The Allen Institute for Artificial Intelligence (AI2) in the United States has just launched OlmoEarth v1.3, a cutting-edge Earth observation foundation model. It has also rolled out precomputed satellite image feature datasets and browser-based analytical tools for worldwide access. This model amalgamates year-round observational data from diverse satellite types, employing advanced technologies like Vision Transformers. It condenses information from 10-meter resolution surface grids into 128-dimensional feature vectors (with an extra 64-dimensional variant for lower storage requirements), acting as 'digital fingerprints' of the Earth's surface. These vectors facilitate surface feature comparison, identification of similar regions, and monitoring of surface changes.

In a comparative assessment involving 27 intercontinental surface identification tasks, pitting OlmoEarth v1.3 against models like Google's AlphaEarth Foundations and Tessera v2, the 128-dimensional version of OlmoEarth v1.3 emerged victorious in 20 tasks. Meanwhile, the 64-dimensional version topped the charts in 18 tasks. Notably, OlmoEarth v1.3 boasts significantly enhanced training efficiency compared to its forerunners.

The accompanying precomputed feature datasets span global landmasses and adjacent coastal waters. Complete global coverage is achieved for data spanning from 2023 to 2025, with partial regional coverage for data from 2017 to 2022, which will be progressively expanded. Leveraging browser-based tools, users can swiftly conduct preliminary explorations, such as surface classification and change monitoring, without the hassle of running large models independently or registering.

Currently, all pertinent resources, code, and technical reports are freely accessible. The model has already found applications in diverse scenarios, including monitoring of illegal mining activities, crop identification, and disaster damage assessment.