Image: Apollo 17 crew / NASA (AS17-148-22727) ( Public domain )
What it is
Prithvi is a family of open geospatial foundation models from NASA and IBM, first released in 2023, trained on Harmonized Landsat and Sentinel-2 (HLS) time series. Fine-tunes target flood mapping, burn scars, and crop tools. Related “AI for science” stacks include NVIDIA Earth-2 (simulation and emulation): a different product, also not a crystal ball.
Problem it targets
Earth-system work drowns in pixels. A shared pretrained model can cut the cost of a flood map in a country that lacks a large labeled set: if someone still checks the map.
How it works
A transformer or similar backbone learns generic spatial-temporal features from satellite cubes. Downstream heads are small labeled tasks. It does not replace GCMs for century-scale climate; it helps with observation tasks.
Status and players
NASA (HLS, IMPACT); IBM Research; Hugging Face / GitHub weights; users in disaster and agriculture agencies. Earth-2 is NVIDIA’s separate emulation line.
Risks and limits
Bias where the training satellites never looked, military dual use, and vendors wrapping a fine-tune as “the climate model.” Read the model card.
Sources
NASA Earthdata Prithvi release (2023); IBM-NASA Hugging Face model cards. A starting weight is not an oracle.