LandsatQuake: A Large-Scale Dataset For Practical Landslide Detection

Paper: LandsatQuake (ICLR 2025 Workshop on Machine Learning for Remote Sensing)
Authors: Vihaan Akshaay Rajendiran, Amanda Roeliza Hunt, Gen Li, Lei Li

Identifying landslides from remote imagery is critical for rapid responses after landslide hazards and for assessing their environmental impacts. LandsatQuake is a benchmark dataset composed of 31 landslide inventories from 21 earthquake-prone regions across the world, covering a total area of 5.56 × 10⁷ km² and spanning the last 40 years. This dataset emphasises practicality by using satellite images acquired by Landsat — the only satellite system that has recorded Earth's land surface for over 40 years.

⚠️ Important: Do NOT use load_dataset()

HuggingFace's auto-parquet conversion does not work correctly for this dataset. It drops all bounding-box annotations and replaces filenames with hashes. Instead, clone the repository directly:

# Install git-lfs first: https://git-lfs.com
git lfs install
git clone https://huggingface.co/datasets/Vihaanakshaay/LandsatQuake
cd LandsatQuake
git lfs pull

Dataset Structure

LandsatQuake/
├── LandsatQuake/                        # Object detection (Pascal VOC)
│   ├── train/                           # 834 images + 834 XMLs
│   ├── test/                            # 105 images + 105 XMLs
│   └── val/                             # 103 images + 103 XMLs
├── LandsatQuake_InstanceSegmentation/   # Instance segmentation (YOLO)
│   ├── image_patches/                   # Image files
│   └── labels/                          # YOLO polygon TXT files
└── LandsatQuake_Polygons.zip            # Raw polygon annotations

Format Details

Images

• Size: 224 × 224 pixels, RGB, JPEG
• Source: Landsat satellite imagery
• Naming convention: {year} {location} {pre/post}-earthquake_patch_{x}_{y}.jpg

Object Detection Labels (Pascal VOC XML)

Each image has a paired .xml file in Pascal VOC format with bounding boxes:

<annotation>
  <size><width>224</width><height>224</height><depth>3</depth></size>
  <object>
    <name>landslides</name>
    <bndbox><xmin>141</xmin><ymin>201</ymin><xmax>150</xmax><ymax>210</ymax></bndbox>
  </object>
  <!-- ... more objects ... -->
</annotation>

Instance Segmentation Labels (YOLO TXT)

Each .txt file contains normalised polygon coordinates:

0 x1 y1 x2 y2 x3 y3 ...

Where 0 = landslides class and coordinates are normalised to [0, 1].

Statistics

Split	Images	Landslide Boxes	Mean Boxes/Image
Train	834	5,626	6.7
Test	105	508	4.8
Val	103	830	8.1
Total	1,042	6,964	6.7

Events Covered (16 total)

Event	Images
1987 Sichuan pre-earthquake	508
2015 Gorkha earthquake	205
1999 Chamoli earthquake	84
2011 Sikkim earthquake	64
2016 Arun rainstorm	34
1991 Limon earthquake	30
2010 Haiti	28
Others (9 events)	89

Quick Start (Python)

import os
import xml.etree.ElementTree as ET
from PIL import Image

def parse_voc_xml(xml_path):
    """Parse Pascal VOC XML and return list of (class, xmin, ymin, xmax, ymax)."""
    tree = ET.parse(xml_path)
    boxes = []
    for obj in tree.getroot().findall("object"):
        cls = obj.find("name").text
        bb = obj.find("bndbox")
        boxes.append((
            cls,
            int(bb.find("xmin").text), int(bb.find("ymin").text),
            int(bb.find("xmax").text), int(bb.find("ymax").text),
        ))
    return boxes

# Example: load a training image with its annotations
img = Image.open("LandsatQuake/train/1987 Sichuan pre-earthquake_patch_2464_11872.jpg")
boxes = parse_voc_xml("LandsatQuake/train/1987 Sichuan pre-earthquake_patch_2464_11872.xml")
print(f"Image size: {img.size}, Landslide boxes: {len(boxes)}")

Citation

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

@inproceedings{rajendiran2025landsatquake,
  title={LandsatQuake: A Large-Scale Dataset For Practical Landslide Detection},
  author={Rajendiran, Vihaan Akshaay and Hunt, Amanda Roeliza and Li, Gen and Li, Lei},
  booktitle={3rd ICLR Workshop on Machine Learning for Remote Sensing},
  year={2025},
  url={https://iclr.cc/virtual/2025/36761}
}