{"id":"us-ai2-olmoearth-v13-satellite-embeddings-20261022","canonicalSlug":"us-ai2-olmoearth-v13-satellite-embeddings-20261022","url":"https://wellficent.com/en/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","language":"en","title":"Ai2 releases OlmoEarth v1.3 satellite embeddings, with global coverage for 2023–25","summary":"The 10-metre dataset turns annual Sentinel and Landsat time series into 128-value pixel features. Ai2 reports benchmark gains, but these vectors are not ground-truth labels, mine or well records, or proof of a physical change.","body":["The Allen Institute for AI (Ai2) released OlmoEarth v1.3 on 22 October, alongside precomputed annual satellite embeddings and a browser-based explorer. Each embedding is a 128-dimensional feature vector for a 10-metre pixel, generated from a year of Sentinel-1 radar, Sentinel-2 multispectral and Landsat imagery. The first 64 dimensions form a complete smaller embedding, so users can choose a lower-volume representation. Ai2 says the archive is roughly 150 terabytes per year and includes provenance tables linking each tile to the satellite scenes used.","The archive is not nine years of global coverage. At launch, Ai2 and Source Cooperative said 2023, 2024 and 2025 were global, while 2017 through 2022 covered selected countries and were being extended over the same footprint. The data page describes a coverage mask based on consistent observations across modalities and years; it excludes Antarctica and notes a narrow gap at the antimeridian. “Global” therefore describes the recent-year release within the archive’s stated coverage, not full nine-year world coverage.","An embedding is a model-derived feature representation, not a land-use label or direct observation of a mine, well or other asset. The files provide annual vectors and source-scene provenance; they do not constitute a verified mining map, physical well inventory or geological survey. Users are expected to train a classifier or regressor on their own labeled points, or compare vectors across years to identify places for further review. Ai2’s launch post illustrates this with change analysis around gold mining in Peru’s Madre de Dios region, but that is the publisher’s example, not independent verification of a mine boundary or a substitute for field evidence.","Ai2’s launch post reports that 128-dimensional embeddings ranked highest on 20 of 27 evaluations and the 64-dimensional version on 18 of 27. Its technical report separately describes results on the AEF benchmark datasets: under its k-nearest-neighbour and ridge protocol, the 128-dimensional vectors outperformed AlphaEarth Foundations and TESSERA v2 on seven of eight datasets; the report notes Ethiopia as an exception where TESSERA v2 consistently performed better. These are evaluations reported by the model developer, not independent validation of accuracy in every region or task. The reported counts use different evaluation summaries and should not be merged into a single universal score.","The weights, dataset and code are distributed under Ai2’s OlmoEarth Artifact License, a custom agreement with explicit restrictions. It permits use, modification and distribution subject to terms, but prohibits use of the artifacts or derivatives for military and defense applications, human surveillance or policing, and extractive activities. The latter restriction expressly includes using them to plan or facilitate oil, natural-gas or mineral extraction through drilling, mining or deforestation. Distributors must attribute Ai2 and pass the restrictions to downstream recipients; the agreement also says it does not grant rights in third-party material contained in datasets. This matters for energy and mining users: a feature vector cannot be treated as permission or evidence for exploration or extraction decisions.","The release’s practical contribution is a reusable feature layer that can reduce the compute and imagery handling needed to test mapping ideas. It does not remove the need for task-specific labels, geographic validation, imagery review or licensing checks. Ai2 says the explorer can compare years, find similar locations, train a lightweight classifier and export a GeoTIFF; those outputs remain analyses built on the embeddings, not observations supplied by the archive. Coverage varies by year, and the model’s own license limits particular uses. For physical well or mine mapping, a separate permitted dataset and verification process would still be required."],"keyPoints":[],"sections":[],"text":"Ai2 releases OlmoEarth v1.3 satellite embeddings, with global coverage for 2023–25\n\nThe 10-metre dataset turns annual Sentinel and Landsat time series into 128-value pixel features. Ai2 reports benchmark gains, but these vectors are not ground-truth labels, mine or well records, or proof of a physical change.\n\nThe Allen Institute for AI (Ai2) released OlmoEarth v1.3 on 22 October, alongside precomputed annual satellite embeddings and a browser-based explorer. Each embedding is a 128-dimensional feature vector for a 10-metre pixel, generated from a year of Sentinel-1 radar, Sentinel-2 multispectral and Landsat imagery. The first 64 dimensions form a complete smaller embedding, so users can choose a lower-volume representation. Ai2 says the archive is roughly 150 terabytes per year and includes provenance tables linking each tile to the satellite scenes used.\n\nThe archive is not nine years of global coverage. At launch, Ai2 and Source Cooperative said 2023, 2024 and 2025 were global, while 2017 through 2022 covered selected countries and were being extended over the same footprint. The data page describes a coverage mask based on consistent observations across modalities and years; it excludes Antarctica and notes a narrow gap at the antimeridian. “Global” therefore describes the recent-year release within the archive’s stated coverage, not full nine-year world coverage.