OlmoEarth: A new state-of-the-art Earth observation foundation model family

OlmoEarth is a family of open foundation models built to make Earth AI practical, scalable, and performant for real-world applications. Pretrained on large volumes of multimodal Earth observation data, OlmoEarth excels at turning raw signals into operational intelligence and insights across space and time. 

OlmoEarth comes in four sizes, all with the same architecture and training approach:

• OlmoEarth-v1-Nano (~1.4M parameters) & OlmoEarth-v1-Tiny (~6.2M)—for fast, cheap inference at scale
• OlmoEarth-v1-Base (~90M)—balanced accuracy and speed for most use cases
• OlmoEarth-v1-Large (~300M)—best performance on challenging tasks

Beyond releasing our research, training, and evaluation stack as open artifacts, along with our technical report, we've embedded OlmoEarth in the OlmoEarth Platform—our solution that enables organizations to leverage our models for Earth observation and accelerate their missions without requiring deep AI or engineering expertise.

State-of-the-art performance

We've extensively evaluated OlmoEarth on dozens of industry-standard research benchmarks and a diverse set of challenging, mission-critical, real-world applications including deforestation-cause alerting in the Amazon, global mangrove loss detection, wildfire risk assessment in North America, smallholder crop-type mapping in Sub-Saharan Africa, maritime detection on the high seas, and fine-grained ecosystem classification. Regardless of evaluation method – whether k-nearest neighbors (kNN), linear probing (LP), or supervised fine-tuning (SFT) – OlmoEarth delivers exceptionally high performance.

OlmoEarth is industry-leading across scene and patch classification, semantic segmentation, object and change detection, and regression in both single-image and time-series domains. OlmoEarth outperforms many recent foundation models from industrial research labs – including Meta's DINOv3, IBM/NASA's Prithvi, and IBM's Terramind – as well as leading academic models like CROMA and Panopticon, and several of our previously-released geospatial models including Satlas and Galileo.

We also compared OlmoEarth to Google DeepMind's AlphaEarth Foundations (AEF). AEF required a different analysis because Google released annualized embeddings, but not the model itself. When we compared the AEF and OlmoEarth embeddings using kNN on three tasks, we found OlmoEarth performed on par or better than AEF. However, once we fine-tuned OlmoEarth, it outperformed AEF substantially. This underscores the importance of a platform that makes fine-tuning and model customization as accessible as possible.

How it works and what's different

OlmoEarth advances on existing methods for training and performing inference with Earth observation data. Our approach combines the scale and utility of self-supervised learning with the value of high-quality labeled data, eliminating the need to choose between approaches or maintain separate training pipelines.

OlmoEarth is a standard vision transformer that processes multimodal time series of satellite images into a unified sequence of tokens. Each token corresponds to a portion of a satellite image, specifying the location and time of that portion along with the sensor type used to capture the image. The model can reason across space, time, and different data modalities simultaneously.

OlmoEarth ingests monthly time series of images from a diverse set of satellite sensors, including optical imagery and radar. It also includes contextual maps like OpenStreetMap, land cover data, and tree canopy height directly into the training process.

We train OlmoEarth by masking parts of a subset of the satellite images, hiding that information from the model. The model learns by using the other visible parts of the satellite images to predict those hidden portions. This self-supervision, combined with weakly supervised learning via map modalities, enables OlmoEarth to build robust general representations from large observational datasets.

Our pretraining dataset consists of millions of samples – approximately 10 terabytes – from around the world, resampled to a uniform resolution of 10 meters per pixel. We use up to 12 monthly timestamps for multi-temporal data; however, many samples contain only a subset of the timesteps and modalities due to the complex and messy real-world data conditions. This flexibility means OlmoEarth can handle incomplete or irregular data with ease.

Deploying OlmoEarth for real-world challenges

OlmoEarth's state-of-the-art performance translates directly to practical benefits. We've already worked with a variety of partners on critical tasks:

• Agriculture and food security: OlmoEarth enables accurate crop-type mapping of smallholder farming in regions such as Sub-Saharan Africa, helping governments and organizations understand what's being grown, where, and when. This supports targeted agricultural interventions, enabling timely, data-driven decisions for improving food security.
• Fire risk assessment: OlmoEarth achieves the lowest error on live fuel moisture content estimates. Operational fire management tools can now approximate vegetation dryness more accurately, improving ignition-risk maps used for prescriptive controlled burns, evacuation planning, and pre-positioning of firefighting resources.
• Object and vessel detection: Better detection in both optical and radar domains reduces missed detections and false alarms, supporting maritime safety, fisheries enforcement, and coastal infrastructure monitoring.
• Land-cover and ecosystem mapping: High performance directly improves habitat protection and land-use planning. Conservation teams can prioritize limited field surveys in the highest-risk areas, while NGOs and governments can allocate restoration funds more confidently using more trustworthy maps.

Because OlmoEarth is competitive with or better than much larger specialist models, organizations gain precision without requiring massive compute. This reduces inference costs, enables continual improvement, and lets teams re-run analyses frequently instead of relying on stale snapshots.

Availability and what's next

OlmoEarth was designed to be easily extensible for bespoke applications across a wide range of problems. It can be leveraged within the OlmoEarth Platform to build customized, highly performant models serving organizations across the entire lifecycle, from raw data acquisition through labeling, fine-tuning, and production deployment.

To demonstrate this in action, we're releasing fine-tuned OlmoEarth models for real-world challenges such as mangrove classification, crop-type and cropland mapping, and forest-fire fuel classification. These were developed with organizations across multiple regions and are ready for adaptation to new areas.

The next family of OlmoEarth foundation models, already in development and expected to be released next year, expands into new sectors like humanitarian response with support for weather data and additional modalities. 

We wish to express deep gratitude to our early collaborators who shared data, expertise, and time to make these models successful in real-world, mission-critical applications: Amazon Conservation, African Wildlife Foundation, CGIAR/International Food Policy Research Institute (IFPRI), Global Mangrove Watch, Global Ecosystem Atlas, ITC University of Twente, NASA Jet Propulsion Laboratory (JPL), and NASA Harvest. We're rapidly onboarding new organizations, focusing on problems that align with Ai2's mission—including food security, wildfire resilience, conservation, sustainability, and ecology and biodiversity.

Sign up to get updates and partner access, and explore our Viewer to see samples of OlmoEarth model outputs.