AI-for-Good-Tutorial-2026

Hands-on material for the AI for Good workshop “Embedding Workflows for Earth Observation Tasks”, organized by Embed2Scale (February 2026).

This tutorial introduces embedding workflows for Earth Observation (EO) using TerraTorch and NeuCo-Bench. It is structured into three self-contained parts:

• Parts 1 & 2: lightweight, designed to follow live during the workshop.
• Part 3: full end-to-end workflow with additional data and (ideally) GPU access.

Before starting, complete the setup in setup.md.

---

Overview

Part 1 — TerraTorch: Embeddings + Lightweight Decoders

You will:

• Generate embeddings from EO data with TerraTorch
• Train and evaluate lightweight decoder‑only models
• Inspect embeddings and compare results

Part 2 — NeuCo‑Bench: No‑Code Evaluations

You will:

• Learn the NeuCo‑Bench benchmarking workflow
• Evaluate pre-computed embeddings fully no‑code

Part 3 — End‑to‑End Workflow

You will:

• Download raw NeuCo‑Bench data
• Compute and store embeddings to run your own NeuCo‑Bench evaluations

---

Repository Structure

AI-for-Good-Tutorial-2026/
│
├── setup.md        # Setup instructions (env, dependencies, downloads)
├── data/           # Provided data + location for additional downloads
│
├── notebooks/      # Hands-on notebooks for Parts 1 & 3
│   ├── 01_terratorch_embedding_generation.ipynb
│   ├── 01_terratorch_embedding_downstream.ipynb
│   └── 03_terratorch_neuco_embeddings.ipynb
│
├── scripts/        # Helper scripts (e.g. data download)
├── workflows/      # TerraTorch CLI YAMLs + NeuCo-Bench scripts
└── results/        # Outputs produced during the tutorial

---

Getting Started

Part 1 — Embedding Generation + Downstream Tasks (TerraTorch)

Run:

• notebooks/01_terratorch_embedding_generation.ipynb → writes embeddings to results/01_embeddings/
• notebooks/01_terratorch_embedding_downstream.ipynb → writes decoder logs + weights to results/01_decoders/

---

Part 2 — No‑Code Benchmarking (NeuCo‑Bench)

In terminal:

chmod +x workflows/02_neuco_benchmark.sh
./workflows/02_neuco_benchmark.sh

This produces results under results/neuco_bench/.

---

Part 3 — Full End‑to‑End Workflow (optional)

For this part you need the optional ~50 GB SSL4EO‑S12 L1C download.

1. Complete the optional data download section in setup.md → downloads to data/neuco_ssl4eo_downstream/.
2. Continue with:
   • notebooks/03_terratorch_neuco_embeddings.ipynb
3. To generate full TerraMind tiny max pooling embeddings via CLI:

terratorch predict -c workflows/03_terratorch_neuco_embeddings.yaml

Make sure to change in the NeuCo‑Bench config: update_leaderboard = true.

Then:

chmod +x workflows/03_neuco_benchmark.sh
./workflows/03_neuco_benchmark.sh

Finally inspect the NeuCo‑Bench results folder.

---

Final Links

• TerraTorch: https://github.com/terrastackai/terratorch
• NeuCo-Bench: https://github.com/embed2scale/NeuCo-Bench
• NeuCo-Bench Dataset: https://huggingface.co/datasets/embed2scale/SSL4EO-S12-downstream
• Embed2Scale Project: https://embed2scale.eu

If you have questions, feel free to reach out: isabelle.wittmann1@ibm.com or open issues in TerraTorch, NeuCo‑Bench, or this tutorial repository.