NASA has released its first open-source artificial intelligence (AI) model for analyzing satellite imagery of Earth, developed in partnership with IBM Research. The model, called the HLS Geospatial Foundation Model (FM), is trained on NASA’s Harmonized Landsat Sentinel-2 dataset.
Foundation models like this can accelerate the analysis of large datasets by enabling faster training of AI models for specific downstream applications. The HLS Geospatial FM has already shown promising results for tasks like flood mapping, burn scar detection, land cover classification, and crop mapping.
The model was trained using computing resources from IBM and NASA, including the IBM Cloud Vela supercomputer and NASA’s Science Managed Cloud Environment. Workshops conducted by NASA’s Interagency Implementation and Advanced Concepts Team (IMPACT) have shown scientists can quickly learn to fine-tune the model for new applications.
This collaborative work aligns with NASA’s Open Science Initiative to build an open and inclusive scientific community. NASA and IBM plan to continue developing AI models for Earth science, including natural language models trained on scientific literature. The release of the HLS Geospatial FM marks an important milestone in applying AI to gain insights from Earth observation data.
The model is available on Hugging Face under the name Prithvi. Scientists around the world can access and build upon this model to advance a wide range of Earth science applications. NASA hopes open-sourcing the model will accelerate innovation in using AI to understand our changing planet.
