Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Numerical representations of imagery for similarity, clustering and lightweight classification

1. Context

This is the final method lesson in SDS320, and it closes the course with a different kind of shortcut than L12’s. Segment Anything let you skip training by prompting a foundation model directly for a specific task. Satellite embeddings let you skip training by reusing a foundation model’s internal representation of a place, a compact numerical summary you compute once and then use for similarity search, clustering, classification, or change detection, largely as fast, lightweight operations on vectors rather than new deep learning pipelines.

Everything from L05 through L12 asked a foundation model, or a trained model, to answer a specific question directly: what class, what box, what mask, what change. This lesson asks a different question first: what does this place look like, numerically, and lets you answer many downstream questions cheaply once you have that answer.

2. Motivation

Every earlier method lesson in this course involved some form of task-specific effort: labeling, training, or at least careful prompting. Embeddings offer a genuinely different economy. A handful of labeled points, sometimes fewer than a hundred, can train a working classifier on top of a pre-computed embedding, because the foundation model has already done the expensive work of learning what matters visually and spectrally about a landscape.

For a course project with limited time and limited labels, this matters directly. If precomputed embeddings already exist for your study area, you may be able to explore, cluster, and even classify your data without training anything from scratch, or committing to a specific architecture at all.

3. Learning objectives

By the end of this lesson, you should be able to:

4. Lesson roadmap

  1. When embeddings fit — typical use cases, and when a task-specific model is still the better choice.

  2. Embedding concepts — what an embedding vector is, why similar places end up close together, and patch- versus pixel-based formats.

  3. The embedding registry — browsing and filtering geoai’s catalog of available embedding datasets.

  4. Clay embeddings — the main hands-on workflow: downloading and loading patch-based Clay Foundation Model embeddings.

  5. Visualizing & searching — PCA visualization, K-means clustering, and cosine similarity search.

  6. Classification & change — training lightweight classifiers and comparing embeddings for change detection.

  7. Pixel & cloud embeddings — TESSERA pixel-based embeddings and AlphaEarth in Google Earth Engine.

  8. Project transfer — deciding whether embeddings can support your project, and planning your analysis.

Pages 1–2 build the conceptual foundation, page 3 helps you find the right dataset, page 4 is the main hands-on workflow, and pages 5–7 extend it with unsupervised exploration, classification, and alternative embedding types. Page 8 transfers everything to your project.

5. Project framing

While reading this lesson, keep your own project in mind and consider:

6. Before class

7. After this lesson

After working through this lesson, and this course, you should have:

Spatial Data Analytics
Project transfer
Spatial Data Analytics
When to use embeddings