Less than seven months after a public/private partnership was forged between IBM Research and NASA IMPACT, the largest ever open-source geospatial AI foundation model (FM) has been released. AI foundation models are pre-trained on large datasets with self-supervised learning techniques. They provide flexible base models that can be fine-tuned for domain-specific downstream tasks. In collaboration with Clark University’s Center for Geospatial Analytics, the European Space Agency (ESA), USGS, and Oak Ridge National Laboratory, IBM and NASA IMPACT developers constructed the cloud-based foundation model using IBM’s watsonx FM stack and Cloud Vela supercomputer. To tailor the model for Earth observation analysis, NASA’s Harmonized Landsat and Sentinel-2 (HLS) imagery was used as the training data for the HLS Geospatial Foundation Model (HLS Geospatial FM). The model was released to the public on Thursday, August 3 via Hugging Face, a data science platform that enables machine learning developers to openly build, train, deploy, and share models. The HLS Geospatial FM pipelines are called Prithvi on Hugging Face and operate as a novel temporal Vision transformer.
The new foundation model has already proven valuable in several applications, including post-disaster flood mapping and detecting burn scars caused by fires. Further modifications are planned to adapt the FM for additional tasks such as predicting crop yields and monitoring greenhouse gas emissions.
More details about the HLS Geospatial FM release and the IBM/NASA partnership achievements and goals can be found in IBM’s press release and on NASA’s Earthdata website.
The Privthi pipeline can be accessed on Hugging Face.
More information about IMPACT can be found on the NASA Earthdata and IMPACT project websites.
