You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

xBD-S12 SUT Wrapper

Custom Inference Endpoints handler that serves the prs-eth/xbd-s12 building-damage models (3-seed localization + damage ensembles, Siamese/UNet on Sentinel-1 + Sentinel-2) behind an HTTP API compatible with Resaro AIP's huggingface_image_classification SUT protocol.

The inference code under src/ is vendored from github.com/prs-eth/xbd-s12 (MIT, © Olivier Dietrich) — see LICENSE.upstream.

How it works

At container boot, the handler downloads the six prs-eth/xbd-s12_{loc,dmg}_seed{1,2,3} checkpoints and syncs a patch-store dataset repo (PATCH_DATASET_REPO env var, default peterachua/xbd-s12-eval-patches).
Per request, it reads image_id from the JSON body (the posted image bytes are ignored), resolves the four co-registered GeoTIFFs from the patch store, runs the ensemble, and reduces the per-pixel damage raster to a patch-level [{label, score}] distribution.
Reduction rule: damaged-pixel fraction over all patch pixels, one-hot into four ordinal buckets — <10% → no_damage, 10–30% → light_damage, 30–70% → moderate_damage, ≥70% → heavy_damage. Patches with no detected buildings have fraction 0 and land in no_damage. Golden-dataset labels must use this same vocabulary.

Patch store layout

Native Zenodo xbd_s12 layout — build the dataset repo by uploading a subset (e.g. the event_split == "test" uids from xbd_s12_metadata.geojson) of the extracted archive. image_id in requests = the xBD-S12 xbd_uid.

s1/<xbd_uid>_pre_disaster_s1.tif     # Sentinel-1, 2 bands (VV/VH), 4 m resampled
s1/<xbd_uid>_post_disaster_s1.tif
s2/<xbd_uid>_pre_disaster_s2.tif     # Sentinel-2 L2A, 12 bands, 4 m resampled
s2/<xbd_uid>_post_disaster_s2.tif

Request / response

curl https://<endpoint>.endpoints.huggingface.cloud \
  -H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" \
  -d '{"inputs": "", "image_id": "guatemala-volcano_00000000"}'
# → [{"label": "no_damage", "score": 0.0}, {"label": "light_damage", "score": 1.0},
#    {"label": "moderate_damage", "score": 0.0}, {"label": "heavy_damage", "score": 0.0}]

Errors (missing image_id, incomplete patch quartet) raise, returning a non-2xx response — intentionally loud so evaluation runs record a SUT error instead of a fabricated label.

Deployment notes

CPU instance is sufficient (checkpoints total ~140 MB); enable scale-to-zero for demo-cadence use and warm the endpoint with one curl before an evaluation run (cold start ≈ 1 min: checkpoint + patch-store download).
Security: Protected. Register in AIP with auth_type="bearer" and an HF token.
This SUT is dataset-coupled: it can only answer for image_ids present in the patch store. A passing connection test proves reachability, not lookup completeness.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support