EarthNetV1

Code for the paper: "Global atmospheric data assimilation with multi-modal masked autoencoders", developed by Zeus AI Inc.

Setup environment with miniconda

conda env create -f environment.yml

Download data and model weights

Test data is included from the subset of available MIRS data. A larger dataset was used to pre-train with GEO, ATMS, and VIIRS alone.

mkdir data checkpoints
aws s3 cp --endpoint https://fly.storage.tigris.dev --no-sign-request s3://zeus-public/earthnet/v1/GEO-ATMS-VIIRS-MIRS.test.tar data/
tar -xvf GEO-ATMS-VIIRS-MIRS.tar
aws s3 cp --endpoint https://fly.storage.tigris.dev --no-sign-request s3://zeus-public/earthnet/v1/earthnet.v1.ckpt checkpoints/

demo.ipynb and inference_earthnet.py show example of how to make predictions with EarthNet.

Cite us

@article{vandal2024global,
  title={Global atmospheric data assimilation with multi-modal masked autoencoders},
  author={Vandal, Thomas J and Duffy, Kate and McDuff, Daniel and Nachmany, Yoni and Hartshorn, Chris},
  journal={arXiv preprint arXiv:2407.11696},
  year={2024}
}