abalhomaid/disaster-damage-assessment

Disaster Damage Assessment Dataset

Binary damage classification dataset (damage vs. no damage) across 4 natural disaster domains, used in:

Unsupervised Domain Adaptation for Rapid Disaster Damage Assessment

Dataset Structure

Domain	Code	Event	Images
Ecuador Earthquake	E	2016	1,724
Nepal Earthquake	N	2015	19,104
Hurricane Matthew	M	2016	333
Typhoon Ruby	R	2014	833

Total: 21,994 images across 4 domains.

Columns:

• image: PIL Image
• label: 0 = no_damage, 1 = damage
• domain: Single-letter domain code (E, N, M, R)
• domain_name: Full domain name
• filename: Original image filename

Usage

Load with datasets

from datasets import load_dataset

ds = load_dataset("abalhomaid/disaster-damage-assessment")

# Filter by domain
nepal = ds["train"].filter(lambda x: x["domain"] == "N")

Extract to directory structure (for training scripts)

To use with the reproduction code, extract to the expected directory layout:

from datasets import load_dataset
import os
from PIL import Image

ds = load_dataset("abalhomaid/disaster-damage-assessment")

domain_dirs = {"E": "ecuador_eq", "N": "nepal_eq", "M": "matthew_hurricane", "R": "ruby_typhoon"}
class_dirs = {0: "no_damage", 1: "damage"}

for row in ds["train"]:
    domain_dir = domain_dirs[row["domain"]]
    class_dir = class_dirs[row["label"]]
    out_dir = f"data/damage/{domain_dir}/images/{class_dir}"
    os.makedirs(out_dir, exist_ok=True)
    row["image"].save(f"{out_dir}/{row['filename']}")

Associated Code

See the reproduction repository for training and evaluation code.

Pre-trained model weights: abalhomaid/disaster-uda-models