Tracks, time series and quality control for every Saildrone, Oshen and Chance Maritime uncrewed surface vehicle published on the NOAA PMEL ERDDAP — 153 deployments, 2017 to now, active and historic.
A static site. There is no server and no database: the observations are fetched from PMEL when the site is built, checked, and written out as plain JSON that your browser reads from this origin. Opening a vehicle is one request for one file — and narrowing to a stretch of a mission fetches that stretch at the rate the instruments actually reported, one minute on most of the fleet.
All 153 records back to 2017, searchable by vehicle, campaign or institution, and filterable by vendor, year and what the quality checks found. The map draws each mission's actual track, coloured by the company that built the vehicle, with a dot at each one's last known position — filled where it is still reporting. Click any track to open it.
- The track, coloured by any variable the vehicle carries, with your own colour scale and range.
- Stacked time series of everything it measured — wind, pressure, air and sea temperature, humidity, salinity, oxygen, chlorophyll, waves — on one shared time axis, so a pressure fall and a wind rise line up.
- A property–property scatter, any variable against any other, coloured by a third: T–S for the surface water mass, wind against pressure for a storm.
- The quality report — every issue found in that record, on the figure and in a table.
Narrowing to a stretch reloads every figure at full resolution. The page opens on an overview of eight thousand points however long the mission, says how many full-rate samples exist, and fetches only the weeks you have windowed into — about 380 KB each. A three-day window on a Saildrone is 4,320 samples where the overview had 720.
Narrow any figure to a stretch of the mission with its own range boxes. The variables on screen, the track's colour, its scale and its range all live in the query string, so every view is a link.
The vehicles that flew together, on one map and one axis: the 2026 hurricane fleet's nine Saildrones and thirteen Oshens, the Nantucket Shoals survey, a TPOS season. Only the quantities the whole cohort carries are offered, and each one was resolved from its vendor's own column name and converted to a common unit first — which is the only reason a Saildrone and an Oshen can share an axis at all. Without it one is in m/s at 3.4 m and the other in knots at 0.66 m, and they look like different weather.
The PMEL archive publishes no QC flags on the 2026 fleet, on any Oshen and on any Chance record, so every quality statement on this site is computed here and says so. Ten checks run over each record when the site is built — gaps, spikes, stuck sensors, impossible values, dead sensors, position dropouts, cadence changes, metadata faults and cohort outliers — and the findings are ranked fleet-wide on the quality page and drawn on the figure they belong to.
Nothing is ever silently altered. A finding marks the data; it does not remove it.
Every figure and the map export a publication-quality PNG — 3× resolution, title and caption drawn in, boxed, on white, and for the map the colour bar and basemap attribution too. Every dataset links to its own ERDDAP page for the full-resolution numbers.
npm installnpm run dataFetches the catalog and every overview from PMEL — about 40 minutes cold,
and the only step that needs the network. One request at a time, because the
server refuses concurrent ones from a single client. npm run data:catalog
alone takes seconds and is enough to build the site with an empty fleet; and
npm run data:series -- --sample 8 builds a spread across the vendors and
eras, which is enough to work on the pages.
npm run data:detail -- --season 2026Builds the full-resolution tier for one season — every sample, in weekly gzipped chunks. It is published from its own repository; see below.
npm run devnpm run verifyBuilds, type-checks, gates the documentation and runs the nine offline test suites — 929 checks. Nothing in it touches the network.
Two tiers, in two kinds of repository.
| what | where | |
|---|---|---|
| overview | 8,000 points per record, every record | built into this site |
| full rate | every sample, weekly chunks | one repository per season |
The full-rate archive is 794 MB gzipped, against GitHub Pages' 1 GB
published-site limit — so it cannot sit beside a site that is itself 250 MB.
It lives in oceansensing/usv-data-<season>, each published as its own Pages
site.
Those are the same origin as this one: a project repository under an
organisation with a Pages custom domain serves under that domain, so
oceansensing.org/usv-data-2026/ sits beside oceansensing.org/usv/. The
detail tier therefore costs no cross-origin request and no relaxation of this
site's content security policy.
A season repository holds no code. Its workflow checks this one out and runs
npm run data:detail, which needs no token because both are public. And
because a closed season cannot change, its shard is built once and never
rebuilt — only the current season is on a schedule.
Published: all ten seasons, 2017 through 2026 — one repository each,
every one serving. This said "published so far: 2026, the other nine are the
same command with a different --season" long after the other nine had been
run; NEXT.md had it right and this file did not.
Note that a season repository's root is a 404 by design — it holds
chunks, not a page — so the way to check one is to ask for a chunk a record
actually spans. sd1005_2017 covers weeks 2487–2523 and every one of those
answers; asking for week 0 would 404 on a perfectly healthy shard, which is
an instrument problem rather than a publication one.
packages/erddap-pmel/ the PMEL tabledap client: catalog, metadata, queries
packages/usv-vars/ 429 vendor column names → 63 canonical quantities,
and which season repository holds a record's full rate
packages/usv-qc/ the ten quality checks, and how a finding is ranked
packages/plot/ the SVG plot engine, colormaps and PNG export
packages/teos10/ seawater properties, from the GSW definitions
scripts/ the data builds, and the test suites
src/ the site
CLAUDE.md is the engineering note — what rests on what, and which of the
obvious designs turned out to be wrong. PLAN.md is the running record of
decisions.
Built by Donglai Gong (Virginia Institute of Marine Science, William & Mary) at the Collaboratory for Physical Oceanography, with Claude (Anthropic).
The observations are NOAA's — PMEL, AOML and NEFSC with Saildrone Inc., Oshen and Chance Maritime Technologies — published on the PMEL ERDDAP, which each page links to and each dataset's own licence and acknowledgement govern. Sibling sites: gliders for the IOOS Glider DAC, and NOAA-USV-analysis for the campaign analysis this visualization was drawn out of.
All rights reserved; see LICENSE.