A year of satellite observation, made lighter to carry.

The human problem

Small research teams may need patterns from large satellite archives without the storage and compute to process every raw scene.

The prior constraint

Earth-observation archives are large, multi-sensor, and costly to preprocess for each downstream task.

AI’s actual role

The model compresses a year of radar and optical observations into reusable per-pixel representations.

The documented result

ESA announced the peer-reviewed CVPR work on 4 June 2026; the paper reports 10-meter annual embeddings, and the linked repository provides implementation material. The available sources do not establish that every artifact was first released on that date.

Why it may matter

Reusable representations can lower the starting cost of a local mapping project while leaving validation with the user.

Limitations

An embedding discards information and is not a universal map.

Local crop, fire, or land-use claims require labels and independent validation.

Repository and data licenses must be checked separately.

Unresolved questions

Which regions and seasons show weaker representations?

How much compute and expertise does local fine-tuning require?

What tasks fail compared with raw imagery?