youngsun05/SDSU_MidWest_Flood_2019

SDSU Midwest Flood Dataset 2019

This dataset provides a benchmark for flood detection in satellite imagery. It contains true-color satellite images from the 2019 Midwest USA flooding event, with each image paired with a binary mask indicating flooded areas.

Important Update

This dataset has been revised to provide segmentation-ready image-mask pairs through the Hugging Face datasets library.

Each sample now contains:

• image: RGB satellite image
• mask: binary flood mask image
• file_name: original image filename
• mask_file_name: corresponding mask filename

The previous Hugging Face viewer displayed 1,000 rows because the original Image/ and Mask/ folders were interpreted as separate image-folder classes. The corrected version contains 500 image-mask pairs.

Dataset Contents

• Images: True-color satellite imagery
• Masks: Binary masks for flooded areas
• Number of image-mask pairs: 500
• Image size: 700 × 700 pixels
• Mask values: 0 for non-flooded areas and 255 for flooded areas
• States covered: Iowa, Kansas, Montana, Nebraska, and South Dakota
  • Note: The state abbreviation used in the original filenames follows the naming convention from the original dataset files.
• Images per location: 10

Usage

You can load this dataset directly using Hugging Face Datasets:

from datasets import load_dataset
import numpy as np

dataset = load_dataset("youngsun05/SDSU_MidWest_Flood_2019")

sample = dataset["train"][0]

image = sample["image"]   # PIL RGB image
mask = sample["mask"]     # PIL binary mask image

print(image.mode, image.size)
print(mask.mode, mask.size)

mask_array = np.array(mask.convert("L"))
mask_array = (mask_array > 0).astype("uint8")

print(mask_array.shape)
print(np.unique(mask_array))

For semantic segmentation models, use the image column as the input image and the mask column as the target segmentation mask.

Paper

If you use this dataset in your research or project, please cite the following paper:

A Novel Dataset for Flood Detection Robust to Seasonal Changes in Satellite Imagery Youngsun Jang, Dongyoun Kim, Chulwoo Pack, and Kwanghee Won Presented at the ACM RACS 2024 conference, Pompei, Italy, November 5–8, 2024.

arXiv Version

This paper was presented at ACM RACS 2024, but the session was not published in the official ACM proceedings. This Hugging Face repository and the arXiv version serve as the official references.

Citation

If you use this dataset, please cite:

@misc{jang2024midwestflood,
  title={A Novel Dataset for Flood Detection Robust to Seasonal Changes in Satellite Imagery},
  author={Youngsun Jang and Dongyoun Kim and Chulwoo Pack and Kwanghee Won},
  year={2024},
  note={Presented at ACM RACS 2024. Available at \url{https://huggingface.co/datasets/youngsun05/SDSU_MidWest_Flood_2019} and \url{https://arxiv.org/abs/2507.23193}},
}

License

This dataset is distributed under the CC-BY-4.0 License, allowing sharing and adaptation with appropriate credit.

Links

• Dataset GitHub Repository: https://github.com/youngsunjang/SDSU_MidWest_Flood_2019
• arXiv Paper: https://arxiv.org/abs/2507.23193