Quality
The PMEL archive publishes no quality flags on almost all of this data. The 2026 hurricane fleet, every Oshen and every Chance Maritime record carry no QC column at all. Ten older Saildrone datasets carry RH_QC, TEMP_AIR_QC,WND_QC and a data-mode column, and that is the entire QC content of 153 records.
So every quality statement on this site is one this site made. Nine checks run over each record when the site is built, at the finest resolution the build fetched, and each record's page prints the resolution its checks ran at — a check that did not look at a rate must not imply that it did.
A finding marks the data. It never removes or alters it.Everything flagged below is still drawn on its vehicle's page, still in the numbers, exactly as PMEL published it. What is offered is a second opinion, not a correction.
The ten checks
gap The vehicle or its link stopped reporting.
Measured against the record's own median cadence, so ten minutes of silence is a fault on a one-minute Saildrone and normal on a five-minute Oshen.
Not: Not a sensor failing. A sensor that stops while the vehicle keeps reporting is a dropout, and the two send you to different places.
spike A step out and an immediate step back, in one sample.
The threshold is the larger of six robust standard deviations of the step distribution and a floor chosen per quantity — 3 hPa for pressure, 1 °C for sea temperature. The scale is the median absolute deviation, because a standard deviation is computed from the very outliers it is meant to find.
Not: Not a step that stays. A front, a squall or a vehicle entering an eddy moves a record and leaves it moved; nothing physical at these scales moves and returns inside one sampling interval.
stuck One value repeated for hours.
Measured in elapsed time, not in samples.
Not: Not quantization. An Oshen publishes sea temperature to 0.05 °C and humidity to 1 %, so a calm night legitimately repeats a value for many consecutive samples — a check counting identical readings reports every Oshen in the fleet as broken.
range A value outside what the quantity can physically be.
One wide table serves the whole archive, from an Arctic October to a Caribbean August. What it catches is an undecoded −999 sentinel and a unit that was never converted.
Not: Not "unusual". 880 hPa is a category-5 core, and being inside one is the entire point of this fleet.
dropout A sensor missing where the vehicle is not.
Dead and intermittent are reported separately: 90 % of the final twelve hours empty is dead; a high missing fraction with a live tail is intermittent. Sensors that stop within the same day are grouped, because a payload losing power is one event and not twelve.
Not: Not sparseness. A Saildrone's CTD reports every five minutes into a one-minute record — 80 % of rows empty, and a perfectly healthy instrument.
timeorder The clock runs backwards between consecutive rows.
Every consecutive pair is compared. A record substantially out of order is reported more severely than one with a stray row: one in a thousand is the line.
Not: Not wrong values, and not a reason to distrust a mean or a range — those do not depend on the order the rows arrive in. It matters for anything that assumes the clock only moves forward: a track through consecutive fixes, a difference between neighbours, a rate, a spectrum. It is also not reported for a record that interleaves several vehicles, where stepping backwards is the shape of the table rather than a fault.
cadence The reporting interval changed part-way through the record.
Compared as a ratio, so 1→2 minutes counts as much as 30→60.
Not: Not a fault in itself. It matters because it changes what every other statistic over the whole record means.
position Missing fixes, the null island, and jumps no USV could make.
A Saildrone makes about 3 m/s downwind; anything over 8 m/s between two valid fixes is a bad fix, not a fast vehicle. 0°N 0°E is called out on its own — it is in the Gulf of Guinea, and it drags a track across an ocean the vehicle never entered.
Not: Not a reason to distrust the measurements in those rows.
metadata Units missing, damaged, or contradicted by the values.
Reported, never silently corrected. Oshen humidity declares “units = 1” and publishes percent; Chance publishes no units at all on pressure and chlorophyll; two datasets carry “¡C”, which is “°C” written as Mac Roman and read back as Latin-1.
Not: Not a criticism of anyone. These are near-real-time products from active missions.
silent A record that was recently reporting and has stopped.
Only within a month of the last report. Beyond that a record is archived, not silent, and the catalog already shows when it ended.
Not: Not a lost vehicle. A recovery, a link failure and a finished mission look identical from here.
What they found
| Record | Campaign | Findings | Worst | What |
|---|
Counts come from the catalog, which each record's page expands into the findings themselves. The about pagesays where the data came from and what was converted on the way.