IBM is collaborating with the European Space Agency (ESA) to develop an AI system that uses satellite data to monitor climate change and other global challenges, such as water scarcity, in real time.
According to IBM, the new system, called TerraMind, is the most advanced Earth observation model ever created. IBM trained the open-source model using TerraMesh, the largest set of freely accessible geospatial data.
IBM reports that TerraMind is a multimodal model that uses an innovative encoder-decoder architecture based on symmetric transformers. The model can handle pixel, token, and sequence-based inputs. It is capable of learning correlations between different data forms. Although it has been trained on more than 500 billion tokens of data, it is relatively small and lightweight. It consumes about ten times less computing power per modality than existing models. This makes it possible to scale TerraMind at significantly lower costs.
Wide range of factors
According to IBM, TerraMind’s multimodal capabilities are essential for Earth observation tasks. This is because many factors must be taken into account when assessing the impact of climate change.
For example, when predicting the risk of water scarcity, factors such as climate, temperature, precipitation, and agricultural activities must be taken into account. Previously, researchers used different models to study each of these factors separately. TerraMind brings all this information together in a single model. This provides a more complete picture of what is happening at a given location.
Juan Bernabé-Moreno, director of IBM Research UK and Ireland, says that TerraMind stands out because it goes beyond simply processing Earth observations with computer vision algorithms.
Minimal bias
To compile the TerraMesh dataset that forms the basis of TerraMind, IBM researchers collected data on biomes, land use, land cover types, regions, and other factors. Their aim is to ensure that the model can make predictions worldwide with minimal bias.
TerraMesh contains more than nine million globally distributed and spatially and temporally aligned data samples across nine core modalities. These include observations from ESA satellites, information about surface structures such as vegetation and land use, and geometric data describing locations and characteristics, such as latitude and longitude, elevation, and simple text descriptions of regions.
IBM claims that tests show TerraMind to be the best-performing fundamental model for Earth observation ever created. The ESA led these evaluations, comparing TerraMind with twelve other models in its class on the PANGEA benchmark. This is a standard used to measure AI performance on realistic tasks. Think of land cover classification, change detection, environmental monitoring, and analysis across multiple time points. According to ESA, TerraMind outperformed competing models by at least 8% in all PANGEA tests.
Model generates its own training data
TerraMind also uses a new technique called Thinking-in-Modalities tuning, which is similar to the Chain-of-Thought method used in reasoning models such as OpenAI’s GPT-4o and DeepSeek’s R1. This approach, abbreviated as TiM, involves the model independently generating synthetic training data from other modalities to further improve its performance beyond what would be possible with regular fine-tuning.
IBM said TerraMind can be used to solve long-term problems such as disaster management, environmental monitoring, precision agriculture, urban and regional planning, critical infrastructure monitoring, forest management, and biodiversity monitoring, as well as short-term problems such as monitoring forest fires and floods.
Read also: IBM aims for simplicity with AI agents and automation