| Summary: | In 2023, NASA teamed up with IBM Research to create an AI geospatial foundation model. Trained on vast amounts of NASA's widely used Harmonized Landsat and Sentinel-2 (HLS) data, the model provides a base for a variety of AI-powered studies to tackle environmental challenges. In keeping with open science principles, the model is freely available for anyone to access. Foundation models serve as a baseline from which scientists can develop a diverse set of applications, enabling powerful and efficient solutions. "Foundation models only know what things are represented in the data," explained Manil Maskey, the data science lead at NASA's Office of the Chief Science Data Officer (OCSDO). "It's like a Swiss Army Knife—it can be used for multiple different things." Once a foundation model is created, it can be trained on a small amount of data to perform a specific task. To date, the Interagency Implementation and Advanced Concept Team (IMPACT) along with collaborators have demonstrated the geospatial foundation model's capabilities by fine-tuning it to detect burn scars, to delineate flood water, and to classify crop and other land use categories. Because of the computational resources required to create the initial foundation model, a partnership was necessary for success. In this case, NASA brought the data and scientific knowledge, while IBM brought the computing power and AI algorithm optimization expertise. The team's shared commitment to making their research accessible through open science principles ensures that their model can be useful to as many researchers as possible. "To build a foundation model at scale, we realized early on that it's not feasible for one institution to build it," Maskey said. "Everything we have done on our foundation models has been open to the public, all the way from pre-training data, code, best practices, model weights, fine-tuning training data, and publications. There's transparency, so... |