sentinel3-data-repo

Sentinel-3 OLCI ocean color for the US East Coast — chlorophyll-a, and the harmful algal bloom indicators built on it — published as the map data contract's grid files.

It publishes, since 2026-08-31. pipeline/products.toml declares one product and the workflow builds it; the first green run put a 7-day merged composite at /map/chl-s3.json with a native-resolution tile tier beside it, and the site draws it as Chlorophyll-a (Sentinel-3).

The schedule is on, since 2026-09-19: three attempts a day, at 03:41, 11:41 and 19:41 UTC, the first to find a new overpass builds it and the others stop on the probe. (It had been left off for nineteen days after the two dispatched runs of 2026-08-31 met the condition for turning it on, and the map drew a 2026-08-29 overpass the whole time; PLAN.md has the record.)

PLAN.md is the founding plan and running record — the upstream, the measurements, and what is open. DECISIONS.md indexes the dated one-way decisions. Which document gets what, and what "update docs" means across all ten repositories, is the doctrine block at the top of CLAUDE.md — the same text in all ten, held equal by the site's check:docs.

What it publishes

Chlorophyll-a over the US East Coast, as a 7-day merged composite:

source	NOAA CoastWatch, Copernicus Sentinel-3A and 3B OLCI, near-real-time
datasets	noaacwS3AOLCIchlaSectorFIDaily, noaacwS3BOLCIchlaSectorFIDaily
variable	chlor_a, mg m⁻³
native resolution	0.0025° ≈ 278 m
extent	lat 29.81–45.19, lon −80.04 to −59.96 — Florida to Nova Scotia
upstream cadence	one frame per day per satellite, ~14:30–15:20Z
upstream latency	2–5 days — S-3A 2, S-3B 5, measured 2026-08-31
upstream history	90-day rolling window — not an archive

The tributaries are mostly empty, and that is the upstream, measured 2026-09-02. On a clear day the Bay mouth carries a value in 76% of its pixels and the James River in 4.7%, because the OC4Me retrieval fails within about 5 km of every shore. Nothing here masks them; PLAN.md section 4 has the measurement and section 5 the alternatives.

Both open questions closed on 2026-08-31, as DECISIONS.md D5 and D6. The window is seven days; the published product is Sector FI whole at the instrument's own 0.0025°, as a tile tier under a 0.05° overview, with values linear at two decimals.

What that costs, measured the day it was decided:

overview	402 × 308, 602 KB
per-cell age, same grid	568 KB
tiles, native 0.0025°	367 tiles, 233 MB
coverage of observable water	76.8%
per-cell age of the freshest observation	median 2 d, p90 4 d, max 6 d
upstream per build	~1.7 GB, one native frame is 188 MB in 10.7 s

Why a composite, in one table

Coverage of everything ever observable, by composite length, measured 2026-08-30 across 36 frames and five weeks (both satellites, 0.15° sampling, lat 30–45, lon −80 to −66):

composite	S-3A only	S-3B only	merged A+B
1 day	11.8%	5.9%	17.3%
3 days	31.6%	18.1%	43.5%
5 days	46.6%	30.3%	62.5%
7 days	58.1%	40.5%	74.2%
10 days	73.5%	55.9%	86.6%
14 days	87.3%	70.7%	96.3%

Single-day coverage over a Chesapeake box across six consecutive days was 19.9%, 12.6%, 0.0%, 7.4%, 7.2%, 0.5% — one day in six completely blank. That is what makes the composite a requirement rather than a preference, and merging the two satellites is worth roughly a doubling of the single-satellite window.

The full method, and the caveat that late August is neither the cloudiest nor the clearest season on this coast, are in PLAN.md.

How it will run

The same arrangement as espc-model-repo, and for the same reasons:

• No code here. The orchestrator (pipeline/orchestrate.py) comes from realtime-data-repo; the fetchers and the published-file contract (schema.ts) come from oceansensing.github.io. Both are checked out at run time, so a change to either lands here on the next run rather than on any push here.

• This repository carries pipeline/products.toml — what it publishes, and nothing else executable.

• Its own cron and its own fault domain. A cloudy fortnight over the Atlantic must not hold back the currents, and HYCOM's outages must not hold back this.

• Dispatch a run: Actions → the publish workflow → Run workflow, or gh workflow run publish.yml --repo oceansensing/sentinel3-data-repo. The schedule runs it three times a day since 2026-09-19; a dispatch is for not waiting.

• Read the health: curl -s https://oceansensing.org/sentinel3-data-repo/status/status.json | python3 -m json.tool

• What it publishes: chl-s3.json and chlage-s3.json, the 7-day composite and its per-cell age, with a native-resolution tile tier beside each — see pipeline/products.toml.

Structure

PLAN.md         the founding plan and running record; measurements live here
CLAUDE.md       what must not be got wrong here, and the shared doc doctrine
DECISIONS.md    dated one-way decisions, D1 onward
pipeline/       products.toml — the one product and its two roots

Licensing and attribution

The published data derives from Copernicus Sentinel-3 OLCI, processed by EUMETSAT and served by NOAA CoastWatch. Any map drawing it owes that attribution, and the site's attribution control is where it goes. The upstream's own terms travel with the data: CoastWatch's ERDDAP states the data may be used and redistributed freely but is not intended for legal use and may contain inaccuracies.