Introducing OlmoEarth v1.1: A More Efficient Family of Models for Remote Sensing
Allen Institute for AI releases OlmoEarth v1.1, a family of models that cuts compute costs by up to 3x while maintaining performance on remote sensing tasks.

["The Allen Institute for AI has released OlmoEarth v1.1, an updated family of models that brings significant efficiency gains to remote sensing tasks. Since the release of OlmoEarth v1 in November 2023, partners have applied the models to a wide range of tasks, from tracking mangrove change to classifying drivers of forest loss. The new release moves closer to the institute's mission of bringing state-of-the-art AI to organizations working to protect people and the planet.", 'Efficiency is crucial when processing satellite imagery to make predictions across large areas.
OlmoEarth v1.1 achieves this by reducing compute costs by up to 3x while maintaining the performance of OlmoEarth v1 on a mix of research benchmarks and tasks constructed with partners. The models are transformer-based, a dominant architecture in machine learning, and process remote sensing data by converting it into a sequence of tokens.', 'Two key factors control efficiency in transformer-based models: model size and token sequence length. Compute costs scale quadratically with token sequence length, making small reductions in sequence length critical to cutting costs.
The OlmoEarth models are released in a family of sizes, allowing users to choose the size that fits their compute budget. The new models have lower MACs (multiply-accumulate operations), which estimate the computation needed for one model forward pass.', "One of the key challenges in developing OlmoEarth v1.1 was determining what a token should represent in remote sensing data. The institute's approach involves splitting Sentinel-2 imagery into resolution-based patches, creating a token per timestep per resolution.
While using a unique token per resolution is common, it can lead to high token counts and increased compute costs. The institute's researchers found that merging tokens without impacting performance required modifications to the pre-training regimen, which are detailed in their paper.", 'The result is a model family that achieves similar performance to OlmoEarth v1 while requiring significantly less compute. OlmoEarth v1.1 runs up to three times cheaper than OlmoEarth v1, making frequent, planet-scale map refreshes more affordable for teams running OlmoEarth.
The institute has released the weights and training code for OlmoEarth v1.1, including the Base, Tiny, and Nano models, which can be accessed on Hugging Face and GitHub.', 'By isolating the effect of methodological changes, the institute hopes that OlmoEarth v1.1 will advance understanding of scientific principles when pretraining models for remote sensing. The release of OlmoEarth v1.1 marks an important step towards making state-of-the-art AI more accessible to organizations working to protect the planet.']
Source: Hugging Face