ABCD dataset

Name: ABCD (AIST Building Change Detection) Dataset
Purpose: To build and evaluate post‑tsunami building damage detection systems, specifically to identify whether buildings have been washed away.
Content: Contains 8,506 fixed‑scale and 8,444 resized building image pairs; each pair includes pre‑ and post‑tsunami aerial patches centered on a target building.
Image Source: RGB aerial imagery from Japan's Tohoku region captured before and after the 2011 Great East Japan Earthquake.
Image Resolution: Pre‑earthquake images at 40 cm resolution; post‑earthquake images originally at 12 cm (resampled to 40 cm).
Image Format: Fixed‑scale images are 160 × 160 pixels; resized images are uniformly cropped to 128 × 128 pixels based on building size.
Label Source: Labels derived from on‑site surveys conducted by the Ministry of Land, Infrastructure, Transport and Tourism (MLIT), which assessed over 220,000 buildings in the affected area.

Directory Layout:
- fixed-scale/: Fixed‑scale image pairs.
  - patch-pairs/: Stores fixed‑scale pairs in .tif format.
  - 5fold-list/: CSV files for 5‑fold cross‑validation.
- resized/: Resized image pairs.
  - patch-pairs/: Stores resized pairs in .tif format.
  - 5fold-list/: CSV files for 5‑fold cross‑validation.
File Format: Each .tif file contains six channels; the first three are pre‑tsunami RGB, the last three are post‑tsunami RGB.
Label Format: In the CSV, the first column is the file name and the second column is the class label ("1" for "washed away", "0" for "survived").

Contact Person: Aito Fujita
Institution: National Institute of Advanced Industrial Science and Technology (AIST), Japan
Email: fujita.713[at]aist.go.jp