6 Inputs
Fusion uses three daily satellite sources plus the gold maps to test against. Every source is turned into one value per 0.01° cell per day: flood, dry or “couldn’t see”. “Couldn’t see” is never counted as dry.
| input | what it is | pixel | years | read from | cell is flood when | cell is “couldn’t see” when |
|---|---|---|---|---|---|---|
| GFM | Copernicus Global Flood Monitoring, Sentinel-1 radar | 20 m, read at 80 m | 2015 on | EODC STAC (stac.eodc.eu), collection GFM: layers ensemble_flood_extent, exclusion_mask, advisory_flags |
at least 30% of its observed pixels are flood | under half its pixels observed: no Sentinel-1 pass that day, excluded pixels, or GFM’s own snow/frozen-ground flag |
| FloodScan | AER flooded fraction (SFED) from passive microwave | about 9 km | 1998 on; outside Africa to 2026-03-31 | Africa: prod blob raster/floodscan/daily/v5/processed/ (daily COGs). Asia-Australia and South America: dev blob raster/floodscan/daily/v5/raw/ (one NetCDF per region) |
the 9 km cell it sits in is at least 30% flooded | outside FloodScan’s regions, or after its archive ends (2026-03-31 outside Africa). Inside, it has no cloud mask and always answers |
| GFDS | JRC/GDACS Global Flood Detection System, passive microwave | about 9 km (0.09°) | 1997 to 2026-09-14 (feed ended); used from 2015 | gdacs.org/flooddetection/DATA: daily signal GeoTIFF plus GFDS’s long-term mean |
the signal’s drop below its long-term mean implies at least 30% water | days GDACS has no file for, and pixels GFDS masks itself |
| microwave vote | FloodScan and GFDS combined into one vote, since they measure the same thing | either says flood | both couldn’t see (possible only outside FloodScan’s regions) | |||
| VIIRS | NOAA VIIRS flood map, optical | 375 m | 2012–2020 and 2023-08 on (no files 2021-01 to 2023-07) | AWS s3://noaa-jpss/, VFM_1day_GLB |
at least 30% of its observed pixels are flood | under half its pixels observed: cloud, shadow, snow or ice, no data |
| gold maps | analyst flood maps: Copernicus EMS rapid mapping and UNOSAT | vector, burned at 0.001° | 2015 on | dev blob global/copernicus_ems/flood/gold/ and global/unosat/gold/ |
see Gold labels | not examined by the analyst |
From pixels to cells. Each product pixel goes to the cell that holds its centre. A cell is “couldn’t see” if fewer than half its pixels were observed that day. Otherwise it is flood if at least 30% of the observed pixels are flood, and dry if not. FloodScan and GFDS pixels are larger than a cell (about 9 km), so every cell copies the value of the 9 km pixel it sits in.
Why the sources have gaps.
- GFM: Sentinel-1 passes over a given place only every 6 to 12 days, so GFM has no image on most days.
- VIIRS: it sees reflected sunlight like a camera, so cloud blocks it, and floods come with cloud.
- Passive microwave (FloodScan, GFDS): it measures the energy the ground itself gives off. It sees through cloud every day, but only at about 9 km.
The in-cell rules (at least half observed, at least 30% flood) are the same for every source and for gold. A river that floods 10% of a cell is therefore dry everywhere. MODIS was tested only as an experiment; see Other approaches tried. FloodScan is used under AER’s commercial licence, and whether it covers the non-Africa regions is still to be confirmed.