Prithvi-EO-2.0

Prithvi-EO-2.0 is the IBM and NASA geospatial foundation model, a transformer with 3D spatiotemporal embeddings pretrained on about 4.2 million samples from NASA's Harmonized Landsat and Sentinel-2 archive. The 300M and 600M parameter variants add temporal and location embeddings, so the model reasons season by season and transfers to new geographies with little labeled data.

Weights, fine-tuning workflows, and training data are openly released under Apache 2.0 through Hugging Face and IBM's TerraTorch toolkit. The 600M model averaged 75.6% on GEO-Bench, an 8% improvement over its predecessor, and the family has been demonstrated on flood mapping, wildfire burn scars, crop classification, and biomass estimation. Limitations: the 30m HLS training resolution means fine detail still needs higher-resolution inputs, and it is a vision backbone for EO tasks rather than a general chatbot-style model.