ly-ai/Disaster_Recognition_RemoteSense_EN_CN_JA

xView2 Multi-Language Disaster Recognition Dataset

This dataset is derived from the xBD (xView2) Building Damage Assessment Dataset and has been reformatted for Vision-Language Model (VLM) training with multi-language support.

📊 Dataset Overview

This dataset contains satellite imagery paired with multi-language conversational annotations for disaster recognition tasks. It supports three languages: English, Chinese (中文), and Japanese (日本語).

Dataset Splits

• Training Set (tier3): 9,168 image pairs → 55,008 conversations
• Test Set: 933 image pairs → 5,598 conversations
• Total: 10,101 image pairs → 60,606 multi-language conversations

Each image pair consists of:

• Pre-disaster satellite image
• Post-disaster satellite image
• Corresponding segmentation masks
• Building damage labels
• Metadata (capture date, sun position, sensor info)

🗂️ Dataset Structure

Downloadable Files (Available on HuggingFace)

The dataset is provided as compressed archives to facilitate downloading:

xview2/
├── xview2_train.tar.gz             # Training split (8.04 GB compressed)
├── xview2_tier3.tar.gz             # Additional training data (17.79 GB compressed)
├── xview2_test.tar.gz              # Test split (2.67 GB compressed)
├── xview2_train_tier3.json         # Training metadata
├── xview2_test.json                # Test metadata
├── xview2_train_tier3_sharegpt.json # Training conversations (ShareGPT format)
├── xview2_test_sharegpt.json       # Test conversations (ShareGPT format)
├── verify_dataset.py               # Dataset integrity verification script
├── README.md                       # This file
└── samples/images/                 # Sample images for preview
    ├── guatemala-volcano_00000000_pre_disaster.png
    ├── guatemala-volcano_00000000_post_disaster.png
    ├── hurricane-florence_00000004_post_disaster.png
    └── santa-rosa-wildfire_00000000_post_disaster.png

After Extraction

Once you extract the compressed archives, the structure will be:

xview2/
├── train/                          # Training split (extracted from xview2_train.tar.gz)
│   ├── images/                     # Satellite images (pre/post disaster)
│   ├── masks/                      # Segmentation masks
│   ├── color_masks/                # Visualization masks
│   └── labels/                     # Building annotations (JSON)
├── tier3/                          # Additional training data (extracted from xview2_tier3.tar.gz)
│   ├── images/
│   ├── masks/
│   ├── color_masks/
│   └── labels/
├── test/                           # Test split (extracted from xview2_test.tar.gz)
│   ├── images/
│   ├── masks/
│   ├── color_masks/
│   └── labels/
└── ... (metadata and conversation files)

🌍 Disaster Types

The dataset covers 6 types of natural disasters:

Type	English	中文	日本語	Examples
volcano	Volcano	火山	火山	Guatemala volcano
flooding	Flooding	洪水	洪水	Hurricane Florence, Hurricane Harvey
wind	Wind damage	风灾	風災	Hurricane Matthew, Hurricane Michael
earthquake	Earthquake	地震	地震	Mexico earthquake
tsunami	Tsunami	海啸	津波	Palu tsunami
fire	Fire	火灾	火災	Santa Rosa wildfire, SoCal fire

🖼️ Sample Images

The samples/images/ directory contains example images for preview:

• Guatemala Volcano (Pre-disaster): guatemala-volcano_00000000_pre_disaster.png
• Guatemala Volcano (Post-disaster): guatemala-volcano_00000000_post_disaster.png
• Hurricane Florence (Post-disaster): hurricane-florence_00000004_post_disaster.png
• Santa Rosa Wildfire (Post-disaster): santa-rosa-wildfire_00000000_post_disaster.png

💬 Conversation Format

The dataset uses ShareGPT format with two-turn conversations:

Real Example: Guatemala Volcano (Post-disaster - Chinese)

Image: train/images/guatemala-volcano_00000000_post_disaster.png (see sample above)

{
  "id": "xview2_train_000000_post_zh",
  "images": ["data/xview2/train/images/guatemala-volcano_00000000_post_disaster.png"],
  "messages": [
    {
      "from": "human",
      "value": "<image>\n这张图片中发生了灾害吗？"
    },
    {
      "from": "gpt",
      "value": "是的，这张图片中发生了灾害。这是灾后图像。"
    },
    {
      "from": "human",
      "value": "这张图片中发生了什么类型的灾害？"
    },
    {
      "from": "gpt",
      "value": "这张图片中发生了火山灾害。"
    }
  ]
}

Real Example: Guatemala Volcano (Pre-disaster - English)

Image: train/images/guatemala-volcano_00000000_pre_disaster.png (see sample above)

