OlmoEarth v1.1 cuts AI costs by up to threefold

OlmoEarth v1.1 cuts the computational cost of its satellite image analysis models by up to threefold, according to Allen Institute for AI. The new family aims to maintain performance similar to OlmoEarth v1 while processing maps of huge areas with less time and fewer resources.

The model is designed for tasks such as tracking mangrove changes, identifying the causes of forest loss, and creating crop maps at national scale. Its partners have already used OlmoEarth in projects covering national, continental, and global areas.

Lower costs for analyzing more territory

Processing satellite images involves more than running a model. The full workflow includes exporting the data, preparing it, making the prediction, and processing the results. Of all these stages, the computation required to run the model is the most expensive part.

Reducing that cost has a practical effect: it allows planetary maps to be updated more frequently and makes the technology more accessible to teams without large-scale infrastructure. It can also speed up the process of adapting the model to specific tasks, such as distinguishing crop types in a region.

OlmoEarth v1.1 offers models in three sizes: Base, Tiny, and Nano. This lets each organization choose between greater capacity and lower consumption, depending on its available computing budget.

The change is in how it reads images

OlmoEarth uses an architecture called a transformer, which is common in many current artificial intelligence systems. To analyze a satellite image, it first divides it into fragments and turns each one into a token, a small unit of information that the model can process.

In Sentinel-2 images, the original system generated separate tokens for each spatial resolution. Because images can include multiple dates, resolutions, and color bands, the number of tokens grows quickly. And in transformers, the cost increases quadratically as that sequence grows.

Version 1.1 combines information from different resolutions into a single token. This reduces the number of tokens by approximately threefold and, as a result, the work required during training and execution.

The change was not automatic. A straightforward combination caused a drop of 10 percentage points on a standard remote-sensing image classification benchmark called m-eurosat kNN. To prevent that, the team modified how it trained the model before releasing it.

What changes for people already using OlmoEarth

Allen Institute for AI says OlmoEarth v1.1 delivers performance similar to the previous version across a combination of research benchmarks and tasks created with partners. Still, it acknowledges that some regressions have been observed, so results may vary depending on the use case.

The recommendation for people using OlmoEarth v1 is to test the new version in their own projects. If it maintains the required performance, the benefit could be considerable: up to one-third of the computing cost and faster execution during fine-tuning and inference.

The importance of this release goes beyond a cheaper version. By training v1.1 on the same dataset as v1, the team can better study how much the architecture and training method influence results. The next step will be to determine which tasks retain their accuracy when the number of tokens is reduced and which still require the original design.