The Arena's data is missing. Build it with
uv run python scripts/build_site.py and open _site/arena.html.
floodfusion v1.2 · Arena
Every model, the same floods, and Copernicus's own maps as referee
Instead of asking is this pixel water?, v1.2 asks is this water unusual for this
place and season? Here that idea meets its opponents on equal terms:
a foundation model trained on the same labels, the operational Copernicus flood service, and a blind test on
floods no model had seen, scored only after the models were frozen.
1 · Scoreboard
Copernicus EMS maps drawn on radar satellites, 2023–2026
Every model maps every flood zero-shot
and is scored against the Copernicus EMS delineation of the same day. Five ways to score it:
at the model's own threshold (τ*), the same on the binned scores, after a per-scene recalibration of the
flood prior, at each flood's best threshold chosen in hindsight, and threshold-free
(AP). A win that flips between them is not a win.
2 · Maps
Drag the divider: two models, one flood
Floods drawn by a seeded rule, blind to every score. Inside the region Copernicus
observed (yellow edge), each map is coloured against its delineation.
hit · false alarm ·
miss · edge of the region Copernicus
observed · background: Sentinel-1 VV after the event.
The map quicklooks are rendered from the models' predictions and
will appear once scripts/arena_assets.py has run.
3 · Research landscape (October 2026)
Why a bigger network will not win, and what might
Numbers in this section are reported by others, on their own benchmarks; they are
set apart from this project's results and cited.
The encoder is not the bottleneck
On GEOID-Flood's 219 Copernicus floods, every foundation model lands at 0.45–0.48 flood IoU, and a
32M-parameter Swin-T matches them. The authors conclude that training design matters more than the
encoder; models trained on Kuro Siwo transfer almost as well as on their own set (0.54–0.57 against
0.59–0.60).
Copernicus GFM maps every Sentinel-1 scene with three algorithms and a majority vote, within hours,
and its archive is open. Its next version (2027) adds a deep-learning member and hydraulic flood
expansion. Beating a foundation model on a benchmark matters less than beating this.
New training sets now cover the same Copernicus activations as this project's test set (GEOID-Flood
reaches EMSR871; IBM and ESA's ImpactMesh covers hundreds of events). A model trained on them has seen
the test. Only floods no model was trained or tuned on, scored after it is frozen, stay clean: hence
the blind test below.
What separates a flood from a dark field is what the place normally looks like: its radar a year
earlier, its land cover, where water can flow. Land-cover priors improved every backbone on held-out
floods, TerraMind included.
Two defects found while building v1.2, and what they change
Both were found after v1.1's results were reported. The published numbers stand as
computed; corrected runs are reported beside them, never instead.
5 · Blind test
Floods no model had seen
Every eligible Copernicus flood activation from EMSR899 on (events from late July 2026)
is scored by the frozen models, under rules committed before any of them was looked at. None was seen during
development; each is scored only after the freeze. Scores stay sealed until the stopping rule is met: at
least 20 activations in 15 clusters, or 31 December 2027. That detects gaps of about 0.06 IoU per activation
and 0.07 per cluster; 10 activations would detect only about 0.10.