Method

How an address becomes a flood record

Every figure on this site should be reproducible by someone who does not trust it. This page is the working: which products, which queries, which rules when the data disagrees with itself, and the places where the method is weakest.

The constraint everything else follows from

Watermark reports observations. It does not model, forecast, score or rank risk. Where an observation is missing it says so and says why, rather than filling the gap with a plausible number. The recurring failure mode in this kind of work is not being wrong — it is being confidently wrong about something nobody checked, so most of what follows is guards.

1. Where has water been?

Source. DEA Water Observations from Space, Geoscience Australia. Every Landsat pass over Australia since 1986, classified per 30-metre pixel. Three products are used: the per-scene layer for individual passes with dates, the calendar-year summary for counts per year, and the all-time summary for the total.

The query. A WMS GetFeatureInfo in EPSG:3857 with a 90-metre box and a 3×3 grid, reading the centre cell. Point queries are quantised onto the analysis grid before storage, so every request landing inside one pixel shares one stored answer.

The date of a pass is not the date in the file

Landsat crosses Australia at about 10am local time, which is 23:00 the previous day in UTC. Taking the UTC date at face value dated every observation one day early, and then joined it to the river gauge's reading for that earlier day. On a rising river, one day is metres.

So the local day is derived from the pixel's own longitude — never from the server's clock, which may be anywhere. This single correction changed the wet count at a reference property from two to one, which then matched Geoscience Australia's own published total for the same pixel: an independent confirmation that it is now right.

Classification order, which matters

The order is: obscured, then masked, then wet, then dry.

Cloudy or Cloud shadow makes a pass obscured — even where the classifier has also flagged water. The classifier runs regardless of cloud, and under cloud its answer is not an observation. Counting those made the site disagree with Geoscience Australia's own totals for the same pixel.
Terrain shadow and High slope are masked.
Only Clear and Wet makes a pass wet. The raw decision-tree flag on its own is not enough: it asserts that the classifier saw water, not that the sensor could see the ground. A pass carrying it with no visibility information is recorded as unknown, which keeps it out of the clear-pass count rather than putting it into the wet count. These records are permanent, so a wet observation on somebody's home has to be earned.
Obscured and masked are never treated as dry.

The neighbourhood ring

An address point can land on a roof or a driveway, and a flood can stop at a fence. So seventeen points are sampled — the centre, eight around a 40-metre ring and eight around a 100-metre ring. This is what lets the site say no water on this exact spot, but water has been observed 100 metres away, which is more useful than either half alone.

Any sentence naming a distance quotes the radius the scan actually reached, never the one it asked for. A partly-sampled ring is reported as partly sampled: saying "nothing nearby" on the strength of two pixels out of seventeen is the same false claim as saying it on the strength of none, only harder to notice.

2. Where is water now?

Source. Bureau of Meteorology Water Data Online, 6,692 stations. The hourly series gives the current reading and trend; the quality-controlled daily series is used to rank a level against the station's own history.

Honesty about latency. This is the Bureau's quality-controlled archive, not its real-time telemetry. Daily readings run roughly ten days behind — so the most recent event is routinely the one still missing. Where a day water was seen has no readable gauge reading yet, that is reported as not yet readable, which is a different fact from the Bureau holds no reading, and the two are never merged.

Context instead of invented thresholds. Official minor, moderate and major flood classifications exist for many stations and are used where they do. Where they do not, Watermark does not invent one. It reports where a level sits against that station's own record — a percentile — because that is an observation rather than a judgement.

3. Ground height

SRTM 1-arc-second, decoded from height-encoded tiles. Roughly ±5 m vertical, 30 m horizontal — enough to tell floodplain from hillside and to rank road low points against each other, nowhere near enough to compare a floor to a kerb. It was flown in February 2000 and is a surface model, so it measures canopy and rooftops over forest and dense housing. See sources.

4. Escape routes, and cut-off height

The question. In a rising flood, do the roads out go under before the house does? A graph is built over every driveable road within 2.5 km, routes out are found, and the lowest point on each is sampled for height.

What it is not. A hydraulic model. It reports a height difference in metres — a measurement — and stops there. An earlier draft finished the sentence with "in a rising flood the roads go under first", which reads as fact but is a hydraulic claim: it assumes a level water surface across the whole area. That holds in a backwater valley and fails wherever the low point sits in a different sub-catchment. The difference in metres is the observation; what water does with it is not ours to assert.

5. The watch level

This is the number that is hardest to get anywhere else, and the one with the most guards around it.

Method. Take the dates water was observed on this ground. Read the gauge for each — at the satellite's own overpass hour where hourly data exists, because on a day a river rose four metres the daily mean and the 10am reading are not the same fact. Then find the highest reading on a day this ground was seen dry. The water arrived somewhere between the two.

Which gauge. Not simply the nearest. Where the nearest named watercourse is known, a gauge on that same watercourse is preferred, because the nearest gauge by straight-line distance is frequently on a different river.

Shown versus acted on

These are answered separately, and collapsing them loses the distinction that matters. A figure may be good enough to display as an observation while not being good enough to act on as a level to watch. Where the gauge's full record could not be read, the observations disagree with each other, or the evidence is a single pass, the number may still be shown as what was seen — but it is not offered as something to set a watch against, and the reason is named.

Known weaknesses. It rests on a handful of observations. Rivers change — channels shift, levees are built, gauges are moved and re-levelled — so a range derived from 1990s observations may not describe today's river. Gauge datums are local, so nothing is ever compared between stations.

6. Address search

Nominatim, run by the OpenStreetMap Foundation, biased to Australia. What you type is sent from this server rather than your browser, and is not stored here. See privacy.

7. When a source cannot be reached

Every upstream answer is classified into one of three states before anything is built from it: answered, answered with nothing, or could not be asked. The third must never be rendered as the second. "No water has ever been seen here" and "we could not reach Geoscience Australia" look identical to a reader if the code lets them, and the first is a claim about somebody's home made on the strength of a bad minute on a network.

An empty answer is also never stored permanently: a point outside the Landsat footprint today may be inside it next year, and a "no" that can never be revisited is indistinguishable from a lie.

8. Reproducing any figure

Every number here comes from a public endpoint you can call yourself. The satellite record for a single pixel comes straight from Geoscience Australia's WMS; the gauge readings come straight from the Bureau's Water Data Online. If a figure on this site disagrees with the agency it came from, the agency is right and I would like to know: siroj.8b@gmail.com.

What this method still cannot tell you is set out at limits, and is worth reading before relying on anything here.