9 Gold labels
A gold map gives each cell it examined one label for its map date: flooded, dry or none. Cells it did not examine are never scored.
Each map has two parts: the flood polygons and the area the analyst examined. Both are drawn onto a 0.001° grid, so each 0.01° cell holds 10 × 10 = 100 sub-pixels.
| label | rule |
|---|---|
| flooded | at least half the cell was examined, and at least 30% of the examined part is flood |
| dry | at least half examined, under 30% flood, and under 30% “flood possible” |
| none | under half examined; or examined, under 30% flood and at least 30% “flood possible” |
Maps of the same event on the same day are merged, and flood wins over dry.
Where the thresholds come from. The 30% flood share and the at-least-half rule are the same for gold and for every product, so a cell is “flooded” by one definition on both sides. As a result, a flood covering under 30% of a cell is invisible to scoring (see Not yet validated). These thresholds, the 30-day and 60-day periods, the 10% daily fading and the P ≥ 0.5 cut-off were all set in the project plan before any result was seen. None is tuned to the results.
9.0.1 The days before and after each event
Gold maps exist only on flood days. To learn how often each source raises a false alarm on ordinary days, the model also needs dry labels away from the flood. These come from the 30 days before an event’s first map and the 60 days after its last.
- In those days, every cell the maps examined is labelled dry,
- except cells that were flooded on any map day in their window. Those get no label, because floodwater takes weeks to drain. In open land in the first run (D-39), fusion still called flood on 48% of days 1–30 after the event in those cells, against 5% in cells that never flooded. This rule was decided on 14 October (ADR 0006); the old rule’s scores are shown alongside.
These days are used to fit the false-alarm rates and to check reliability. Recall, precision and F1 are scored on gold-map days only.
What the labels cannot tell us. Analysts map what they saw; unmapped water inside the examined area counts as dry. Many Copernicus maps are drawn from Sentinel-1, the same satellite as GFM, which favours GFM. Results are therefore also reported for the events whose maps are not from Sentinel-1.