Auto-City-Research - Damage Is Not Need

Auto-City-Research

Code: github.com/Ireliya/auto-city-research
Data: huggingface.co/datasets/Ireliya/auto-city-research
Project: ireliya.github.io/auto-city-research

This repository is the lightweight, privacy-safe reproducibility dataset for the Urban Cup 2026 Competition 2 project:

Damage Is Not Need: Auditing Post-Disaster Priority Disagreement with Multi-Source Urban Evidence

Research Scope

The study audits where a ranking based only on satellite-observed building damage disagrees with transparent multi-source priority scenarios using population exposure, road accessibility, critical facilities, and urban form. It does not provide a ground-truth unmet-need label or an operational emergency allocation model.

Four xBD/xView2 event footprints are included: Hurricane Harvey, Mexico earthquake, Palu tsunami, and Santa Rosa wildfire.

Headline Verification

• xBD building records: 99,629
• Reference 500 m cells: 1,448
• WorldPop 100 m percentile-consensus disagreements: 73
• WorldPop 100 m exact Top-20% disagreements: 115
• Legacy WorldPop 1 km checks: 67 and 109
• Cells passing all fixed non-temporal gates: 4, all in the Mexico footprint
• Final candidates with supportive historical OSM persistence: 0
• Harvey external-proxy coverage: 149 SVI/NFIP tracts, 10,134 aggregated NFIP claims, and 41 FEMA assistance ZIP aggregates

The four candidate keys are fixed in data/derived/final_consensus_v1/final_consensus_candidates.csv. The external proxy results are mixed and are released without relabeling them as validation success.

Main Contents

data/derived/xbd_core_v1/                    Parsed xBD building labels
data/derived/xbd_damage_grid_v1/             Reference 500 m damage grid
data/derived/worldpop_context_v1/             WorldPop 1 km comparison
data/derived/worldpop_context_100m_v1/        WorldPop 100 m primary population join
data/derived/priority_mismatch_v1/            Legacy 1 km priority audit
data/derived/priority_mismatch_100m_v1/       Primary 100 m priority audit
data/derived/evidence_hardening_*/            Four baselines and 20,000 weight draws
data/derived/multiscale_*/                    Independent 250/500/1,000 m reconstructions
data/derived/population_resolution_audit_v1/  100 m versus 1 km diagnostics
data/derived/historical_osm_v1/               Pre-event/current OSM sensitivity
data/derived/harvey_external_validation_v1/   Aggregated NFIP boundary test
data/derived/external_proxies_v1/             CDC SVI and FEMA RI-IHP tests
data/derived/final_consensus_v1/              Fixed-gate candidate table and manifest
data/derived/study_overview_v1/                Event, scale, and overview manifests
reports/figures/                               12 result figures plus the study overview
reports/pdf/                                   English paper and Chinese competition material
records/final_evidence_freeze_20260715-1612.md Canonical result record
MANIFEST.csv                                   Byte counts and SHA-256 hashes

Reproduce From GitHub

git clone https://github.com/Ireliya/auto-city-research.git
cd auto-city-research
conda env create -f environment.yml
conda activate city
python scripts/reproduce_core.py --profile final

The command downloads the dataset revision pinned by the GitHub repository, verifies MANIFEST.csv, recomputes both population-resolution analyses and the fixed consensus, checks the four exact candidate keys, and regenerates Figures 11–12. Raw satellite imagery is not required for this deterministic reviewer route.

Privacy And Redistribution Boundary

Included artifacts are derived grids, aggregated tables, selected derived geospatial layers, figures, reports, source/result manifests, and evidence ledgers.

This repository excludes:

• raw xBD satellite imagery and the full xBD archive;
• raw WorldPop GeoTIFFs;
• bulk OSM extracts, tile caches, and raw ohsome response caches;
• individual NFIP claims and person/household assistance records;
• credentials, private paths, and temporary files.

Upstream Terms

Because these artifacts combine multiple sources, the dataset card uses license: other; no blanket license overrides upstream terms.

Source	Public terms and treatment here
xBD/xView2	Official challenge page; only derived labels, grids, and aggregates are released, not imagery
WorldPop	WorldPop licensing FAQ, CC BY 4.0 attribution retained
OpenStreetMap/ohsome	ODbL and attribution; derived OSM files remain identifiable and maps carry attribution
CDC/ATSDR SVI	Official SVI documentation; only audit fields and aggregate metrics are used
OpenFEMA NFIP and RI-IHP	NFIP Claims v2 and RI-IHP v2; only non-person-level aggregates are released
Census TIGER/Line	Official technical source; Census attribution retained

See reports/data_access_license_notes.md and data/manifests/auxiliary_data_sources.csv for the detailed source ledger.

Citation

Auto-City-Research. Damage Is Not Need: Auditing Post-Disaster Priority Disagreement with Multi-Source Urban Evidence. Urban Cup 2026 Competition 2.