allenai/olmoearth_pretrain

Earth system foundation model data, training, and eval

What it solves

OlmoEarth provides a family of multi-modal, spatio-temporal foundation models designed for Earth Observation. It addresses the need for scalable planetary intelligence by providing a flexible framework to process raw satellite data and move from research and development to fine-tuning and production deployment.

How it works

The models are trained on a diverse set of satellite modalities, including Sentinel 2, Sentinel 1, and Landsat, as well as six derived maps (such as OpenStreetMap and WorldCover). The architecture consists of an encoder-decoder structure, with various model sizes available ranging from Nano to Base/Large to suit different computational needs.

Who it’s for

This project is for researchers and developers working with satellite imagery and geospatial data who need foundation models that can be fine-tuned for downstream tasks like segmentation or embedding computation.

Highlights

  • Multi-modal support for multiple satellite sources (Sentinel 1, 2, and Landsat).
  • Spatio-temporal foundation models available in multiple sizes (Nano, Tiny, Small, Base, Large).
  • End-to-end platform for planetary intelligence, from pretraining to production.
  • Comprehensive dataset of over 285,000 global samples of 2.56km x 2.56km regions.

Related

  • Project
    PozzettiAndrea/ComfyUI-MotionCapture

    A ComfyUI custom node package that uses GVHMR to extract 3D human motion and SMPL skeletal parameters from video.

  • Project
    MrGiovanni/AbdomenAtlas

    A large-scale multi-organ CT dataset and toolset for efficient abdominal segmentation, reducing annotation time from decades to weeks using human-in-the-loop AI.

  • Project
    zae-bayern/elpv-dataset

    A benchmark dataset of 2,624 annotated electroluminescence images of solar cells used for the visual identification of defects that reduce power efficiency.

  • Project
    Luo-Yihao/FaithC

    A near-lossless 3D voxel representation system that replaces SDF and Marching Cubes with Faithful Contour Tokens (FCTs) to preserve sharp edges and internal structures.