Bulk training data

Bulk NRW training imagery

Canonical source: versioned documentation. This wiki copy is maintained manually.

The bulk collector prepares additional photos for experimental training review. The registered protocol targets 1,000 distinct photos: 600 current captures at original PV4GER training locations and 400 ALKIS geometry crops separated from original training centres. The selection record binds the reference metadata and coordinates before the first new RGB request.

The collection is now complete: 1,000 photos, independently verified, with forty review batches per lane. Capture dates span 2022–2026; current source availability does not mean every crop is newly flown. The raw collection retains all-ignore placeholders. Subsequent mask curation and quality exclusions created a separate V4 training snapshot with 482 additional partial masks; the model is unchanged. See the collection audit and final mask audit.

Current captures at known training locations may show changed roofs or arrays. They are additional photos of existing training locations, not independent new roofs. No old mask is read or transferred. Geometry crops use the existing small-roof rules; neither path guarantees PV presence. This is a convenience sample for training, not a representative evaluation set.

Acquisition

Install the geo and train extras. Run from the repository root:

uv run --locked --extra train --extra geo pvextent data fetch-training-bulk \
  data/pvextent-training-v2 data/nrw-training-bulk-v1 \
  --protected data/nrw-negative-holdout-candidates-v1 \
  --identity-only data/nrw-training-schleiden-v1 \
  --identity-only data/nrw-training-duelmen-v1 \
  --identity-only data/nrw-training-marsberg-area-v1 \
  --protocol docs/research/nrw-training-bulk-protocol-v1.json \
  --cache-dir data/cache/nrw-training-bulk-v1 --json

The protocol fixes quotas, seed, 30 geographic boxes, candidate bounds, distances, concurrency and request budgets. Known locations are ranked deterministically and interleaved across projected 10 km cells, without reading source masks or using model predictions. The operational plan is sealed before RGB collection. Every crop must have a dated metadata tile covering its full 32 m extent and confirming native 0.1 m resolution. Accepted images preserve the exact 320 × 320 RGB PNG response, without resizing.

Reference IDs, encoded image hashes and building IDs prevent identity reuse. The original non-training records and all twenty external-pilot cases are protected by at least 5000 + 32√2 metres. Their raster paths are never opened. Only training images are decoded to reject pixel-identical recaptures. The existing 22 training-review images and new crops have at least 80 m centre spacing; known-location candidates also have 200 m seed spacing. Geometry candidates keep at least 80 m from original training centres.

Successful official responses enter a verified immutable cache with URL, exact bytes, SHA-256 and original retrieval time. A persistent ledger reserves every network attempt before requesting it. Up to four workers share a minimum 0.35 s request interval, at most three attempts per URL and a 5,000-request run budget. There are no additional retries inside the cache. Retrieval time and snapshot assembly time are distinct.

On interruption, rerun the same command with the same references, protocol and cache. Cached responses and attempt counts survive; corrupt entries fail rather than silently downloading replacements. The final dataset is published only when both quotas are met. --allow-partial explicitly permits fewer images; it must not be described as a completed 1,000-image collection. Existing output directories are never replaced. Per-image acquisition evidence stays outside Git.

Review queue

uv run --locked --extra train pvextent data review-queue \
  data/nrw-training-bulk-v1 artifacts/nrw-training-bulk-review-v1 \
  --batch-size 25 --seed 20261021 --json

For 1,000 photos this prepares 40 batches in each of two reviewer lanes. Each lane contains every image exactly once. Opaque IDs and a restricted image manifest hide acquisition origin, inherited labels and model predictions. Queue export reads metadata only. It creates pending templates, not reviews: full_image_inspected starts false, no polygons are present and no labels change.

Assign distinct reviewers, inspect original pixels, record native-pixel PV and clear-background polygons independently, and retain uncertain areas as ignore. The current proposal schema approves positive cases; clearly negative cases need the separate negative curation workflow. Empty polygons do not confirm PV absence. Two completed reviews, consensus curation and a new immutable training snapshot are required before a separate training experiment. AI review is explicitly experimental, not human-confirmed benchmark truth.

All raw collection masks contain only 255; its samples remain unreviewed and unassigned. The separate final V4 snapshot now contains 482 qualifying partial masks (246 positive and 236 background-only), with 518 new cases still entirely ignored. Raw source and review evidence remain unchanged. Regular checkpoint resume continues to reject changed dataset fingerprints.