eo-data-embedding

Status: Released Stack: Python 3.11 · PyTorch · Clay v1.5 · FAISS · Gradio Lifecycle: ECSS-E-ST-40C tailored

Similarity search demo

Overview

eo-data-embedding embeds Sentinel-1/2 imagery once with a frozen Clay v1.5 Vision Transformer, then serves every downstream task cheaply over the stored vectors — similarity search, few-shot classification, and bitemporal change detection — with no per-task fine-tuning and no GPU at query time.

The design goal is label-efficient, honest EO ML: embed-once architecture, CPU-only query path, and negative results reported rather than hidden.

Architecture

flowchart LR
    Tiles[EO tiles S1/S2] --> Extract[Clay v1.5 extract]
    Extract --> Store[(Embedding store)]
    Store --> Search[Similarity search]
    Store --> Probe[Few-shot probe]
    Store --> Change[Change detection]

Key properties:

Frozen foundation model — Clay v1.5 ViT; no per-task fine-tuning at query time.
CPU-only demo — eo-data-embedding demo downloads a public bundle and serves a Gradio UI without GPU.
ECSS-tailored lifecycle — ECSS-E-ST-40C engineering + DRD set, ECSS-E-HB-40-02A ML V&V, ECSS-Q-ST-80C product assurance.

Key technical work

Multi-modal tile ingestion and normalisation for Sentinel-1 SAR and Sentinel-2 MSI inputs.
Vector extraction pipeline with deterministic, reproducible embedding storage.
FAISS-backed cosine similarity search over precomputed embeddings.
Few-shot linear probe with label-efficiency evaluation across training-set sizes.
Bitemporal change detection via supervised Δembedding classifier; zero-training distance baseline explicitly rejected when at chance level.

Results

Task	Metric	Result
Similarity search	mAP@10	0.774
Similarity search	precision@10	0.822
Few-shot probe	macro-F1 (50 labels/class)	0.895 ± 0.011
Few-shot probe	macro-F1 (full train pool)	0.92
Change detection	supervised Δembedding F1	0.510 (honest, validation threshold)
Change detection	ROC-AUC	0.640

eo-data-embedding

Overview

Architecture

Key technical work

Results

Links

Documentation (GitHub Pages)

User documentation

Project documentation