Introduction

eo-data-embedding is not a classical L0/L1/L2 instrument processor; its data-processing model is a batch-then-serve embedding pipeline: a one-time, GPU-oriented encode pass turns Sentinel-1/2 imagery into vector embeddings that are persisted once, after which every downstream capability (similarity search, few-shot probe, bitemporal change detection) operates cheaply over the stored embeddings.

Processing chain

        flowchart LR
    tile["tile<br/>(C,256,256)"] --> embed["embed<br/>frozen Clay v1.5 ViT<br/>(N,1024)"]
    embed --> store["store<br/>Parquet"]
    store --> search["search<br/>FAISS cosine"]
    store --> probe["probe<br/>few-shot vs CNN"]
    store --> change["change<br/>Δembedding, bitemporal"]

    classDef core fill:#e8f5e9,stroke:#2e7d32,color:#11270f;
    classDef out fill:#fff3e0,stroke:#ef6c00,color:#3a2400;
    class tile,embed,store core;
    class search,probe,change out;
    

Processing chain — a one-time encode pass, then a fan-out to three consumers over the stored embeddings.

1. Tile — raw scene (C,H,W) (S2 reflectance / S1 backscatter) split into fixed (C,256,256) tiles (change.tile_image); EuroSAT arrives pre-tiled.

2. Embed — encode normalizes per verified band stats and runs the frozen encoder under torch.no_grad() → class-token (N,1024) (or per-patch tokens for spatial change maps).

3. Store — one Parquet row per tile (id, modality, vector[, label]); reloaded via load_embeddings + stack_vectors into the dense (N,1024) matrix.

4. Fan-out — the stored matrix feeds three independent consumers: search (FAISS IndexFlatIP, cosine), probe (few-shot LogisticRegression vs a ResNet-18 baseline), change (zero-training delta-embedding distance + supervised probe, held-out threshold).

Detailed per-module algorithms, parameters and equations are in the Software Design Document and the Interface Control Document; this DPM captures the top-down decomposition and processing flow.