The entire Sentinel 2 archive translated into stac-geoparquet enabling scalable bulk access using any tool that understands Parquet. Syncs daily with Earth Search API and uses its COG's as the assets.
Guidance for AI agents and automated clients working with the Sentinel-2 L2A STAC-GeoParquet mirror.
One rule governs each edit to this file. A claim here is either quoted from a source or measured from the data. If you cannot point at where a fact came from, it does not belong in this file. An agent acting on an invented column name or an invented join key produces a confident wrong answer that nothing downstream catches.
Prefer https:// for one file you can name, and s3:// when you need
DuckDB to expand a year=*/*.parquet glob. DuckDB refuses a glob over plain
https://.
This catalog publishes two item indexes, each with a stats collection beside it.
sentinel-2-c1-l2a mirrors Earth Search's Sentinel-2 Collection 1 index, which
is ESA's reprocessing of the archive. It is sorted tile-major and joined on
_tile rather than s2:mgrs_tile. Its collection.json counted 30,402,025
rows, 2015-10 to 2026-09, when read on 2026-09-22. stats-c1 aggregates its
MGRS-tile statistics. Query Collection 1 first, because it is the uniform
record. Upstream:
Earth Search sentinel-2-c1-l2a.
sentinel-2-l2a mirrors the original Earth Search Sentinel-2 L2A item index as
year-partitioned STAC-GeoParquet. Earth Search counted 51.25 million items on
2026-09-15. stats aggregates its MGRS-tile statistics. Upstream:
Earth Search sentinel-2-l2a.
Read the agent guide of the collection you query before you query it. Each one
documents its schema, its query pattern, the NULLs to expect and the contract
for the assets JSON-string column. This page repeats none of that.
Coverage in sentinel-2-l2a is partial before December 2018, with no items for
2015 and 2016 and part of 2017 and 2018. That gap belongs to Earth Search's
record. Do not report a missing year as an observation about Sentinel-2 itself.
sentinel-2-c1-l2a reaches back to October 2015 instead, because ESA
reprocessed the archive to one baseline.
This catalog mirrors its upstream. Earth Search, run by Element 84 on the AWS Registry of Open Data, produces the item index republished here, and no step in this pipeline filters, reclassifies or interpolates it.
Assets and structural links resolve relative to the object that carries them.
Catalogs here publish no self link, so a client tracks its own location.