Auto-City-Research

Urban Cup 2026 · Competition 2

Damage Is Not Need

When an urban AI ranks recovery only by satellite-observed building damage, which places disappear from view?

A reproducible audit of ranking disagreement across four disasters. It does not estimate true unmet need or prescribe dispatch.

Code Data Paper
Harvey Mexico Palu Santa Rosa
4disaster events
99,629labeled buildings
1,448reference cells
4 → 0robust → temporal

01 · Research question

Visible damage is evidence. It is not the whole decision.

If a post-disaster AI uses only remote-sensing damage to rank recovery priority, does it systematically omit places selected when population exposure, road access, critical services, and urban form are also inspected?
What we observe Ranking disagreement under transparent scenarios
What we do not observe True unmet need or a correct rescue allocation

02 · Event atlas

One audit, four distinct urban contexts

Select an event to inspect its own 500 m footprint, damage pattern, and evidence state.

Hurricane Harvey xBD footprint with 500 metre damage cells
Hurricane Harvey · independently rebuilt 500 m analysis cells
Flooding

Hurricane Harvey

not assessable

29.748° N · 95.580° W

Buildings
23,014
500 m cells
612
Percentile diagnostic
49
Exact top-20%
52
Cross-definition robust
0
Temporal support
0

Provisional disagreement remains visible, but no cell passes every fixed cross-definition gate.

03 · Study overview

Geography, scale, and evidence in one frame

The same graphical overview appears in the English paper and Chinese competition report.

Study overview showing four disaster locations, the same Mexico area at three grid scales, and the fixed-gate evidence funnel
Four selected event footprints, independently rebuilt analysis grids, and the fixed audit that reduced diagnostic disagreements to four non-temporal candidates and zero temporally supported candidates.

04 · Cross-scale explorer

Same place. Different analytical support.

These are independently reconstructed grids around the same real Mexico candidate, mexico-earthquake_500m_3_38. Empty mesh cells contain no xBD building labels; shaded cells entered the analysis.

Not retained

The candidate area does not meet the common two-population-product support gate at 250 m.

xBD building labels containing cell
Mexico candidate area rebuilt with a 250 metre grid
250 m reconstruction · common geographic crop

05 · Audit framework

Disagreement must survive fixed evidence gates

The framework compares rankings; it never converts a scenario score into a ground-truth need label.

Observed physical condition Damage-only rankings 4 alternative baselines
Transparent policy scenarios Multi-source priorities Population · access · services
4damage baselines
20,000weight draws
2population products
3rebuilt scales
1,448audited cells
73 / 115percentile / exact diagnostics
4cross-definition robust, all Mexico
0supported by historical OSM

NFIP, SVI, and IHP remain outside this funnel. Their results are reported as mixed, construct-specific external evidence.

06 · Evidence

The strongest result is the narrowing

Open a figure for a larger view. Every chart is generated from frozen derived tables.

Multiscale robustness results
Scale changes counts and area shares. The analysis is rebuilt at 250, 500, and 1,000 m rather than resampled from one grid.
Fixed consensus audit results
Only four cells pass every non-temporal gate. None survives the separate historical OSM support test.
External proxy divergence results
External proxies disagree by construct. Insured loss, social vulnerability, and household assistance do not define one universal target.

07 · Validity boundary

A useful audit is honest about what remains unknown.

Supported

Damage-only and multi-source rankings can diverge, and the apparent signal is sensitive to policy weights, population resolution, spatial scale, and map time.

Not established

The analysis does not identify actual unmet need, causal urban-form mechanisms, or an ethically correct rescue and recovery allocation.

Appropriate use

Use robust disagreements as locations for human review and additional local evidence, not as automatic dispatch instructions.

08 · Open reproduction

Trace every number back to data, code, config, and log.

The public repository contains the core analysis, fixed configs, evidence manifests, and one-command reproduction entry point. Raw xBD imagery is not redistributed.

Source codeGitHub repository Derived dataFixed Hugging Face revision Chinese reportCompetition narrative
core reproduction
git clone https://github.com/Ireliya/auto-city-research.git
cd auto-city-research

conda env create -f configs/environment_city.yml
conda activate city

python scripts/reproduce_core.py --profile final

Expected fixed totals: 73 percentile, 115 exact top-20%, 4 non-temporal robust, 0 temporal support.