README.md

metadata

title: Flood mapping
emoji: 🌊
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
short_description: In-browser Sentinel-2 flood mapping with WorldFloods v2

Flood mapping — WorldFloods v2 · Sentinel-2 · in-browser

Draw a study area, pick a pre-flood baseline scene and a post-flood scene, and run the ml4floods WorldFloods v2 U-Net entirely in your browser — no server, no account, no API keys.

• Scenes: public Sentinel-2 L2A archive on AWS Open Data, searched via the Element84 Earth Search STAC API and read as cloud-optimized GeoTIFFs with HTTP range requests (geotiff.js).
• Model: WF2_unetv2_bgriswirs (6 bands: B2, B3, B4, B8, B11, B12) exported to ONNX with normalization baked in, executed with onnxruntime-web (WebGPU when available, WASM otherwise).
• Flood definition: water probability ≥ threshold in the post scene on pixels that were land in the baseline scene, excluding thick cloud, clipped to the drawn polygon.
• Outputs: flood polygons (GeoJSON), flood mask (GeoTIFF), overlay PNG, stats JSON — all generated client-side.

Caveats

• The model was trained on Sentinel-2 L1C (TOA); here it runs on L2A (surface reflectance) from the open archive. Results are close but validate against known events before operational use.
• Optical method: clouds are blind spots. Scene ranking uses scene-level cloud cover; the app reports the cloud share inside your AOI per run.
• Study areas are capped at 1,000 km² (browser memory). A run downloads ~50–120 MB of imagery.

Full method description, assumptions and references: METHODOLOGY.md.

Built by IWMI. Model and method: spaceml-org/ml4floods.