Introduction to AI for EO data

Hosted by NEODAAS (Workshop 1)

Learn how machine learning can unlock insights from Earth Observation data. In this introductory workshop, we’ll break down the key steps in preparing EO data for ML models and explore how different problem types—classification, regression, and segmentation—apply to real environmental challenges. Using Marine Primary Production as our case study, we’ll walk through how to design meaningful ML workflows and demonstrate a practical example using a simple Random Forests model.

Learning Outcomes

Have an awareness of best practices around pre-processing data for machine learning, including sampling, normalisation, and splitting.
Understand the difference between classification, regression, and segmentation problems.

Prerequisites

None

Introduction to EO Foundation Models: Inference and Fine‑Tuning

Hosted by IBM (Workshop 1)

This workshop introduces participants to Earth Observation foundation models and provides hands‑on experience running inference and performing a lightweight fine‑tuning workflow. Attendees will work with Prithvi and TerraMind to explore their capabilities, compare outputs, and adapt a model to a simple downstream task. The session builds practical intuition for how EO foundation models generalise and how fine‑tuning improves task‑specific performance.

Learning Outcomes

Understand EO foundation model concepts
Run inference with Prithvi and TerraMind
Perform a simple fine‑tuning workflow

Prerequisites

Google account for Colab, or a laptop with local dependencies installed
Basic Python (variables, functions, simple scripts)
Some familiarity with Jupyter/Colab notebooks
Exposure to remote sensing concepts (bands, resolution) and geospatial machine learning (data formats, error metrics) is helpful but not required

Where your models can be trusted: Evaluating spatial machine learning reliably

Evaluating Hosted by Jakub Nowosad

Understanding how well spatial predictions perform and where they can be trusted requires proper validation and awareness of the model’s spatial limits.
This workshop introduces methods for robust evaluation, including kNN distance matching cross-validation and the Area of Applicability (AoA), using examples in R with the CAST package.
Participants will learn to interpret spatial model performance realistically and apply these concepts broadly, beyond R.
The session combines demonstrations, optional hands-on exercises, and discussion, giving participants both practical skills and conceptual understanding.

Learning Outcomes

Evaluate spatial predictions using kNN distance matching CV and AoA
Understand and interpret where predictions are reliable
Apply these evaluation concepts to their own workflows

Prerequisites

Basic familiarity with machine learning and spatial data concepts
Some experience with R is helpful but not required
Participants who want to follow the examples should have R and an R IDE (e.g., RStudio, Positron, etc.) installed, along with the CAST package

Introduction to the TESSERA Geospatial Foundation Mode: Hands-on Earth Intelligence with Embedding-as-Data

Hosted by Zhengpeng (Frank) Feng

TESSERA is a pixel-wise foundation model for Earth Observation that learns robust representations from multi-modal Sentinel-1/2 time series. Unlike traditional foundation models that require heavy fine-tuning and GPU infrastructure, TESSERA pioneered the “Embeddings-as-Data” paradigm: we release precomputed, global, annual, 10m resolution embeddings that can be downloaded with a single pip install geotessera command and used immediately for downstream tasks with only lightweight models (e.g. a small MLP or scikit-learn classifier). This means you don’t need to deploy or fine-tune any foundation model — just treat the embeddings as analysis-ready geospatial data.
In this hands-on workshop, participants will experience the full TESSERA workflow, from downloading embeddings for any region of interest to building a working land cover classifier. We will start with a brief overview of the TESSERA model and the motivation behind temporal sampling invariance for handling cloud-corrupted satellite imagery. Then, participants will use the GeoTessera Python library to retrieve embeddings for a region of their choice, explore them interactively using the Tessera Embeddings Explorer, a browser-based tool for visual exploration, similarity search, and point-and-click classification, and train a simple classifier on the embeddings with minimal labelled data.

Learning Outcomes

Understand the Embedding-as-Data paradigm and how it differs from traditional foundation model fine-tuning workflows
Download and work with TESSERA embeddings for any region on Earth using the GeoTessera Python library and CLI
Perform interactive visual exploration and point-and-click classification using the Tessera Embeddings Explorer
Build a lightweight downstream model (e.g. land cover classification, solar panel detection) on top of precomputed embeddings with minimal labelled data and no GPU required

Prerequisites

A laptop with a web browser (for the TEE online demo at https://tze.geotessera.org/)
Basic familiarity with Python (installing packages, running scripts) is helpful but not essential
No GPU, no cloud account, no deep learning framework needed — everything runs on a standard laptop
Bring your curiosity and, if you like, coordinates of a region you care about — we’ll map it together!

Introduction to MLOps for Earth Observation

Hosted by Juan Ginzo

What happens when a GeoAI workflow needs to run reliably on new imagery every week, across different seasons and geographies, without someone manually checking every output? This workshop introduces MLOps thinking to the EO community: the engineering practices that take a working pipeline and make it robust, reproducible, and production-aware. Through a hands-on session, participants will explore how to validate inputs, track experiments, test outputs, and detect when something silently breaks. No cloud accounts or infrastructure setup required.

Learning Outcomes

Understand the key principles of MLOps and how they apply to EO workflows
Be able to identify common failure modes when moving from experimentation to production
Apply basic input validation, experiment tracking, and output testing to their own pipelines

Prerequisites

A laptop with a Google (Gmail) account for Colab access, OR the ability to run python notebooks locally in your laptop.
Familiarity with EO data concepts (bands, resolution, indices) is helpful but not required
Some experience with Python

PlotToSat Workshop

Hosted by Milto Miltiadou

This is a practical workshop on how to use PlotToSat for efficiently generating monthly aggregated Sentinel-1 and Sentinel-2 time-series across distributed field-plot regions and polygons defined in a CSV file or Shapefile, respectively.
Associated paper
Github: https://github.com/Art-n-MathS/PlotToSat
BlueSky: @plottosat.bsky.social

Learning Outcomes

Understand the capabilities and limitations of PloToSat
Learn how to efficiently generate aggregated EO time-series
Recognise and resolve typical errors

Prerequisites

PlotToSat uses the cloud platform Google Earth Engine and having an account (https://earthengine.google.com/signup) and VS code installed (https://code.visualstudio.com/) are the prerequisites.

Advanced Geospatial AI: Multimodality, TerraTorch Iterate, and TerraKit

Hosted by IBM (Workshop 2)

This advanced session expands into multimodal geospatial AI and expert workflows with TerraMind. Participants will experiment with combining inputs (e.g. optical + DEM), explore how modality choices affect performance, and learn how to use TerraTorch Iterate for hyper-parameter optimisation. The workshop also introduces TerraKit for dataset preparation, enabling participants to design and execute more sophisticated geospatial AI pipelines.

Learning outcomes

Understand and run advanced TerraMind multimodal fine-tuning and inference
Use TerraTorch Iterate for performance optimisation
Prepare fine-tuning data with Terrakit

Prerequisites

Google account for Colab, or a laptop with local dependencies installed
Completion of Workshop 1 or equivalent experience
Comfortable with Python and notebook‑based workflows
Basic understanding of EO data types (optical vs SAR, DEMs, metadata)