# xBD Damage Inventory QA Dataset

## Overview
A question-answering dataset for building damage assessment
from post-disaster satellite imagery, built on the
[xBD / xView2](https://xview2.org) dataset.

## Dataset Statistics
- 300 scenes selected across 5 damage buckets
- 1,656 QA samples covering 8 question templates
- 100% validation pass rate
- Disaster types: Wind, Flooding, Fire, Tsunami, Volcano, Earthquake

## QA Templates
| Template | Description | Difficulty |
|----------|-------------|------------|
| XBD-Q1 | Full damage inventory report | Medium |
| XBD-Q2 | Worst quadrant by severe area | Medium |
| XBD-Q3 | Highest severity 4×4 cell | Hard |
| XBD-Q4 | Cross-scene comparison | Hard |
| XBD-Q5 | Severe polygon filter + area | Easy |
| XBD-Q6 | Event-level severity ranking | Hard |
| XBD-Q7 | Spatial dispersion analysis | Medium |
| XBD-Q8 | Full scene change summary | Hard |

## Splits
| Split | Samples |
|-------|---------|
| Train | 1,212 |
| Val   | 203 |
| Test  | 241 |

## Damage Buckets
| Bucket | Scenes |
|--------|--------|
| no-low-damage | 50 |
| minor-moderate | 50 |
| major-heavy | 70 |
| destroyed-heavy | 70 |
| mixed-severe | 60 |

## Area Computation
All footprint areas computed as:
`area_m2 = shapely_polygon.area (px²) × GSD²`
GSD sourced from metadata.gsd in post-disaster JSON.
lng_lat polygons used for visualization only.

## Validation
Every QA sample independently validated by recomputing
ground truth from raw source JSON files.
Pass rate: 100% (1,656 / 1,656)