bernabé -Moreno, director of IBM Research in the United Kingdom and Ireland -. At present, Terramind is the founding model of The best performance for land observation according to the reference points established by the community ”.
In an ESA evaluation, Terramind It was compared with 12 popular land observation models. The evaluation showed that Terramind surpassed other models in these tasks in 8% or more.
“Terramind combines information from various training data modalities to increase the accuracy of its results – says Simonetta Cheli, director of Earth Earth Observation Programs and responsible for Esrin –. Its ability to intimutively integrate contextual information and generate scenarios never seen before is a fundamental step to unlock value of the ESA data ”.
In practice, to predict the risk of water shortage, for example, scientists must take into account many different factors, such as land use, climate, vegetation, agricultural activities and location. Before Terramind, All these data were dispersed and stored separately. Gathering this information allows users to make more precise predictions about the potential risk of water shortage based on a broader vision of conditions on Earth.
The data set you handle Terramind includes 9 million globally distributed data samplesspace-time aligned in nine main modalities: observations made by satellite sensors, the geomorphology of the land surface, the surface characteristics that are important for life on earth (vegetation and land use) and basic descriptions of locations and its characteristics (latitude, length and simple text descriptions).
“With the science and technology of observation of the Earth and international collaboration, we are releasing all the potential of the spatial data to protect our planet – says Nicolas LongepeEarth’s observation data scientist -. This project is a perfect example in which the scientific community, large technology companies and experts have collaborated to take advantage of this technology for the benefit of land sciences. Magic occurs when they bind Experts in Earth Observation Data, Automatic Learning Experts, Data Scientists and HPC Engineers”