OlmoEarth Tutorial — ICLR ML4RS Workshop

Getting started with OlmoEarth: from embeddings to fine-tuning for land use / land cover classification, using an African Wildlife Foundation (AWF) dataset in southern Kenya near Amboseli National Park.

Contents

• OlmoEarthTutorial.ipynb — follow-along notebook with code cells to run (recommended).
• OlmoEarthTutorialCompleted.ipynb — fully executed notebook with outputs, for reference.
• OlmoEarthTutorial.pdf — static PDF of the executed notebook.

Running the tutorial

The tutorial is designed for Google Colab with a GPU runtime (T4 is sufficient). It also runs locally on a machine with a CUDA GPU or Apple Silicon (MPS).

Option 1: Google Colab (recommended)

1. Open colab.research.google.com.
2. File → Upload notebook and select OlmoEarthTutorial.ipynb from this repo.
3. Runtime → Change runtime type → GPU.
4. Run the cells top to bottom. The first cell installs dependencies; the dataset (~1.8 GB) is downloaded from HuggingFace inside the notebook.

Option 2: Local (Jupyter)

git clone https://github.com/allenai/olmoearth_ml4rs_tutorial.git
cd olmoearth_ml4rs_tutorial

python -m venv .venv && source .venv/bin/activate
pip install olmoearth_pretrain rslearn scikit-learn matplotlib einops \
            huggingface_hub 'jsonargparse[signatures]>=4.27.7' \
            jupyter rasterio scipy

jupyter notebook OlmoEarthTutorial.ipynb

Time & resources

Approach	Time	GPU memory
Embeddings + kNN / linear probe	minutes	~2–3 GB
Fine-tune (4 epochs)	~15–20 min	~4–6 GB
Fine-tune (30 epochs)	~2–3 hours	~4–6 GB

Default settings complete in roughly 30–45 minutes end-to-end.