Purpose#

eo-data-embedding is a multi-modal geospatial embedding search & change-detection toolkit built on a frozen Vision-Transformer foundation model (Clay v1.5). It turns Sentinel-1/2 imagery into reusable vector embeddings and exposes four capabilities over them:

Similarity search — find look-alike scenes/tiles by cosine similarity (FAISS).
Few-shot classification — a linear probe on frozen embeddings that reaches strong accuracy with very few labels per class (vs a from-scratch CNN baseline) — the foundation-model label-efficiency benefit.
Bitemporal change detection — zero-training Δembedding distance and a supervised change probe.
Plug-and-play CPU demo — a Gradio UI requiring no GPU and no model at runtime.

../_images/demo_search.png — Similarity search in action — each query tile (left) and its nearest neighbours retrieved by cosine search over frozen Clay v1.5 embeddings (EuroSAT). See the test report for the per-class retrieval, confusion-matrix and label-efficiency figures.#

Benefits: embed once (the only heavy/GPU step), then run every downstream task cheaply over the stored embeddings; label-light adaptation; reproducible, config-driven runs; portable across laptop/CPU, Colab, Kaggle and cloud.