\n\nAn embedding is a model-derived feature representation, not a land-use label or direct observation of a mine, well or other asset. The files provide annual vectors and source-scene provenance; they do not constitute a verified mining map, physical well inventory or geological survey. Users are expected to train a classifier or regressor on their own labeled points, or compare vectors across years to identify places for further review. Ai2’s launch post illustrates this with change analysis around gold mining in Peru’s Madre de Dios region, but that is the publisher’s example, not independent verification of a mine boundary or a substitute for field evidence.\n\nAi2’s launch post reports that 128-dimensional embeddings ranked highest on 20 of 27 evaluations and the 64-dimensional version on 18 of 27. Its technical report separately describes results on the AEF benchmark datasets: under its k-nearest-neighbour and ridge protocol, the 128-dimensional vectors outperformed AlphaEarth Foundations and TESSERA v2 on seven of eight datasets; the report notes Ethiopia as an exception where TESSERA v2 consistently performed better. These are evaluations reported by the model developer, not independent validation of accuracy in every region or task. The reported counts use different evaluation summaries and should not be merged into a single universal score.\n\nThe weights, dataset and code are distributed under Ai2’s OlmoEarth Artifact License, a custom agreement with explicit restrictions. It permits use, modification and distribution subject to terms, but prohibits use of the artifacts or derivatives for military and defense applications, human surveillance or policing, and extractive activities. The latter restriction expressly includes using them to plan or facilitate oil, natural-gas or mineral extraction through drilling, mining or deforestation. Distributors must attribute Ai2 and pass the restrictions to downstream recipients; the agreement also says it does not grant rights in third-party material contained in datasets. This matters for energy and mining users: a feature vector cannot be treated as permission or evidence for exploration or extraction decisions.\n\nThe release’s practical contribution is a reusable feature layer that can reduce the compute and imagery handling needed to test mapping ideas. It does not remove the need for task-specific labels, geographic validation, imagery review or licensing checks. Ai2 says the explorer can compare years, find similar locations, train a lightweight classifier and export a GeoTIFF; those outputs remain analyses built on the embeddings, not observations supplied by the archive. Coverage varies by year, and the model’s own license limits particular uses. For physical well or mine mapping, a separate permitted dataset and verification process would still be required.","category":"economy","region":"world","topics":["ai","mining"],"eventDate":"2026-10-22","eventDateBasis":"Ai2 published the OlmoEarth v1.3 release on 22 October 2026.","publishedAt":"2026-10-25T21:01:30Z","modifiedAt":"2026-10-25T21:01:30Z","publicationBasis":"first-publication","preparedAt":"2026-10-25T20:53:36Z","sourcePublishedAt":"2026-10-22","translatedAt":"2026-10-25T20:53:12Z","sources":[{"url":"https://allenai.org/blog/olmoearth-v1-3","name":"Allen Institute for AI (Ai2)","licence":"No open reuse licence stated for article text; factual reporting only.","checkedAt":"2026-10-25","licenseUrl":"https://allenai.org/","attribution":"Allen Institute for AI (Ai2); original factual summary, no source prose or image reproduced.","publishedAt":"2026-10-22"},{"url":"https://allenai.org/papers/olmoearth-v1-3","name":"OlmoEarth v1.3 technical report","licence":"OlmoEarth Artifact License applies to model artifacts; article factual summary only.","checkedAt":"2026-10-25","licenseUrl":"https://data.source.coop/olmoearth/olmoearth-v1-3-embeddings/LICENSE.txt","attribution":"Ai2 technical report; publisher-reported architecture and evaluation details."},{"url":"https://source.coop/olmoearth/olmoearth-v1-3-embeddings","name":"OlmoEarth v1.3 Embeddings dataset documentation","licence":"OlmoEarth Artifact License; license terms summarized in article.","checkedAt":"2026-10-25","licenseUrl":"https://data.source.coop/olmoearth/olmoearth-v1-3-embeddings/LICENSE.txt","attribution":"Source Cooperative dataset documentation and provenance notes.","publishedAt":"2026-10-22"},{"url":"https://data.source.coop/olmoearth/olmoearth-v1-3-embeddings/LICENSE.txt","name":"OlmoEarth Artifact License","licence":"OlmoEarth Artifact License.","checkedAt":"2026-10-25","licenseUrl":"https://data.source.coop/olmoearth/olmoearth-v1-3-embeddings/LICENSE.txt","attribution":"Actual artifact license terms reviewed and summarized."},{"url":"https://huggingface.co/allenai/OlmoEarth-v1_3-Base","name":"OlmoEarth v1.3 Base model card and license","licence":"OlmoEarth Artifact License.","checkedAt":"2026-10-25","licenseUrl":"https://huggingface.co/allenai/OlmoEarth-v1_3-Base/blob/main/LICENSE.txt","attribution":"Ai2 model card confirms the custom artifact license."}],"translations":{"en":"https://wellficent.com/en/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","tr":"https://wellficent.com/tr/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","ar":"https://wellficent.com/ar/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","fr":"https://wellficent.com/fr/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","es":"https://wellficent.com/es/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","ru":"https://wellficent.com/ru/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022","pt":"https://wellficent.com/pt/news/us-ai2-olmoearth-v13-satellite-embeddings-20261022"},"corrections":[],"locations":[],"type":"news","classification":{"version":"1","dataKinds":["technology-release"],"geographies":[{"role":"issuer-location","countryCode":"US","subdivisionCode":null}],"evidenceBasis":["reported-observation"],"measurementPeriod":null}}