Coverage

Availability gates. The measured share only warns. Everything else on the page follows from that, and it is a deliberate choice.

A span carries two coverage figures, and they answer different questions.

Availability

What share of the span carries a value at all. For a gap-filled product that is 100 % wherever the product covers the span, and 0 % where it does not.

Measured share

What share came from the instrument rather than from the gap-filling model.

Every statistical gate reads availability

Normals, ranks, anomalies, badges and trends are computed wherever the product covers the span, at whatever measured share. TA_F, NEE_VUT_REF, GPP_NT_VUT_REF and the rest are the series the community publishes and analyses. A page that quietly declined to use them would describe a sparser record than the file is of.

Two further gates apply to the comparisons. They are about how many years there are, not how much of each:

• a calendar-month normal, and every anomaly, z-score and rank taken from it, is withheld below MIN_NORMAL_YEARS (8) qualifying years, so a sparse month is ranked against nothing and can never be “the driest on record”;

• a slope is not stated below TREND_MIN_YEARS (10) qualifying years, and the two-epoch split needs EPOCH_MIN_YEARS (5) complete years in each half.

The measured share warns

Below its variable’s warning line, a span is hatched on the grid, carries the sparse badge, is marked tile-thin in the month panel, and is counted in the build’s console output. From the CH-LAE example:

warning: SW_IN is under 50 % measured in 10 of 252 months (lowest 0 % in January 2005);
  the gap-filled values are used and those months are marked on the grid

The figure is used either way, and the reader is told what it rests on.

Two lines, because two things are being measured

Key	Variable	Unit	A span is its	Warns under
TA	Air temperature	°C	mean	50 %
PREC	Precipitation	mm	total	50 %
SW_IN	Incoming shortwave radiation	W m⁻²	mean	50 %
VPD	Vapour pressure deficit	kPa	mean	50 %
RH	Relative humidity	%	mean	50 %
SWC	Soil water content	%	mean	50 %

Key	Variable	Unit	A span is its	Warns under
NEE	Net ecosystem exchange	g C m⁻²	total	20 %
GPP	Gross primary productivity	g C m⁻²	total	20 %
RECO	Ecosystem respiration	g C m⁻²	total	20 %
LE	Latent heat flux	W m⁻²	mean	20 %
H	Sensible heat flux	W m⁻²	mean	20 %

Meteorology is measured almost continuously when it is measured at all, so a month that has fallen under half measured is a real outage and worth flagging.

A turbulent flux cannot reach that, and not because anything failed. Half of every record is night, and u* filtering rejects the calm nights by design, so a half-hourly NEE that is around half measured is a good one. Warning a flux at the meteorological line would flag every month of every flux record ever produced, which says nothing. The line sits instead where a month stops being a thin measurement and becomes very largely model.

What this replaced

Both gates used to read the measured share, at thresholds set for meteorology. That gave two silences.

• No month of any flux record reaches the meteorological threshold, so no flux normal, rank, anomaly, trend or badge was ever computed, and the sparse badge landed on every tile.

• Even at a threshold lowered for the fluxes, some calendar months had fewer than MIN_NORMAL_YEARS qualifying years, so they had no normal at all and stood blank on the anomaly grid.

Both are gone. Every variable of a real record now publishes a trend over its whole span, where before most were withheld for want of qualifying years.

What using the filled values costs

The measured share is not constant through a record. Where it rises over the years, as it does at CH-Oe2, annual GPP correlates about as strongly with coverage as it does with time, so the GPP and RECO slopes are entangled with the coverage improving.

The trend note used to claim that a slope is never a picture of changing coverage. With the old gates gone, so is that guarantee, and the note now says which variables lean on the filling and by how much.

The case that shows the gate still bites

RH has no QC column, so its gaps are genuinely missing rather than filled. Its availability is 73.5 %, and its trend is withheld at 6 complete years. Correctly.

The normal is not stationary, and the page says so

Every anomaly, standard score and rank is taken against the mean of the whole record. That mean is not a fixed climate. Over twenty-one years the record itself moves, so a baseline drawn from all of it sits between the early years and the late ones, and a late month is compared against a climate that is partly no longer current. Unstated, the effect reads as weather.

So the trend is computed rather than assumed, published beside the grid, and never folded into the comparison. Theil-Sen slope and Kendall’s tau are taken down each column of the grid and over the record as a whole, by fluxatlas.stats.trend(). The CH-LAE dashboards draw the same estimator, so the two cannot disagree about the slope of a series. The foot row carries the slope under each calendar month. The two-epoch split states the same thing as two numbers by halving the record, which makes the size of the effect legible next to the single normal the rest of the page compares against.

The baseline is not selectable, and that is deliberate. Badges are evaluated against the whole-record normal, so a page that let a reader re-baseline the tiles would show tiles and badges disagreeing about the same month. One baseline serves every claim, and the trend is published as the fact that qualifies it.

Where this lives in the code

thin_spans computes the counts, report_thin_spans prints them, meta.thin carries them to the page, and variables[].cov carries the thresholds because the renderer reads them too. fluxatlas.variables.coverage() answers the thresholds for a key from the registry alone.