Promote focused small-building detector

2026-07-13 12:15:27 +02:00
parent b3f7c3ca63
commit 1689dce928
20 changed files with 546 additions and 61 deletions
@@ -7071,3 +7071,68 @@ Open:
 - Embedded PostGIS live migration smoke passed with PostGIS `3.6`, required tables/indexes, database collation and the single Alembic head `202606120900`.
 - Browser validation confirmed that the recommended profile selects the exact active asset and threshold `0.15`, while live preflight displays CPU dependency/model readiness and the expected missing-manifest guard before dataset handoff.
 - Browser console warnings/errors: `0`.
+
+# Sprint 174 - Focused small-building recovery and promotion
+
+## Data and training evidence
+
+- Converted the Sprint 173 persistent false-negative audit into one focused real-data experiment instead of extending the same corpus blindly.
+- Added Beerse, Rijkevorsel, Hoogstraten and Vorselaar as training AOIs and Vosselaar/Grobbendonk as independent tile-level validation AOIs.
+- Kept Turnhout, Retie and Westerlo outside the tile corpus as operation-level holdouts.
+- Exported `/app/storage/operator-data/yolo-building-aoi1024-smallbld-minpx3vis035` from an explicit 23-sample manifest subset:
+  - 198 retained tiles;
+  - 180 positive and 18 negative tiles;
+  - 58,820 real GRB-derived labels;
+  - 48 visually reviewed tiles;
+  - zero invalid labels, missing images, missing label files or low-variance review selections.
+- The accepted `min-label-px=3` corpus retained 1,228 more genuine small-building labels than the comparable `min-label-px=4` export.
+- Trained one inactive 30-epoch CPU candidate from the previous active local model:
+  - model: `geointel-building-yolov8s-smallbld-minpx3-img640-ft30.pt`;
+  - model SHA256: `a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1`;
+  - dataset-summary SHA256: `49b2a07d2105d08356431757b83eafc1498eaf1fb76965b1efe05b776824942a`;
+  - dataset-YAML SHA256: `3a2ea97c35a18072a1ab6738cd673c0ecec5344b19461c91d72a15e138d46e8d`;
+  - no model download and no fake training or QA data.
+
+## Persisted promotion evidence
+
+- Evaluated the exact fixed profile `tile=512`, `overlap=64`, `confidence=0.15` with QA match IoU explicitly fixed at `0.25`.
+- Seven positive AOIs produced:
+  - mean precision `0.5898197518`;
+  - mean recall `0.5769921004`;
+  - mean F1 `0.5824578632`;
+  - minimum per-AOI F1 `0.5527837436`.
+- Every AOI improved F1 relative to the previous balanced model. Turnhout improved from `0.4897494305` to `0.5527837436`.
+- The strict pure-empty gate covered Postel, Lommel and Arendonk and produced zero detections for every sample.
+- The formal promotion report recommended the exact key `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt|512|64|0.15`.
+- Fixed-reference object evidence used identical GRB feature identities and reduced false negatives from `7,753` to `6,182`:
+  - 1,571 fewer total false negatives;
+  - 745 fewer misses in the 25-100 m2 bucket;
+  - 181 fewer misses below 25 m2;
+  - all seven AOIs improved.
+- Remaining persistent misses total `5,838`, concentrated in Turnhout, Herentals and Geel and still dominated by small buildings.
+- Mean precision decreased from `0.6470590036` to `0.5898197518`. The new profile is therefore a recall-balanced operator default with a higher false-positive review load, not ground truth.
+
+## Repository hardening
+
+- Added explicit `--samples` / `OPERATOR_YOLO_SAMPLES` corpus selection with selected/excluded sample provenance and unknown-sample rejection.
+- Added persistent false-negative area statistics, size buckets and combined GeoJSON review evidence.
+- Copied the complete operator evaluation/promotion toolchain into the all-in-one image and added a regression that rejects every Docker `COPY scripts/...` source that does not exist.
+- Removed Pydantic protected-namespace warnings for legitimate `model_*` API fields while preserving all schema field names and response contracts.
+- Updated Detection Lab profiles: the new small-building profile is recommended, the previous expanded profile remains the higher-precision legacy choice, and the background-aware `0.35` profile remains conservative.
+- Guarded activation first returned `ready_to_apply`; the reviewed `--apply` pass updated only `GEOINTEL_INSTALL_AI`, `YOLO_ENABLED`, `YOLO_MODELS_DIR` and `YOLO_MODEL_PATH` in the Tower environment.
+
+## Local validation
+
+- `python -m compileall backend/app`: passed.
+- `python -m pytest`: 472 passed.
+- `python -m ruff check` for all changed Python modules/tests: passed.
+- `npm run typecheck`: passed.
+- `npm run build`: passed; app bundle `215.64 kB`, MapLibre bundle `801.82 kB` before gzip.
+- `bash scripts/run_readiness_check.sh`: passed with 472 tests.
+- `python -m alembic heads`: one head, `202606120900`.
+- `python -m alembic upgrade head --sql`: complete migration chain rendered successfully.
+- Shell syntax checks passed for live migration and the full operator evaluation/promotion chain.
+
+## Next recommended pass
+
+- After redeploy, verify the active model SHA, local model-load preflight, live PostGIS migration smoke and browser profile selection. Then review false-positive evidence and the remaining 5,838 persistent misses before any further training.