The Allen Institute for AI (AI2) has announced OlmoEarth v1.1, marking a significant step forward in applying AI to satellite imagery. The most prominent highlight of this model line is its ability to reduce compute costs by up to 3x compared to its predecessor, paving the way for planet-scale data analysis at a fraction of the cost.
New Token Structure
In transformer models designed for remote sensing, performance and cost depend heavily on model size and token sequence length. In version 1.1, the research team redesigned the data representation by packing different resolutions (such as Sentinel-2 data) into a single token rather than keeping them separate.
This change reduces the total token count by threefold. While conventional token merging often leads to severe performance drops, AI2 mitigated this by adjusting the pre-training regimen, enabling the model to maintain its accuracy across environmental and agricultural benchmarks.
Why It Matters
Processing satellite data to track climate change, manage forests, or monitor agriculture typically requires massive computing infrastructure. Slashing operating costs by 3x enables conservation organizations and the research community in Vietnam to deploy national- or regional-scale resource monitoring maps more frequently.
OlmoEarth v1.1 was trained on the exact same dataset as v1, allowing researchers to easily isolate and evaluate the impact of methodological changes. The weights for the Base, Tiny, and Nano sizes have been publicly released, offering maximum support to the open-source community.