https://fieldsofthe.world/ftw-webgpu/
This app runs the Fields of The World CC-BY PRUE field boundary model on chips from the FTW benchmark, entirely in the browser with onnxruntime-web. It picks a random chip from the benchmark on Source Cooperative, stacks its two Sentinel-2 windows, runs the model, and draws the predicted field and boundary classes over the imagery. Each prediction is scored against the chip's 3-class label. Changing a slider reruns inference right away, which shows how the predictions depend on the normalization divisor, the input scale, and the sliding-window patch size.
- Model. The FTW PRUE CC-BY U-Nets from the FTW v3.1 release (EfficientNet-B3, B5 or B7), exported to ONNX and pinned to commit
ffe4452. They were trained only on the CC-BY and CC0 countries of the dataset and are licensed CC-BY-4.0. The model downloads once and is cached in the browser. - Input. Sentinel-2 bands B04, B03, B02 and B08, with window B first and window A second. This is the order
ftw_toolsstacks them in training (temporal_options="stacked"). Reflectance is divided by the normalization divisor (3000 in training). - Output. Per-pixel logits for background, field and boundary. The app shows their argmax.
- Chips. The FTW benchmark: 256 × 256 GeoTIFFs for
window_a,window_bandlabel_masks/semantic_3class, fetched directly fromdata.source.coop. The default split is test, which the model never saw. Label licenses vary by country and are shown with each chip.
- Normalization divisor. What reflectance is divided by before inference.
- Window order. Feeds the windows in training order or swapped.
- Input size. The chip is bilinearly resampled to this size (a multiple of 32, from 128 to 1024 px) before inference. Upsampling approximates a larger tile, or imagery at a finer resolution than the model was trained on.
- Patch size and overlap. An input larger than the patch runs as a sliding window. Overlapping logits are blended with a linear ramp before the argmax.
The URL hash keeps the chip and settings, so a link reopens the same view.
The WebGPU backend is used when the browser has it. Otherwise inference runs on WebAssembly, using threads when the page is cross-origin isolated. GitHub Pages can't set the COOP/COEP headers that isolation needs, so coi-serviceworker adds them, and the page reloads once on its first visit.
Raster-to-polygon conversion of the predictions isn't implemented yet.
npm ci
npm run dev # http://localhost:5173/ftw-webgpu/
npm test # unit tests (Node's built-in runner)
npm run build # typecheck + production build into dist/public/chips.json is the per-country chip index (ids by split, window dates, label license). Rebuild it from Source Cooperative with python scripts/build_chip_index.py, which needs duckdb.
Pushing to main deploys to GitHub Pages through .github/workflows/deploy.yml.
The code is MIT. The model weights are CC-BY-4.0 from Fields of The World. Chip imagery is Copernicus Sentinel-2 data, and each country's labels keep their source license.