IBM and NASA open-source massive geospatial AI model
IBM and Hugging Face just open-sourced a geospatial AI foundation model built in partnership with NASA. Trained on a year of satellite data across the continental U.S., the model beats state-of-the-art benchmarks by 15% using half the labeled data. Why it matters: It turns massive satellite archives into ready-to-use tools for tracking floods, fires, and deforestation.
IBM and NASA just dropped their joint geospatial AI foundation model on Hugging Face, making it freely available for anyone to download and build on. It turns raw, complex satellite archives into an off-the-shelf engine for mapping changes across planet Earth.
What it can do
In plain English, the model is trained to look at satellite imagery and instantly recognize what is happening on the ground.
On tasks like mapping floods and delineating wildfire burn scars, it beats previous state-of-the-art methods by 15% while requiring only half as much labeled training data.
That drastically cuts down the manual work needed to build specialized applications. With a bit of fine-tuning, developers and researchers can point it at tracking deforestation, forecasting agricultural crop yields, or monitoring greenhouse gas hotspots.
IBM and NASA are also collaborating with Clark University to expand its reach into time-series segmentation and land cover similarity tracking.
How it was built
The fundamental problem in Earth science isn't a lack of data, it's the sheer impossibility of labeling it all. NASA projects that upcoming missions will generate 250,000 terabytes of data by 2024.
To bypass that bottleneck, the team trained a self-supervised transformer across a full year of Harmonized Landsat Sentinel-2 (HLS) satellite observations covering the continental United States.
Because the foundation model already understands the physics, colors, and patterns of Earth's surface, downstream users only need small labeled datasets to achieve high-accuracy results.
The open-source weights are hosted directly on Hugging Face for research and development. If you need enterprise-grade tooling, IBM plans to package a commercial version into its Environmental Intelligence Suite later this year.
Why it matters
If you work in GIS, remote sensing, or climate tech, this is worth downloading immediately. Cutting the need for labeled data by 50% removes the most expensive roadblock in training environmental computer vision models.
We also see this as a healthy forcing function for open science. Geospatial modeling has historically stayed locked behind government supercomputers and proprietary software suites. By publishing the weights openly on Hugging Face, IBM and NASA set a new benchmark for how public satellite data should be put to work.
Sources (1)
- newsroom.ibm.com(opens in a new tab)Originalnewsroom.ibm.com