Getting Started with OlmoEarth: From Embeddings to Fine-tuning

Joseph Redmon, Henry Herzog, Favyen Bastani, Yawen Zhang, Gabriel Tseng, Hadrien Sablon, Piper Wolters, Ryan Park, Jacob Morrison, Alexandra Buraczynski, Karen Farley, Josh Hansen, Andrew Howe, Patrick Alan Johnson, Mark Otterlee, Ted Schmitt, Hunter Pitelka, Stephen Daspit, Rachel Ratner, Christopher Wilhelm et al. (7 additional authors not shown)

Abstract: We present a hands-on tutorial for OlmoEarth, a foundation model for Earth observation. The tutorial covers two approaches for land cover classification: (1) extracting embeddings from the frozen encoder for rapid prototyping with simple classifiers (kNN, linear probe), and (2) end-to-end fine-tuning for higher accuracy. Using the African Wildlife Foundation land cover dataset from Kenya, learners work through the full pipeline from data loading through model evaluation. The accompanying iPython notebook runs in approximately 2-3 hours on Colab with default settings and a T4 GPU, or roughly 45 minutes on an Apple M4. All code, data, and pre-trained weights are openly available.

Submission Number: 5