Photos: How IBM and NASA's new geospatial model is changing our view of the world
The open-source foundation model, Prithvi-EO-2.0, was co-developed with Germany’s Jülich Supercomputing Centre and shows improved performance on a range of environmental mapping tasks.
The open-source foundation model, Prithvi-EO-2.0, was co-developed with Germany’s Jülich Supercomputing Centre and shows improved performance on a range of environmental mapping tasks.
IBM and NASA’s latest foundation model for analyzing geospatial data is here. The expanded version has improved performance on mapping tasks that involve transforming raw satellite data into pictures of Earth’s changing landscape.
At 600 million parameters, Prithvi-EO-2.0 is six times bigger than its predecessor, which was the world’s largest geospatial AI model when it was released in August 2023. It was pre-trained in collaboration with Germany’s Jülich Supercomputing Centre on a more geographically extensive dataset, and it includes embeddings that capture rich relationships in the data extending through time and space.
“Prithvi 2.0 can now better capture long-term processes season-by-season, and also assimilate high-resolution information,” said Paolo Fraccaro, an IBM researcher who led the team that developed Prithvi-EO-2.0. These innovations are detailed in a new technical report on ArXiv.
Prithvi-EO-2.0 is available on Hugging Face and IBM’s TerraTorch toolkit as part of NASA’s Office of the Chief Science Data Officer’s initiativemake satellite data more accessible through new technologies. Trained on 4.2 million data points from NASA’s Harmonized Landsat and Sentinel (HLS) dataset, Prithvi-EO-2.0 outperformed other geospatial models on the GEO-bench framework, which evaluates how well AI models can classify and segment satellite images.
HLS data is captured at a resolution of 30 meters per pixel, but Prithvi 2.0 can zoom in much closer, picking out crops and tree-species in satellite images, and cattle and solar panels in drone data. Overall, Prithvi 2.0 achieved an average score of 75.6% on GEO-bench, an 8% improvement over the previous model.
IBM researchers attribute the improvement to the early involvement of NASA experts who helped design the model, gather its training data, and measure its performance. NASA experts also introduced new applications, including carbon flux estimation, landslide detection, and land cover classification, at different scales.
“We hope that this spirit of open collaboration can be a model for developing other tools aimed at unlocking very large datasets,” said Juan Bernabé-Moreno, the IBM Research Europe director who also leads research initiatives in climate and sustainability.
To recognize this latest milestone, IBM researchers applied Prithvi to climate-related tasks on three continents — in the Bolivian Amazon, in coastal Spain, and a northeastern U.S. city. These are just a few examples of how Prithvi can help us better understand our evolving planet.
Trees have an outsize influence on the health of our planet. They absorb excess carbon dioxide from the atmosphere, help to regulate climate, and provide habitat for 80% of land-dwelling species. Stashed in their leaves and roots, and the soils below, are 860 gigatons of carbon, the equivalent of a century’s worth of industrial carbon emissions.
Globally, forests are under increasing pressure from a growing, urbanizing population. In the tropics, demand for timber and agricultural land has led to a steady chipping away of rainforests which are home to millions of species found nowhere else. A carbon-credit system that pays people to protect the forest instead of cutting it down has been proposed as one way to reduce deforestation. But the system can only really work if there’s a way to verify that the forest stays intact.
Toward this end, IBM fine-tuned Prithvi to estimate canopy height from NASA and European Space Agency (ESA) satellite images. In this animation spanning 2016 to 2024, you can see how a swath of land in Bolivia transitions from tree canopy in dark green to developed areas in white and light green corresponding to roads, buildings, and crops. The fine-tuned canopy height model is now available through IBM Granite.
Torrential rain in eastern Spain on Oct. 29, 2024, triggered flash flooding that claimed more than 200 lives. In just eight hours, more rain fell on the Valencia region than typically falls in a year. IBM mapped the extent of the floods using a geospatial foundation model built with NASA and fine-tuned on NASA Sentinel satellite data for Spain.
Bad weather partially obscured Sentinel-2’s view of the flooding, but Sentinel-1’s radar imaging pierced the clouds to capture the full scope of the damage. Combining data from both satellites, IBM and NASA’s AI model put the flood’s extent at about 120-square kilometers. IBM researchers estimate that the accuracy of their map is about 30% greater than a traditional map drawn with Sentinel-2 data alone. The new flood detection model is now available through IBM Granite.
Climate change is raising temperatures globally, but cities especially are feeling the heat. Buildings and paved surfaces trap energy from the sun, creating ‘heat islands’ that can feel 20°F hotter than places with more trees and vegetation.
Baltimore is one of the cities most at risk of extreme heat in the United States, according to the nonprofit research center, Climate Central. This past summer, the city experienced several bouts of prolonged heat that not only endangered public health but strained the power grid, contributing to blackouts.
As the planet continues to warm, tracking heat in places like Baltimore will be vital for immediate response efforts as well as longer term heat-mitigation measures like tree planting. While NASA’s Landsat satellite captures land-surface temperatures at a scale of 30 meters per pixel, it only passes over Baltimore twice a month. Other NASA satellites pass by daily, but their coarse-grained resolution misses important details.
Prithvi provides the best of both worlds with hourly maps at the zoomed-in scale of Landsat images. In the above animation from July 30 to August 1, 2023, Baltimore sustained a three-day heat wave while leafier suburbs stayed relatively cool. The fine-tuned land surface temperature model is now available through IBM Granite.