source: Hugging Face Blog: The OlmoEarth Platform: Geospatial inference at planetary scale
level: technical
The Allen Institute for AI (Ai2) introduced the OlmoEarth Platform, a system for running geospatial foundation models at planetary scale. It supports the full lifecycle from fine-tuning and evaluation to large-scale inference. The platform targets governments, NGOs, and mission-driven groups that lack the engineering resources to manage data labeling, model training, and distributed computing on their own. Ai2 built it based on a decade of experience operating tools like Skylight and EarthRanger.
The platform can process continent-sized areas in about a day, handling dozens of terabytes of imagery at fractions of a penny per square kilometer. A recent North America wildfire risk map used 19,600 CPUs and 994 GPUs in parallel, cutting a 4,737-hour serial job to 30.5 hours. It divides work into CPU-based data prep, GPU inference, and CPU postprocessing to keep expensive hardware busy. Failures are handled with automatic retries and idempotent tasks, while a custom metadata index avoids overwhelming public satellite catalogs.
OlmoEarth models are pretrained on roughly 10 terabytes of multimodal satellite data and are already used for deforestation monitoring, food security, and wildfire risk. The platform pulls imagery from multiple providers using cloud-optimized formats like COG and Zarr, reading only needed pixels. Ai2 plans to add automated runs, change detection alerts, agentic interfaces, faster models, weather data, and precomputed embeddings. The system runs on Google Cloud but is designed for multi-cloud and on-premise deployments.
why it matters: It gives resource-constrained environmental organizations practical access to large-scale geospatial AI, reducing the infrastructure barrier for tasks like deforestation monitoring and disaster response.
source: Hugging Face Blog: The OlmoEarth Platform: Geospatial inference at planetary scale