The dataset supports an Urban Cup 2026 Competition 2 project that audits whether a disaster-response urban AI system would under-rank high-need places if it only used satellite-observed building damage.

It contains derived tables, figures, report PDFs, and evidence ledgers used by the GitHub code repository. It does not redistribute raw xBD satellite imagery.

It also excludes individual NFIP claim rows. Only privacy-safe Census-tract aggregates for the Harvey study footprint are released.

Main Contents

data/derived/xbd_core_v1/              Parsed xBD building label tables
data/derived/xbd_damage_grid_v1/       500 m damage grid
data/derived/osm_context_v1/           OSM road/facility context
data/derived/worldpop_context_v1/      WorldPop population context
data/derived/priority_mismatch_v1/     Main damage-vs-need mismatch outputs
data/derived/mismatch_drivers_v1/      Driver profiles for mismatch cells
data/derived/robustness_v1/            Threshold/scenario robustness checks
data/derived/strict_budget_v1/         Exact top-k budget checks
data/derived/osm_building_form_v1/     Independent OSM building-form robustness
data/derived/evidence_hardening_v1/     Four baselines and 20,000 weight draws
data/derived/multiscale_v1/             250/500/1000 m prepared grids and results
data/derived/harvey_external_validation_v1/  Aggregated NFIP boundary validation
reports/figures/                       PNG/PDF report figures
reports/pdf/                           PDF exports of paper and reports
records/evidence_index.csv             Claim-to-evidence ledger
MANIFEST.csv                           File size and SHA-256 manifest

Headline Verification Numbers

The downloadable tables should verify:

xBD building records: 99,629
500 m cells: 1,448
Primary stable mismatch cells: 67
Exact top-20 stable mismatch cells: 109
OSM building polygons joined for robustness: 813,352
Four-baseline top-20 mismatch ranges: Harvey 27--51; Santa Rosa 18--41
Policy-plausible weight draws: 10,000, plus 10,000 unconstrained draws
Multiscale 500 m exact top-20 mismatch cells: 109
Harvey external validation: 149 intersecting tracts and 10,134 aggregated NFIP claims

The NFIP result is mixed rather than confirmatory: need-aware rankings do not consistently outperform damage-only rankings against insured property losses. NFIP is therefore treated as a boundary test, not as ground truth for unmet rescue or recovery need.

Download From GitHub Code Repository

git clone https://github.com/Ireliya/auto-city-research.git
cd auto-city-research
python -m pip install huggingface_hub pandas
python scripts/download_data.py
python scripts/smoke_reproduce.py

License And Source Boundary

Code in the paired GitHub repository is MIT licensed. Project documentation is released under CC BY 4.0 unless otherwise noted.

The data artifacts here are derived from multiple upstream sources, including xBD/xView2, WorldPop, OpenStreetMap, OpenFEMA NFIP Claims v2, and US Census tract boundaries. Use remains subject to upstream data terms. Raw satellite imagery, source rasters, map-tile caches, and individual insurance claim rows are not redistributed here.

For details, see:

reports/data_access_license_notes.md
data/manifests/auxiliary_data_sources.csv

Downloads last month: 27

Total file size:

115 MB