The live page's data is missing. Build it with
uv run python scripts/build_site.py and open _site/live.html.
floodfusion · Live
Flood maps from Sentinel-1, made by a frozen model, by themselves
Every six hours a scheduled job asks which floods need a map, finds the Sentinel-1 radar passes over them, and maps each one with the model frozen for the blind test: the same weights, byte for byte, and the same threshold. The job reads no Copernicus product, and no map here has been checked against one.
1 · Maps
The maps so far
Pick a map: the flood the model found on that pass (magenta), inside the area it was asked to map (dashed). One radar pass rarely images a whole area: the part it did not image is greyed out, because "not seen" is not "dry". The polygons and STAC items can be downloaded per tile; the full COG layers stay with the job that made them.
No map has been published yet: the watcher's first run publishes here.
2 · How it works
One command, any place and date
- What to map. A list of seen Copernicus floods to show, and every new Copernicus flood activation, watched for 30 days after its event. Of an activation the job reads only its code, name, category, date and the outlines of its areas.
- Which passes. The Sentinel-1 passes over each area, from Microsoft Planetary Computer.
- The inputs, rebuilt exactly as for the benchmark: the post-event pass, its same-orbit reference, the seasonal reference a year earlier, terrain, and Sentinel-2 where it is clear. On a replayed benchmark tile all 14 input bands were identical.
- The map, on CPU, with the frozen weights downloaded from Hugging Face Hub and checked file by file against the freeze manifest.
uv run --frozen python -m floodfusion.live map --bbox 11.8 44.15 12.2 44.3 --date 2023-05-17 --out maps/forli
3 · Blind floods
The newest floods: the model's map, and nothing to grade it
Copernicus flood activations from EMSR899 on are the blind test: the frozen models score them, sealed, until 20 activations in 15 clusters are scored or 31 December 2027. While a second model was in development, this page could show no blind flood, so that no output on one could reach it. That development has ended: fine-tuning on Copernicus's own labels scored lower than the frozen model, so there is no v1.3 model. Since 13 November 2026, on the author's go, blind floods appear here as they come: the model's map alone, with no Copernicus outline and no score. Their scores are computed only by the blind test's own runner, and stay sealed.