[RafaelCardoso990]

What this adds

models/unet_roads/: a fine-tunable base model that segments roads in VHR RGB imagery and returns the road surface as polygons. Open call category: Transportation infrastructures. It supports both inference and fine-tuning.

Only models/unet_roads/ is touched.

How it works

• Network: U-Net with a ResNet34 encoder (segmentation-models-pytorch), one road probability channel per 256 x 256 chip.
• Inference: the probability maps of all chips in a request are pasted into one mosaic, thresholded and vectorised, so a road is one polygon across chip edges. Polygons are simplified and filtered by area.
• Training: polygon labels are burned as drawn; line or point labels are buffered to road_width_m. BCE + Dice loss, flips and rotations, plus zoom and colour jitter. Spatial block split.
• Weights: https://huggingface.co/Rafael990/unet-roads (checkpoint and ONNX, CC-BY-SA-4.0).

Training data and results

Base stage on SpaceNet 3 roads (Khartoum, Paris). Fine-tuning stage on OpenAerialMap drone imagery with OpenStreetMap highways (Maputo, Freetown, Dar es Salaam).

Held-out set	IoU	Relaxed F1 (3 px)
SpaceNet 3 Khartoum	0.53	0.44
SpaceNet 3 Paris	0.56	0.20

The OSM reference on the drone sets is incomplete, so I also reviewed the predicted masks by eye, chip by chip. Chip-level accuracy among chips with a road: 80% on a hard sample of 60 Khartoum validation chips, 78% on 90 Maputo chips (mostly in the fine-tuning set) and 69% on 120 Banepa chips, which the model never saw in training. Details and limitations are in the model card.

What I ran locally

Same commands as the workflows, on this branch, with no changes outside models/unet_roads/:

• style: ruff check, ruff format --diff, ty check, pytest for fair (526 passed, 1 skipped)
• validate-stac: validate_stac_items.py (including the asset URL checks) and validate_model.py
• build-model-images: test image, 11 step tests inside it, models/test_integration.py, runtime and inference images, and both inference smoke tests (published ONNX and fine-tuned ONNX) on the Banepa bbox
• test-model: I ran the flow on a local kind cluster (just up, just example unet_roads, scripts/test_serve.py). Registration, fine-tuning in the pod, promotion, batch prediction and the serving test all passed, with one deviation: the image is not in ghcr.io/hotosm/fair-models, so for that run I pointed mlm:training and mlm:inference at the digest of the image loaded into kind. The STAC item in this PR uses the normal :latest tags.

What I expect to fail in CI, and why

As in #68 and #77, a PR from a fork cannot push to ghcr.io/hotosm/fair-models:

• Build Model Images stops at "Push runtime image".
• Test Single Model stops at registration with Image digest resolution failed: ghcr.io/hotosm/fair-models/unet_roads:latest ... 403 Forbidden.

Could a maintainer run these from a branch on your side?

Notes for reviewers

• fair-py-ops version: the Dockerfile pins 0.3.17, like the other models. With 0.3.25 the integration test fails for this model at promotion (Evaluation metrics do not match the base model's fair:metrics_spec): models/conftest.py mocks the metric logging for every test, and promotion now validates those metrics. With the mock disabled the test passes on 0.3.25.
• Polygons, not centrelines: the contributing guide lists LineString for roads, but examples/run.py trains every model on the buildings-banepa-* polygon sample, which a line model fails the compatibility check against. I kept this PR to polygons so it runs with the existing sample data. A centreline variant of the same network is ready (mask skeletonised and traced into LineStrings), and I can open a follow-up PR if you tell me how you would like line models tested.
• Road width: the training labels are centrelines buffered to 6 m, so the polygons show where a road runs, not its measured width.

Checklist

• stac-item.json declares fair:metrics_spec, fair:split_spec and fair:hyperparameters_spec
• keywords has a feature tag, a task value and exactly one geometry type
• license is a supported SPDX identifier (CC-BY-SA-4.0)
• just validate passes for the model and STAC item
• README follows the model card sections

[RafaelCardoso990]

Hi @kshitijrajsharma, this PR adds a road surface segmentation model for the Open Call for EO GeoAI Models. The application form and the email to fair@hotosm.org were sent on 5 November.

The workflows are waiting for maintainer approval. Could you approve them when convenient? As with #68 and #77, "Build Model Images" and "Test Single Model" will also need the model image in ghcr, which a PR from a fork cannot push. Everything else passes locally, as described in the PR.

Thank you!