OlmoEarth Overview
Introduction
OlmoEarth is a highly specialized project and platform developed by the Allen Institute for Artificial Intelligence (Ai2), a non-profit artificial intelligence research institute based in Seattle, Washington. Operating as a critical component of Ai2's open science mandate, OlmoEarth introduces a state-of-the-art, multimodal, spatio-temporal foundation model engineered specifically for earth observation, remote sensing, and planetary intelligence.
Initially launched as OlmoEarth v1 in November 2025, the technology was expanded with the release of the optimized OlmoEarth v1.1 family. The overarching mission of OlmoEarth is to bridge the "last-mile" gap between frontier machine learning research and the real-world environmental organizations, researchers, and communities that require actionable geospatial insights to combat climate change, track ecosystem decline, and manage global resources.
Continue…
Core Capabilities and Architecture
Unlike traditional Large Language Models (LLMs) that are optimized purely for generative text workflows, OlmoEarth is structured as a spatio-temporal transformer-based encoder-decoder model. It is pretrained on a massive, globally distributed dataset encompassing hundreds of thousands of multi-temporal samples covering vast geographic regions over multi-year spans.
The underlying technical mechanics of OlmoEarth focus on:
- Multimodal Earth Data Fusion: The model tokenizes and natively ingests data across diverse satellite modalities (including Sentinel-2, Sentinel-1, and Landsat) alongside derived geospatial layers such as OpenStreetMap features, WorldCover data, the USDA Cropland Data Layer, SRTM Digital Elevation Models, and canopy height indexes.
- Spatio-Temporal Modeling: By dividing continuous remote sensing streams into explicit patch tokens layered with 2D sin-cos spatial, sinusoidal temporal, and learnable modality embeddings, the model tracks how landmasses, forests, and waterways evolve over months or years.
- Highly Efficient Latent Image Modeling: OlmoEarth utilizes advanced token-merging techniques and custom pre-training regimens. The refined v1.1 family compresses token sequences, reducing multiply-accumulate operations (MACs) and cutting compute cost-to-run considerations by up to 3x without sacrificing target performance metrics.
- Open Science Integrity: True to Ai2's institutional commitments, the entire training ecosystem?including model weights, pre-training data, training frameworks, and comprehensive evaluation scripts?is fully open-sourced under permissive licensing models.
Products and Platform Solutions
1. OlmoEarth Studio
OlmoEarth Studio acts as an end-to-end, no-code web platform tailored for organizations that require geospatial intelligence but lack internal data engineering or deep machine learning expertise.
- Guided Model Fine-Tuning: Users can upload localized geographic interest vectors, configure training hyper-parameters through an intuitive user interface, and fine-tune foundation models for specific classification or segmentation maps.
- Integrated Annotation Tools: Provides visual annotation mechanisms directly in-platform to label imagery, guiding target model performance over specific territories or landscape anomalies.
- Data Preparation Automation: Programmatically manages the heavy data-engineering lifting, including discovering, exporting, cloud-clearing, and preprocessing raw Earth observation data feeds.
2. OlmoEarth API
For developers and enterprise data engineering teams, the OlmoEarth API serves as a programmatic gateway to scale planet-level intelligence.
- Automated Predication Streams: Enables developers to build datasets, deploy custom fine-tuned variations of OlmoEarth, and fetch high-volume inference predictions directly into existing application backends.
- Embedding Generation: Provides tools to calculate dense spatial and temporal embeddings over monthly or yearly boundaries, allowing third-party machine learning applications to ingest refined earth-feature vectors directly.
3. Open-Source Codebases and Model Families
For researchers and ML engineers, Ai2 distributes the complete architectural blueprint through dedicated software packages:
olmoearth_pretrain: The foundational GitHub repository housing the core data extraction workflows, masking logic, self-supervised pre-training scripts, and historical benchmark evaluations.olmoearth_projects & rslearn: High-level python toolkits that automate complex patch windowing, inference post-processing, and integration loops with rslearn (Ai2's geospatial ingestion engine).
Practical Deployment Configurations
Through collaborative fine-tuning arrangements, the OlmoEarth framework has been actively deployed across several specialized environmental classifications:
- Mangrove Mapping: Tracking sensitive coastline ecosystems and measuring real-time blue carbon habitat changes.
- Forest Loss Classification: Grouping and cataloging active drivers of deforestation across continental forest zones.
- Country-Scale Crop-Type Mapping: Programmatically identifying agricultural footprints, yield dynamics, and crop types within a matter of days over massive nationwide grids.
- Live Fuel Moisture Content Mapping: Assessing wildfire risks by observing systemic water stress signatures across regional vegetation profiles.
