Context overview

eo-data-embedding sits between Sentinel imagery and the analysis layer: it is the stage that turns pixels into reusable vector embeddings. Its inputs are raw Sentinel-1 backscatter / Sentinel-2 reflectance scenes (or pre-tiled benchmark imagery such as EuroSAT); its output is a persisted embedding store that every downstream capability reads from. It is therefore not an inline instrument processor in the L0/L1/L2 sense but a batch-then-serve stage whose product (the embedding store) is the interface to all consumers.

Within the processing chain its role is a single encode pass followed by a fan-out:

• Upstream — a raw scene (C, H, W) is split into fixed (C, 256, 256) tiles, normalized per verified band statistics, and passed through the frozen Clay v1.5 ViT encoder under torch.no_grad() to yield a class-token embedding (N, 1024) per batch.

• Persistence — the embeddings are written once to a Parquet store (id, modality, vector[, label]), which is the boundary between the one-time GPU encode pass and the cheap, repeatable serving phase.

• Downstream — the stored matrix is consumed independently by three capabilities: search (FAISS cosine nearest-neighbour), probe (few-shot classification), and change (bitemporal delta-embedding detection). None of these re-runs the encoder; they operate purely over the stored vectors.

The end-to-end processing flow and the precise step-by-step decomposition are given in the DPM Introduction; the key processing parameters are tabulated in the Parameters data list.