Satellite data has a new representation

For fifty years we have analysed satellite imagery as reflectance: how bright each pixel is in each band. In 2025 and 2026, Google, ESA, the University of Cambridge and others released something different: embeddings. Every 10 m pixel on Earth, described by 64 or 128 numbers that a foundation model learned from years of Sentinel-1, Sentinel-2 and Landsat data. Nearby numbers mean similar places.

With embeddings, a land-cover map needs tens of labels instead of thousands. Finding every place that looks like this one becomes a dot product. Change between years becomes a distance. And the data is free.

This course teaches you to understand, access and use these products. You work with real data over a real study area (East Attica, Greece, including the 2018 Mati wildfire) in seven Python notebooks that run on an ordinary laptop. No GPU and no paid account required.

24 video lessons · 2 h 10 min · 7 hands-on notebooks · 6 quizzes · capstone

What you will learn

By the end of the course you can:

  • Explain what an embedding is, and why models trained without labels produce useful ones

  • Compare AlphaEarth (Google DeepMind), TESSERA (Cambridge), Major TOM (ESA Φ-lab / CloudFerro) and Clay by unit, time, inputs, dimension, format and licence

  • Download embeddings for any area from open cloud buckets and Hugging Face, with no account

  • Visualise, search, cluster and classify embeddings with scikit-learn and FAISS

  • Map land cover with 5 to 500 labels per class and validate it honestly

  • Detect and date change across nine years of annual embeddings

  • Search Sentinel-2 imagery with a sentence, using vision-language embeddings

  • Run an open foundation model (Clay v1.5) on your own imagery, on a CPU

  • Know the limits: temporal granularity, domain shift, interpretability, licences, versions

  • Design a workflow that runs every year at regional or national scale

Who is this course for?

  • Remote sensing and GIS professionals who want to add foundation-model embeddings to their toolbox
  • Researchers and students in geography, environmental science, agriculture or forestry who work with Sentinel or Landsat data
  • Analysts in public agencies and companies evaluating AlphaEarth, TESSERA or Major TOM for operational use
  • Anyone who knows what a pixel, a band and a projection are, can run Python in a notebook, and has no machine learning background. The course covers the little you need.

Hands-on details. All notebooks run locally in Python 3.11 on macOS, Linux or Windows (WSL). Setup takes about ten minutes. Downloads are 50–200 MB per notebook from public sources. Notebook 7 downloads the 5 GB Clay checkpoint once and needs about 8 GB of free RAM. Every notebook comes in a learner version with exercises and a solution version with outputs.

Course curriculum

  1. 01

    • 0.1 Welcome and how to use this course

    • Course materials: notebooks, data and setup guide
  2. 02

    • 1.1 Why Earth observation needs a new representation

    • 1.2 Machine learning essentials for EO people

    • 1.3 What is an embedding?

    • 1.4 EO concepts the models rely on

    • Quiz · Foundations: from pixels to vectors
  3. 03

    • 2.1 Self-supervised learning: learning without labels

    • 2.2 Anatomy of an EO foundation model

    • 2.3 The product landscape

    • 2.4 Precomputed embeddings vs running the model yourself

    • Quiz · How embeddings are made
  4. 04

    • 3.1 The access landscape

    • 3.2 AlphaEarth: Earth Engine and the open GCS bucket

    • 3.3 TESSERA, Major TOM and Earth Index in Python

    • 3.4 The European side: CDSE, CREODIAS and CloudFerro GeoAI

    • Hands-on 1 · Setup and your first embeddings

    • Hands-on 2 · Same area, second product: TESSERA

    • Quiz · Where and how to access embeddings
  5. 05

    • 4.1 Seeing embedding space

    • 4.2 Similarity search: find more like this

    • 4.3 Clustering: unsupervised land cover

    • 4.4 Few-shot classification: labels are cheap now

    • 4.5 Change detection across years

    • Hands-on 3 · Similarity search

    • Hands-on 4 · Few-shot land cover

    • Hands-on 5 · Change detection

    • Quiz · Working with embeddings: the core toolkit
  6. 06

    • 5.1 Patch-level embeddings and text search with Major TOM

    • 5.2 Running Clay v1.5 on your laptop

    • Hands-on 6 · Text search over Major TOM

    • Hands-on 7 · Make your own embeddings with Clay

    • Quiz · Patch embeddings and open models
  7. 07

    • 6.1 Applications in practice

    • 6.2 Limitations and pitfalls

    • 6.3 Building an embedding-powered workflow

    • 6.4 The road ahead and wrap-up

    • Quiz · Applications, limits and what comes next

    • Capstone · Your own area

Your instructor

Dimitris Sykas is a Remote Sensing and Surveying Engineer with a Ph.D. in hyperspectral remote sensing from the National Technical University of Athens (2016). He has spent fifteen years turning satellite data into operational products: at the NTUA Remote Sensing Laboratory, at Planetek Hellas as Senior EO Engineer, at Neuropublic building machine-learning models for smart farming and CAP monitoring, as a postdoctoral researcher on deep learning for agricultural Earth observation at the National Observatory of Athens, and as CTO of the cloudeo group. In 2024 he co-founded GOVADI, an Athens-based company building sovereign, on-premises AI systems for the Greek public sector and large enterprises, where he leads the technical direction of its products: Greek speech-to-text, retrieval-augmented generation with local models and embedding stores, document OCR and PII anonymisation. He has taken part in more than 25 funded EU and ESA projects, has authored over 23 peer-reviewed publications in remote sensing, and holds an international patent in hyperspectral data compression. He founded GEO University to bring hands-on GIS and Earth observation training online.

Dimitris Sykas

Remote Sensing Engineer, Ph.D. · Co-founder and Director, GOVADI · Founder, GEO University

Frequently asked questions

  • Do I need a machine learning background?

    No. Chapter 1 explains features, models, training and embeddings from scratch, for people who know Earth observation.

  • Do I need a GPU or a paid cloud account?

    No. Every exercise runs on a laptop CPU with public data. An Earth Engine account is not required; AlphaEarth is read from Google's open bucket.

  • Which Python skills do I need?

    Reading and running code in a Jupyter notebook, and basic NumPy or pandas. The notebooks are heavily commented.

  • Is the data free to use in my own projects?

    Yes, under open licences: AlphaEarth CC-BY 4.0, TESSERA CC0, Major TOM embeddings CC-BY-SA 4.0, Clay Apache-2.0. Lesson 6.2 explains what each means in practice.

  • How current is the content?

    Product facts are as of September 2026. The field moves quickly; the course notes link to each product's official page so you can check for updates.

  • How long does the course take?

    About 2 h 10 min of video, plus 6 to 10 hours for the notebooks, quizzes and capstone.

  • Do I get a certificate?

    Yes, a certificate of completion once all lessons, quizzes and the capstone are done.

Turn every pixel into a point

Enrol now and download your first embeddings today.

Enrol now