The OlmoEarth Platform: Geospatial Inference at Planetary Scale

The OlmoEarth Platform: Geospatial Inference at Planetary Scale

The OlmoEarth Platform is an infrastructure designed to transition geospatial foundation models from fine-tuning and evaluation to large-scale inference. It enables organizations to process dozens of terabytes of imagery across continent-scale areas in approximately one day, achieving costs as low as fractions of a penny per square kilometer.

Overcoming Challenges in Satellite Inference

Geospatial inference differs from standard ML tasks due to the massive scale of data and the complexity of inputs. A single fine-tuning job can involve terabytes of data and run for hours, utilizing multiple spectral bands, sensor types, and time steps.

Key technical hurdles include:

  • Data Heterogeneity: Imagery comes from various providers with differing projections, resolutions, and levels of cloud obscuration.
  • Alignment Requirements: Because the output is a map, every prediction must maintain precise alignment with the coordinate grid of surrounding areas.
  • I/O Bottlenecks: Prediction jobs often spend more time downloading and preparing imagery than executing the model, necessitating highly efficient data pipelines.

Optimized Hardware Allocation

To prevent expensive GPUs from being wasted on data preparation, the OlmoEarth Platform divides each inference job into three distinct stages, each mapped to the most cost-effective hardware:

  1. Data Acquisition and Preprocessing (CPU, High I/O): This stage fetches, reprojects, aligns, and normalizes imagery, saving it in a format optimized for fast loading.
  2. Inference (GPU): The platform runs the model's forward pass and writes minimally processed outputs directly to storage.
  3. Postprocessing (CPU): Per-window outputs are stitched together, masks or rescaling are applied, and results are exported in formats such as GeoJSON, GeoTIFF, or Zarr.

Scalable Execution via OlmoEarth Run

OlmoEarth Run, the platform's execution layer, manages large-scale jobs by dividing geographic regions into partitions sized for individual compute instances (workers). These partitions are further subdivided into smaller windows for model processing. Because each window is independent, the platform can execute thousands of forward passes in parallel.

In a recent North America wildfire-risk map generation, the platform demonstrated the following scale:

  • Parallelism: Used approximately 19,600 CPUs and 994 GPUs at peak.
  • Throughput: Network throughput exceeded 168 GB/s.
  • Efficiency: Reduced an estimated 4,737 hours of serial compute to 30.5 hours of wall-clock time, representing a 155× speedup.

Intelligent Data Indexing and Retrieval

To avoid overwhelming external STAC APIs (such as those from ESA or Microsoft Planetary Computer) with thousands of concurrent queries, the OlmoEarth Platform maintains its own internal metadata index.

Metadata Management

  • Update Mechanism: The index is updated via SNS notifications for AWS Open Data or by polling upstream indices every few minutes for other providers.
  • Efficient Retrieval: The index stores scene metadata and pointers to pixels. At runtime, the platform performs windowed reads against cloud-optimized formats (COG or Zarr), retrieving only the necessary bytes for a partition rather than downloading entire scenes.

Best Practices for Data Providers

Based on the platform's development, the team recommends three best practices for publishing Earth observation data: queue-based notifications for new imagery, storage on major cloud platforms without bespoke rate limits, and the use of cloud-optimized formats that support ranged reads.

Fault Tolerance and Recovery

The platform is designed for automatic recovery from the routine failures of distributed computing. Each task is executed within a runner Docker container that is provisioned dynamically.

Because every task is reentrant and idempotent, the platform handles intermittent failures—such as provider unavailability, missing imagery bands, or container crashes—through task tracking, automatic retries, and fallback to alternate providers. A separate monitoring process identifies stalled runners and restarts their tasks.

Future Roadmap

Ai2 is expanding the OlmoEarth Platform to include several new capabilities:

  • Automated Workflows: Scheduling inference jobs or triggering them based on new imagery registration.
  • Alerting Systems: Implementing change detection to notify users of events like flooding or deforestation.
  • Agentic Interfaces: Developing tools to lower the technical barrier for data curation and feature engineering.
  • Efficiency Gains: Researching faster model architectures and developing a dedicated embedding model to replace full forward passes for many tasks.
  • Expanded Modalities: Incorporating weather data (ERA-5) and additional satellite sensors.
  • Cloud Agnostic Deployment: While currently on Google Cloud, the architecture is designed to support multiple clouds and deployments within a partner's own compute environment.

Sources