Building Footprint Extraction website update — September 24, 2026

These notes identify the local research records used for the post. Run identifiers
are provenance references, not public download links. The underlying complete
research dataset and checkpoints have not been deposited as a public release.
CSV metrics are fractions (0–1); percentages in the article multiply them by 100.

Historical baseline:
  Feng96/personal_website, commit 4e01ce28d3c8df71b922f6b4e316d517c9744f98
  content/post/U-NET_building_footprint/index.md and its original six PNGs.

Regional scientific validation and six added figures:
  GBA_V6circle_V7_reference_night_v003_20260911
  manuscript/revision_20260911/tables/regional_primary.csv
  manuscript/revision_20260911/tables/modification_inventory.csv
  manuscript/revision_20260911/figures/Figure_1 through Figure_6
  Filter: scenario=current_reference, evaluation_group=ALL_WITHIN_REGION.
  Regional IoU and BF1 are pooled, not object means. BF1 tolerance is 1 m.
  Figures show September 11 scientific V7, not the later OSM-gated production.
  The main comparison comprises 389 windows. The processed 48,772-object cohort
  includes 60 development cases. Selected image examples are not random samples.

Modified-object comparison:
  PAPER_WRITING_HANDOFF_20260918/05_VALIDATION_AND_REPRODUCTION.md
  NUMERIC_TABLES/modified_fixed_pairs.csv and DERIVATION.json
  1,371 modified objects with fixed matches; 516 other modified objects excluded.
  Mean object IoU: GBA 0.784533278, V7 0.724981061.
  +/-0.01 threshold: 344 increased, 908 decreased, 119 neutral.

Model methods and earlier results:
  PAPER_WRITING_HANDOFF_20260918/02_MODELS_AND_TRAINING.md
  PAPER_WRITING_HANDOFF_20260918/03_INPUT_DATA_AND_PREPROCESSING.md
  PAPER_WRITING_HANDOFF_20260918/04_V7_POSTPROCESSING_AND_PRODUCT.md
  B0_real_NAIP_transfer_night_v001_20260909
  NAIP_rural_continue_v001_20260917
  NAIP_rural_overnight_v002_20260918

Latest completed model experiments:
  NAIP_balanced50k_v003_20260918/TRAINING_DONE.json
  NAIP_balanced50k_resume_v006_20260920/TRAINING_DONE.json
  NAIP_LR_height_factorial_v001_20260921/DONE.json
  NAIP_LR_height_longqueue_v002_20260921/PARTIAL_COMPARISON.json and STATUS.json
  NAIP_B0_R0_hybrid_v001_20260922/DONE.json
  NAIP_targeted_sampling_v001_20260924/DONE.json
  The companion development CSV uses final_validation only, not best_validation.
  naip_val=Orange County; ma_val=Massachusetts; new_val=Colorado/Connecticut.
  These are repeatedly used development data, not independent national tests.
  Architecture routes have different initialization/normalization/LR histories.
  Targeted sampling uses equal updates and exposures, intentionally different
  sample identities. Its conditional second-seed follow-up was not triggered.

Inputs and frozen production probabilities:
  USBF_AC_multinode_v004_20260913/COMBINED_STATUS.json
  USBF_AC_probability_v001_20260914/unified_20260916/STATUS.json
  997,524 registered; 997,522 ready with probabilities; 2 RGB-held windows.
  Complete/partial/no height: 417,927 / 257,999 / 321,596.
  Scope: AL/AZ/AR/CA/CO/CT. Not national completion or training-set size.

Production snapshot:
  USBF_V7_cpu_resume_v002_20260921/STATUS.json
  updated_unix=1790281176.642932 (2026-09-24 16:19:36 EDT)
  2,249 complete delivery cells; 7,462,544 committed buildings; 176,795 modified.
  OSM policy: V7_OSM_circle_confirmation_v1_20260918
  Geometry policy: V7_geometry_source_fallback_v2_zero_area_20260919
  Counts are an explicitly dated internal status snapshot, not a public release
  or measured precision. Later runs may report larger counts.
  Production still uses the frozen September 9 B0/B1 checkpoints.

Source and release boundaries:
  Original GBA geometry, height attributes, and source IDs are retained.
  Modification counts are not certified improvements. Reference sources may
  share lineage with GBA/OSM and may differ in acquisition date.
  The production OSM circle policy has not undergone the same four-region
  re-evaluation. No DOI, full national release, or downstream application gain
  is claimed for USBF in this article.

Companion file SHA256 hashes (unaltered research figures and numeric extracts):
39168f70b27b776efed8087d3b68d88952c60c601a6d6255b5e9768eaa15d227  Figure_1_coverage.png
55d5cf1f9f85ecc5828893c00b5a85ed1d788f20857b9634cc7d31e6e4daaa85  Figure_2_workflow.png
0065c783ab9468e21150e7ce44cf4d01be37330236ca2c8ed1adce097beb34ff  Figure_3_regional_agreement.png
d5415c02075423dda4cb2134c061b72423b16d47e97301598dfa4c2650ace7bb  Figure_4_curved_and_courtyard.png
ced7285c28277361320ff643e42d225e51a82023616a1d4fc6a95db474b70865  Figure_5_local_boundary_changes.png
2f81a01bb8c0205858299ca061297187870a4b8d0a8ca6232926d221f1b64ce6  Figure_6_failure_modes.png
b7a3a69e9547dd34d8519cf43c4264cfc6feda146e94ca6a3fc119c30aae07d7  development-experiments.csv
0b5cc6e22afb5607ff18b5871a276e9f2749c7fd7bb3f455ec7337d117784104  regional-validation.csv