Skip to main content


  • IBM and NASA released an open-source AI model named Prithvi, intended to analyze Earth’s satellite imagery.
  • The model, consisting of 100 million parameters, was trained using US space agency’s HLS program’s one-year worth of images.
  • Prithvi’s 3 versions identify flooding, burns, land use

IBM, in collaboration with NASA, has unveiled Prithvi – an open-source AI model designed to facilitate the analysis of satellite imagery. The model, released under the Apache 2 license, was developed using a database of images gathered over a year by the US space agency’s Harmonized Landsat Sentinel-2 (HLS) program. The Prithvi model is compact with 100 million parameters and offers three variations, each fine-tuned for recognizing flooding, wildfire burn scars, and crop and land use.

The purpose of Prithvi and its variants is to automate the interpretation of satellite imagery. For instance, the model can identify various features in the landscape such as water bodies, forests, crop fields, developed land, and wetlands, among other elements. This functionality can be utilized to study land changes over time, such as tracking the impact of natural disasters. The model follows in the footsteps of previous machine learning efforts to analyze satellite imagery.

The open-source nature of Prithvi allows developers worldwide to download and use the models, potentially magnifying their impact. The models can be downloaded from Hugging Face. Moreover, IBM and NASA have made several online demonstrations of Prithvi available, showcasing its utility for different types of environmental analysis. The performance of Prithvi has been claimed by IBM to be superior to previous methods of geospatial imagery analysis, despite using less labeled data.

Looking forward, Prithvi is expected to assist in tracking climate change and land use, as the quantity of satellite data anticipated to be collected by science probes around the Earth is projected to hit 250,000 terabytes by 2024. IBM has revealed that the model was trained using Vela, its AI supercomputer cluster. The ability to fine-tune the model for detecting flooding reportedly took only about an hour using an Nvidia V100 GPU. A commercial version of Prithvi is also expected to be launched later this year.