In response to the increasing challenges posed by global extreme weather events, NASA has teamed up with IBM Research to develop a novel AI geospatial foundational model based on artificial intelligence and open science principles.
According to recent announcements from NASA, the space agency has partnered with IBM Research to develop a new AI geospatial foundational model based on artificial intelligence and open science principles. This model, trained on NASA's Harmonized Landsat and Sentinel-2 data (HLS), not only provides additional data sources and support for scientific research but also enhances the capability for early prediction of extreme weather events.
Manil Maskey, the Data Science Lead at NASA's Office of the Chief Science Data Officer (OCSDO), described this model as, "The foundational model understands what's represented in the data; it's like a Swiss Army knife that can be used for various situations and handling different things." This versatility means that once the foundational model is created, it can be trained on small amounts of data to perform specific tasks, thereby enabling robust and efficient solutions.
So far, the Interagency Implementation and Advanced Concept Team (IMPACT) and its collaborators have demonstrated the potential of this geospatial foundational model in various fields through fine-tuning and application, including monitoring forest fires, delineating flood zones, and classifying crops and other land use types.
Given that creating the initial foundational model requires massive amounts of data and powerful computing capabilities, the collaboration between NASA and IBM is crucial. In this project, NASA provides data and scientific expertise while IBM contributes computational power and specialized technical expertise in AI algorithm optimization. This collaboration not only accelerates the development process of the model but also ensures that research outcomes can benefit more researchers through the principles of open science.
Manil Maskey emphasized, "To build a large-scale foundational model, we realized early on that relying solely on efforts from one organization would not be successful. This is why we keep the foundational model open to the public, maintaining transparency from pre-training data, code, to practices, model weights, fine-tuning training data, and publications, enabling researchers to trace the reasons for using certain data or model architectures."
Building upon the success of the geospatial foundational model developed by NASA and IBM Research, the two entities continue their collaboration to create a new similar model for weather and climate research. The collaborators for this new project also include Oak Ridge National Laboratory (ORNL), NVIDIA, and several universities. The primary dataset for this project is the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), a vast atmospheric reanalysis dataset spanning from 1980 to the present. Similar to the geospatial foundational model, this new weather and climate model will be developed using open science methods and will be made available to the public in the near future.
