1.1.0 (2026-06-11):
- Detection now runs as a background QgsTask: QGIS stays responsive,
with live progress and Cancel in the task manager. The first-run
weights download is cancellable too.
- New Processing provider (beacongis:detectbuildingdamage): batch
mode, headless qgis_process runs, and Model Designer integration,
with polygon sink + optional damage-mask GeoTIFF outputs.
- Pre/post alignment is now verified geographically (CRS + extent +
size) instead of by array shape — same-size rasters covering
shifted extents are warped onto the pre grid instead of being
silently compared misaligned.
- Automatic pre/post co-registration: global phase-correlation + ECC
translation alignment, plus bounded per-tile phase-correlation
refinement, attacking the parallax/misregistration failure mode on
non-xBD pairs.
- Confidence calibration support: cls logits are temperature-scaled
before softmax when a calibration.json (fit on a held-out xBD
split via calibrate_temperature.py) ships next to the weights.
- Memory: scenes are held as uint8 and ImageNet normalization happens
per tile — several GB lower peak on 100+ MP scenes.
- Removed legacy dead parameters/code paths; misc fixes.
1.0.2 (2026-06-05):
- Plugin Repository security scan compliance: explicit URL scheme
allowlist (http/https only) before every urllib.request.urlopen call
in building_damage_engine and satellite_downloader, with matching
nosec B310 annotations; pragma: allowlist secret comments on the
three published SHA-256 model checksums (which are integrity hashes,
not credentials).
1.0.1 (2026-06-05):
- QGIS 4 / Qt6 readiness: bumped qgisMaximumVersion to 4.99, removed
deprecated supportsQt6 flag, replaced Qt5-only .exec_() calls with
the cross-version .exec() spelling.
1.0.0 (2026-06-02):
- Initial public release on the QGIS Plugin Repository.
- Two-network architecture: LocalizationUNet + SiameseUNet ensemble
(M1+M2).
- SeResNeXt-50 (32x4d) encoder, ImageNet -> Inria Aerial -> xBD
training.
- Combined F1 = 0.7312 (xView2 official scorer, full plugin pipeline).
- Two-model softmax-averaged ensemble for more robust per-pixel
predictions.
- CPU Fast Mode for laptops without GPU (~8x speedup, ~0.5pt F1 cost).
- GSD normalization to ~0.5 m/px training scale.
- Watershed-based touching-building separation with user-tunable
strength.
- Optional 4-way TTA (default off), Hanning-blended tile inference,
AOI clipping.
- GeoPackage / per-pixel mask GeoTIFF / JSON metadata sidecar exports.
- Assessment Reports dock with offline template-driven SitRep
generation (no LLM API).
- Satellite imagery downloader dock (ESRI World Imagery / OSM).
- Distribution: ONNX Runtime (no PyTorch required at runtime).
- Weights distributed via GitHub releases with SHA-256 verification.
- GPU acceleration via onnxruntime-directml or onnxruntime-gpu
(auto-detected).

None