Measurements, where they were taken

effort
per month

by month
of year

0 of 1,415 stations reported this month 0% of the archive's stations

Variable

What the dots show

Frame

Layers, by what they are

Land REFERENCE

Cartography, not measurement — here so the dots can be placed by eye. It is the union of the national land catchments, filled without strokes so no administrative boundary is drawn. Not derived from the water body polygons, whose seaward edges are the boundary under audit.

Station values MEASURED

Class 1–2: sensor value, quantised. Recoverable or boundable error. One dot = one station-month median.

Stations that reported nothing MEASURED

Absence, drawn. Usually invisible on a map, and it is most of the map.

Water body polygons MODEL

Class 7: model-as-datum. A drawn boundary, not an observation. Feature-subspace stability found it adds nothing beyond being a connected region of its size.

Outliers, ringed not removed HEALTH SIGN

An archive with no impossible values is either clean or silently trimmed, and you cannot tell which from inside it. These are the proof it was not trimmed. Amber = beyond the chemistry but explicable; red = gross. Of 599 oxygen values above 15 mg/l, 517 sit between 15 and 20 with a median of 16.2 — bloom supersaturation, not error. An earlier version of this page called them all impossible and was wrong.

Only errors that land outside plausible bounds are ever visible; a decimal slip turning 0.86 into 8.6 is invisible forever. So the ~78 gross values are a lower bound on the gross-error rate and the only observable sample from that process. Trimming them destroys the sole estimator of the archive's own error rate while making it look cleaner — an operation that increases apparent confidence, which is class 7.

Two mechanisms readable in the tail: six bed values at exactly −0.1 mg/l give the sensor's zero offset at anoxia, free. Fourteen clustered tightly at −9.4 to −10.3 are not scattered noise but the signature of a missing-data code leaking through as a value. Neither survives a cleaning pass.

Interpolate between stations YOU ADDED THIS

Inverse-distance weighting. It invents every pixel between the dots. Nothing here comes from the archive — the controls are yours and so is the result.

Key — what every mark means

	Filled dot — one station, at its own position, in the selected month. What it is coloured and sized by is set under What the dots show: the measured value, persistence, or seasonal bias.
	Pale to dark — low to high, always. Value mode: 5th to 95th percentile of that variable. Persistence: 1 to 12 months reported. Seasonal bias: Rayleigh R from 0 to 1.
	Bigger dot — more, in the two count modes: more months reported, or more seasonal. Size is constant in value mode.
	Faint dot — too few records of this variable to characterise (under 12). Not zero; unknown.
	Amber ring — beyond what the chemistry permits but explicable: bloom supersaturation, sensor zero offset at anoxia.
	Red ring — gross value, no benign reading available. Ringed, never removed.
	Small grey dot — a station that exists and reported nothing this month. Off by default, and it is most of the map.
	Solid ground colour — land. Reference only, filled without strokes so no administrative boundary shows.
	Thin amber outline — a water body. A model boundary, drawn only when asked for.
	Soft translucent squares — interpolation you switched on. Every one is invented; none is in the archive.
	Strip bars — observation counts. Upper strip: per month across the record, click to jump. Lower strip: the same counts folded to month of year. The highlighted bar is where you are now.

What this view destroys anyway

depth: 896 → 2 bins
time: day → month
casts: → 1 median
instrument: dropped
time of day: never recorded

Class 3 aggregation, ours, at roughly 3,900:1 against the raw archive. Countable, and counted here rather than left implicit. The measurement-level store that would remove it is about 800 MB.