IBM and NASA have developed and released Prithvia open source AI foundational model created to help researchers and scientists analyze satellite images of Earth. This is a relatively small vision transformer model, having only 100 million parameters. Before its launch, it has trained for a year with images collected by the US space science program HLS 2 (Harmonized Landsat Sentinel-2).
In addition to the main model, three variants of Prithvi are available, also licensed under the Apache 2 license. These three variants are tuned to identify floods, signs of burning from forest fires, and crops and other land uses. Access to available information about the model, as well as its variants, is available via Hugging Face.
Its operation, in essence, is quite simple. It is enough to supply one of the commented models with a photograph obtained via satellite from a great height. When it has it, the model will label zones in the part it includes. For example, if the adjusted variant is used to recognize crop fields, it will probably be able to recognize areas with water, forest, wheat fields, cornfields, and other crops. Also if there are swamps or developable plots in the image.
These models can be useful to automate the study of changes in the Earth’s surface over time. Among other things, they can be used to study the effects of erosion caused by floods, or see how drought or forest fires affect a specific region.
Kevin Murphy, NASA’s Head of Scientific Data, recalled that the entity believes “that foundational models have the potential to change the way observation-based data is analyzed, as well as to help improve understanding of the planet.” . And by moving the models to open source and making them available to the world, we hope to multiply their impact.”
At IBM, on the other hand, they claim that Prithvi is up to 15% better than other high-tech techniques used for geospatial image analysis, despite the fact that it depends less than half of these techniques on labeled data. A few months ago, IBM and NASA announced that they were going to develop models for the study of climate change, and Prithvi is likely to be one of them, since, among other things, it is expected to help record climate change and the use of land. land.
IBM has used its Vela AI supercomputing cluster to train the model, and apparently it only took an hour to tune the model for flood detection, with the help of the Nvidia V100 GPU, so it is likely that the creation of a Prithvi variant does not require huge resources. The company plans to launch a commercial version of Prithvi before the end of this year.
