Researchers from the IMPACT AI foundation model (FM) team within the Office of Data Science and Informatics (ODSI) at NASA’s Marshall Space Flight Center in Huntsville, Alabama noticed a disconnect between the people that build FMs and the experts who need them.

The Prithvi Geospatial Foundation Model (Prithvi) is an open-source AI geospatial foundation model developed through collaboration among NASA, IBM Research, and Jülich Supercomputing Centre.

The ODSI team proposed working collaboratively with end-user scientists and decision makers throughout Prithvi’s development process to customize the model for specific applications. In December 2024, the Prithvi model expanded to include global data to support a broader range of downstream applications.

Evaluating a model’s success involves two key steps. First, researchers compare it to existing models using standardized tests to ensure fair and reproducible results. Second, experts in the field test it on practical applications to see if it improves upon current methods. Typically, AI researchers focus on technical benchmarks, while environmental experts assess real-world performance. However, the Prithvi model breaks this mold by integrating expert feedback throughout the entire process, from design to evaluation.

The Prithvi development team collaborated with six universities to identify several applications of Prithvi in the fields of disaster response, land use and crop mapping, and ecosystem dynamics monitoring. The collaboration involved researchers from Arizona State University, Boston University, Clark University, Oregon State University, Virginia Tech University, and the University of California at Berkeley.

By aligning technical development with the practical needs of the scientific community, the team ensured that the model's architecture was tailored for scientific research. This collaborative ethos not only streamlined the evaluation process but also fostered an advanced level of versatility for the model.

Disaster Response

With foundation models, scientists can now integrate multi-modal data, meaning they can process text, images, and sensor data simultaneously. Traditionally, analyzing satellite imagery for information about the impact of natural disasters has required constructing specific models to detect "building damage" or "flood detection."

Sen1Floods11 is a dataset that contains labeled surface water and land data paired with Sentinel-1 and Sentinel-2 satellite imagery to more accurately map flooding at a global scale. Flood mapping is crucial to flood risk management to help minimize the loss and damage caused by floods. Prithvi was trained with Sen1Floods11 data to classify image pixels based on water coverage. Building a flood model from scratch normally requires thousands of labeled images, but fine-tuning Prithvi requires significantly fewer, making it more accessible for regional and remote disaster response teams.

Image

The burn scars dataset used to train Prithvi contains Harmonized Landsat and Sentinel-2 (HLS) imagery of both intentional and wildfires burning from 2018 to 2021. Burn scar intensity data were also paired with HLS imagery to evaluate fire impacts on vegetation, soil, and watershed functionality to allow response teams to rapidly analyze post-fire conditions to stabilize affected landscapes and mitigate hazards. 

Image

Pond Aquaculture Mapping

Pond aquaculture mapping uses pan-sharpened 15-m Landsat 8 imagery to map brackish water pond aquaculture and mangrove forests in the coastal regions of tropical countries. This application provides a tool that can be used for mapping land cover and land use change for coastal habitats and identifying aquaculture ponds that have been established by deforesting a mangrove site. Conservation and government organizations as well as commercial suppliers can use the maps for sustainable sourcing of seafood.

“The Prithvi 2.0 model shows strong performance in predicting challenging classes like pond aquaculture in regions where the model is not exposed to. By fine-tuning the model in South East Asia, where we have a large high-quality labeled dataset, we were able to map the same classes in Mexico where reference labels are limited. This shows the value of the large-scale pre-training that Prithvi has gone through which enables global applications such as coastal habitat mapping,” said Hamed Alemohammad, associate professor in the Graduate School of Geography and director of the Center for Geospatial Analytics at Clark University.

Image

Ecosystem Dynamics Monitoring

The BioMassters dataset aimed to explore deep learning approaches for predicting yearly above ground biomass (AGB), which is a crucial climate variable that encompasses all living biomass above the soil. This information in this dataset is essential for understanding land use changes and carbon inventories.

Using datasets from FLUXNET, AmeriFlux10, and ICOS networks, Prithvi was used to estimate the gross primary productivity (GPP), or photosynthesis, the critical process of converting solar radiation into energy and materials that sustain life.

"Prithvi significantly improves the predictive skill of ecosystem carbon fluxes across a wide range of biomes compared with state-of-the-art approaches, thanks to its exceptional ability to extract and synthesize information from complex landscapes,” said Yanghui King, an assistant professor in ecosystem and agricultural science at Virginia Tech University. “As an open-source foundation model, Prithvi makes sophisticated AI far more accessible to researchers and practitioners, accelerating model development and enabling new applications that enhance the productivity and resilience of our land."

Prithvi was designed to recognize both short-term changes such as seasonal shifts and long-term patterns all while using less labeled data and being easier to scale across different projects. By making geospatial AI more accessible and efficient, this model helps bridge the gap between cutting-edge technology and real-world applications, allowing more people to benefit from AI-driven Earth observation.
“What makes Prithvi particularly powerful is not only its architectural advances and the broad geospatial knowledge it learns from global Earth observation data, but also its open science foundation, which enables researchers to openly explore, reproduce, evaluate, and extend the model,” said Wenwen Li, an Arizona State University professor specializing in Earth system and geoinformation. “Looking ahead, this openness gives Prithvi unique potential to catalyze new forms of AI-driven scientific discovery.”

Prithvi was developed as part of NASA’s Office of the Chief Science Data Officer’s 5+1 AI for Science strategy to unlock the value of NASA’s vast collection of science data using AI. NASA’s IMPACT AI FM team, based at Marshall, IBM Research, and the Jülich Supercomputing Centre, Forschungszentrum, Jülich, designed the foundation model on the supercomputer Jülich Wizard for European Leadership Science (JUWELS), operated by Jülich Supercomputing Centre. This collaboration was facilitated by IEEE Geoscience and Remote Sensing Society. The HLS dataset is funded by NASA and designed to respond to the needs of the US Federal Government identified by the Satellite Needs Working Group (SNWG) biennial survey.

For more information about NASA’s strategy of developing foundation models for science, visit https://science.nasa.gov/artificial-intelligence-science.