Reproduce “OlmoEarth v1.3: Embeddings” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gabriel Tseng, Favyen Bastani, Patrick Alan Johnson, Yawen Zhang, Piper Wolters, Ando Shah, Hadrien Sablon, Joseph Redmon and 8 more
OlmoEarth v1.3 produces reusable annual satellite embeddings by aggregating observations across dates and sensors into compact vectors for 10-meter locations, while a learned pooling step separates output resolution from the transformer’s coarser processing. Its 128-dimensional embeddings outperform both evaluated baselines on 10 of 11 settings; TESSERA v2 wins in Ethiopia. Generating a global year still costs an estimated 14,000 H100 GPU-hours.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.