| precision on ordinary days |
all precision above is measured inside mapped floods, where about a tenth of labelled cells on map days are flooded (12% in open land); elsewhere it will be far lower |
score the global random sample (53 blocks, 2023–2025) with GFM as stand-in truth, and the false-event check |
| one dominant basin |
four Bangladesh floods in one basin hold 40% of the gold flood, so pooled scores follow them |
Results shows every score with and without that basin; mapped events from other large floodplains would reduce the weight |
| independence from GFM |
many Copernicus maps are drawn from Sentinel-1, GFM’s own satellite |
partly done: Results scores every product on map days without Sentinel-1 maps (GFM falls to F1 0.12, fusion keeps 0.46); a gold source fully independent of Sentinel-1 is still missing |
| river gauges |
an independent check on timing and on floods the analysts never mapped |
not started; the best-documented gauge network with flood thresholds (USGS/NWS) is in North America, now out of scope. Candidates are in docs/gauge_sources.md |
| small floods |
a flood covering under 30% of a 1 km cell is dry for gold and products alike |
a finer grid, at much higher cost |
| GFM on seasonal floodplains |
GFM’s stricter water definition probably costs recall there |
score GFM on normally-wet versus other cells |
| uncertainty |
per-continent counts are small |
intervals by resampling whole events are beside every score; Oceania (13 events) and South America (16) stay thin, and more mapped events would help |
| FloodScan outside Africa |
the licence for the Asia-Australia and South America files is unconfirmed |
check with whoever holds the AER agreement |
| seasonal water in VIIRS and FloodScan |
whether they count it as flood decides how they compare with GFM on floodplains |
compare each source’s flood-call rate on normally-wet and other cells on ordinary days, month by month, using the built events and the global sample |