{
  "id": "xview2_train_000000_pre_en",
  "images": ["data/xview2/train/images/guatemala-volcano_00000000_pre_disaster.png"],
  "messages": [
    {
      "from": "human",
      "value": "<image>\nHas a disaster occurred in this image?"
    },
    {
      "from": "gpt",
      "value": "No, no disaster has occurred in this image. This is a pre-disaster image."
    },
    {
      "from": "human",
      "value": "What type of disaster occurred in this image?"
    },
    {
      "from": "gpt",
      "value": "No disaster has occurred. This image was taken before the disaster."
    }
  ]
}

Real Example: Guatemala Volcano (Post-disaster - Japanese)

{
  "id": "xview2_train_000000_post_ja",
  "images": ["data/xview2/train/images/guatemala-volcano_00000000_post_disaster.png"],
  "messages": [
    {
      "from": "human",
      "value": "<image>\nこの画像では災害が発生していますか？"
    },
    {
      "from": "gpt",
      "value": "はい、この画像では災害が発生しています。これは災害後の画像です。"
    },
    {
      "from": "human",
      "value": "この画像ではどのような種類の災害が発生しましたか？"
    },
    {
      "from": "gpt",
      "value": "この画像では火山災害が発生しました。"
    }
  ]
}

📚 Original Dataset Citation

This dataset is based on the xBD (xView2) Dataset:

@InProceedings{Gupta_2019_CVPR_Workshops,
  author = {Gupta, Ritwik and Goodman, Bryce and Patel, Nirav and Hosfelt, Ricky and Sajeev, Sandra and Heim, Eric and Doshi, Jigar and Lucas, Keane and Choset, Howie and Gaston, Matthew},
  title = {Creating xBD: A Dataset for Assessing Building Damage from Satellite Imagery},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month = {June},
  year = {2019},
  pages = {10-17}
}

Paper Abstract: xBD is a large-scale dataset for the advancement of change detection and building damage assessment for humanitarian assistance and disaster recovery research. The dataset provides pre- and post-event multi-band satellite imagery from a variety of disaster events with building polygons, classification labels for damage types, ordinal labels of damage level, and corresponding satellite metadata. xBD contains ~700,000 building annotations across over 5,000 km² of imagery from 15 countries.

🔗 Data Source

• Original Dataset: https://xview2.org/dataset
• License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)

📋 License

This derivative dataset follows the original license:

CC BY-NC-SA 4.0 - You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made
• NonCommercial — You may not use the material for commercial purposes
• ShareAlike — If you remix, transform, or build upon the material, you must distribute your contributions under the same license

🎯 Use Cases

This dataset is suitable for:

1. Vision-Language Model Training: Multi-modal models that understand disaster imagery
2. Multi-language AI Systems: Models that can communicate about disasters in multiple languages
3. Disaster Assessment: Automated systems for rapid disaster type identification
4. Change Detection: Pre/post disaster image comparison
5. Humanitarian AI: Applications for disaster response and recovery

📦 How to Use

Step 1: Download and Extract

# Download from HuggingFace, then extract
tar -xzf xview2_train.tar.gz
tar -xzf xview2_tier3.tar.gz
tar -xzf xview2_test.tar.gz

Step 2: Verify Dataset Integrity

python verify_dataset.py

Expected Output:

Verifying dataset integrity... 
✅ Dataset is ready

For detailed verification report:

python verify_dataset.py --verbose

Step 3: Load and Use

import json
from PIL import Image

# Load conversations
with open('xview2_train_tier3_sharegpt.json', 'r', encoding='utf-8') as f:
    conversations = json.load(f)

# Get first conversation
conv = conversations[0]

# Load image
image = Image.open(conv['images'][0])

# Access conversation
print(conv['messages'][0]['value'])  # Question 1
print(conv['messages'][1]['value'])  # Answer 1

📚 Citation

If you use this dataset in your research, please cite:

@misc{xview2_multi_language_2024,
  title = {xView2 Multi-Language Disaster Recognition Dataset},
  author = {WayBob},
  year = {2025},
  howpublished = {\url{https://huggingface.co/datasets/WayBob/Disaster_Recognition_RemoteSense_EN_CN_JA}},
  note = {A multi-language Vision-Language Model training dataset derived from xBD (xView2)}
}

And please also cite the original xBD dataset:

@InProceedings{Gupta_2019_CVPR_Workshops,
  author = {Gupta, Ritwik and Goodman, Bryce and Patel, Nirav and Hosfelt, Ricky and Sajeev, Sandra and Heim, Eric and Doshi, Jigar and Lucas, Keane and Choset, Howie and Gaston, Matthew},
  title = {Creating xBD: A Dataset for Assessing Building Damage from Satellite Imagery},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month = {June},
  year = {2019},
  pages = {10-17}
}

📋 License

CC BY-NC-SA 4.0 - Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International

Original dataset: https://xview2.org/dataset