floodfusion v1.2: Sentinel-1 flood maps, frozen for a blind test
The frozen model of floodfusion, a flood-mapping system built from
Sentinel-1 radar, with optical used only where it is clear. These are the weights it was frozen with for its
pre-registered blind test on Copernicus Emergency Management Service floods, three seeds, byte for byte the files
that reports/v12_freeze.json pins at the tag v1.2.0-freeze.
Still floodfusion's current model. Version 1.3 (November 2026) tested fine-tuning these weights on Copernicus EMS's own flood labels. Under an adoption rule fixed in advance, the fine-tuned model scored lower (0.380 against 0.421 macro IoU, cross-validated over 38 activations) and was not adopted, so there are no v1.3 weights.
- Network (row 21): SAR–optical cross-attention fusion with terrain and a seasonal Sentinel-1 reference: the median backscatter of the same orbit around the same date a year earlier, and the change from it. 72.7 MB per seed.
- Training labels: Kuro Siwo plus GEOID-Flood, 7,006 chips from 138 floods in 81 location groups.
- Calibration: temperature scaling fitted on held-out validation groups, plus a separate calibration for windows without usable optical. The decision threshold τ* = 0.25 follows from a stated miss-to-false-alarm cost ratio of 3.
Files
| Path | What it is |
|---|---|
runs/v12/final/all/all/row21_fusion_context_seasonal/seed{42,43,44}/best.pt |
weights (torch.load(..., weights_only=True)) |
…/seed{42,43,44}/run.json |
the full configuration, its hash and the training record |
…/seed{42,43,44}/eval_test.json |
the calibration and τ* |
runs/v12/final/all/all/row21_fusion_context_seasonal__eval_sar_only/seed{42,43,44}/eval_test.json |
the calibration for windows without optical |
SHA256SUMS |
sha256 of every file above, equal to the freeze manifest's |
Use
git clone https://github.com/teo-clerk/floodfusion && cd floodfusion && uv sync --frozen
# the weights, each file checked against the freeze manifest, at the commit that added them to this repository
uv run --frozen python -m floodfusion.weights fetch --repo teo-clerk/floodfusion \
--revision faed3bb78668cf25052211a7e834f34652a4cdf6 \
--run runs/v12/final/all/all/row21_fusion_context_seasonal/seed42 --dest .
# a flood map of any place and date (CPU is enough): COG layers, a STAC item and flood polygons per tile
uv run --frozen python -m floodfusion.live map --bbox 11.8 44.15 12.2 44.3 --date 2023-05-17 --out maps/forli
The weights and live commands are in release v1.3.0 and
later. The project's live page shows the maps a scheduled job
makes with these weights.
The map command reads Sentinel-1 RTC, the Copernicus DEM, JRC surface water and ESA WorldCover from Microsoft Planetary Computer, and Sentinel-2 where it is clear. The live map rebuilds the benchmark's input stack bit for bit: on one replayed benchmark tile (EMSR683 AOI02), all 14 input bands were identical. On the benchmark's own stack, a CPU run gives the stored GPU run's IoU at τ* (0.414). The live stack has no area-of-interest mask, so the network also reads the ground around the area; on the same pixels it scores 0.421.
Results (seen floods, descriptive)
On the 38 Copernicus EMS flood activations of 2023–26 drawn on radar satellites (EMSR664 excluded), scored once after
the freeze (reports/v12_cems_rebuilt.md):
| Model | Macro IoU at τ* | Oracle-threshold IoU | AP |
|---|---|---|---|
| This model (v1.2, row 21) | 0.417 | 0.468 | 0.598 |
| TerraMind-base + LoRA, trained the same way | 0.405 | 0.446 | 0.583 |
| v1.1's best (fusion, Kuro Siwo flood labels) | 0.407 | 0.438 | 0.527 |
| Physics baseline (thresholds on backscatter change) | 0.373 | — | — |
On the 34 of those floods that Copernicus Global Flood Monitoring covers, this model scores 0.419 and GFM 0.321.
Only the gap to GFM is significant (reports/v12_cems_seen_tests.md).
How high could anyone score? Two Copernicus EMS delineations of the same area, made at most 24 hours apart,
agree with each other at 0.520 macro IoU over 34 seen activations (reports/reference_ceiling.md).
The confirmatory test is still running. Copernicus flood activations from EMSR899 on are scored by these frozen
weights and sealed until 20 activations in 15 clusters are scored, or 31 December 2027 (reports/protocol_v1_2.md).
Intended use and limits
- Use: a first-guess flood extent for a rapid-mapping analyst after a flood, from the first Sentinel-1 pass; large-area screening, with an analyst reviewing flagged and low-confidence areas.
- Not seen by C-band radar: water under closed canopy and between buildings. Dense forest and built-up land are flagged not reliably mapped; "dry" there means "not seen". Radar shadow and layover are excluded.
- Not estimated: depth, damage or impact. The map is a snapshot at the acquisition time.
- Do not deny assistance to an area marked "dry" under canopy, in a city, in radar shadow, or outside the acquisition time. Every product ships with its exclusion mask and low-confidence band.
The repository's MODEL_CARD.md lists the known failure modes (partial cloud, threshold transfer, wind-roughened water).
Training data and attribution
- Kuro Siwo annotations (Bountos et al., NeurIPS 2024), CC BY 4.0, pinned at
Orion-AI-Lab/KuroSiwo-annotations@b930dd27. - GEOID-Flood labels (Chiriaco et al., 2026), CC BY 4.0, pinned at revision
868407460bf3oflinks-ads/geoid-flood: "contains modified Copernicus Sentinel data and Copernicus EMS information". - Sentinel-1 RTC from Microsoft Planetary Computer (CC BY 4.0). Contains modified Copernicus Sentinel data.
- Copernicus DEM GLO-30, produced using Copernicus WorldDEM-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA. ESA WorldCover 2021, CC BY 4.0.
- Evaluation only: Copernicus Emergency Management Service rapid-mapping delineations and Global Flood Monitoring, © European Union. Contains modified Copernicus Emergency Management Service information (2023–2026).
- No Sen1Floods11 data or labels were used to train these weights.
Licence and citation
The weights are released under CC BY 4.0, the licence of their training labels; the code is MIT. Cite the
repository with its CITATION.cff.