This dataset follows the TACO specification.

WorldFloods: A Global Dataset for Operational Flood Extent Segmentation

Name: worldfloods
Creator: tacofoundation
License: https://choosealicense.com/licenses/cc-by-nc-4.0/

Database

WorldFloods is a public dataset containing pairs of Sentinel-2 multispectral images (Level-1C) and corresponding flood segmentation masks. As of its latest release, it comprises 509 flood events worldwide, requiring approximately 300 GB of storage if fully downloaded. The primary goal is to facilitate automatic flood mapping from optical satellite data by providing high-quality training and validation samples.

Each image-mask pair captures a satellite scene (with all 13 Sentinel-2 bands) and the corresponding binary flood masks, refined or curated from authoritative flood extent products by services such as Copernicus EMS and UNOSAT. Flood events range from 2016 to 2023, encompassing diverse hydrological conditions and global regions.

Sensors Used

Sentinel-2 MSI (Multispectral Instrument):
- 13 spectral bands spanning from visible (VNIR) to shortwave infrared (SWIR).
- Spatial resolutions of 10 m, 20 m, or 60 m, which are harmonized to 10 m in the dataset for consistent pixel-size alignment.
- Scenes are Level-1C reflectance products, further processed with custom or standardized atmospheric corrections (e.g., Sen2Cor) to ensure accurate radiometric signatures.

Original dataset

Version 1.0.0
- Released with the article: Towards global flood mapping onboard low cost satellites with machine learning (Mateo-García et al., 2021).
- Covers events from 2016 to 2019, generated semi-automatically, with only validation and test subsets manually curated.
- Offers flood mask products with fewer curated maps than v2.
- Total download size: up to ~300 GB if all subsets (train, val, test) are included.
- Google Drive: A subset of data and pretrained models are shared in a public Drive folder.
- Google Cloud Storage: The GCS bucket is: gs://ml4cc_data_lake/2_PROD/2_Mart/worldfloods_v1_0/(Requires a GCP requester pays project).
Version 2.0.0
- Latest release accompanying the article:
  Global flood extent segmentation in optical satellite images (Portalés-Julià et al., 2023).
- Flood events from 2016 to 2023, including manually curated flood masks based largely on Copernicus EMS products.
- Each sample has two binary reference channels: (1) clear/cloud and (2) land/water. This allows cloud-aware flood segmentation.
- Total download size: ~76 GB.
- Hugging Face: The dataset is hosted under isp-uv-es/WorldFloodsv2.
- Zenodo: Manually curated flood masks, metadata, and additional products (including Pakistan 2023 flood event map) are available at Zenodo, DOI: 10.5281/zenodo.8153514.

Taco dataset

To facilitate streamlined data access and usage, the WorldFloods dataset has been organized into TACO “tortillas.” Each tortilla file contains three samples:

S2: The multi-band Sentinel-2 image (Level-1C or Level-2A), harmonized to 10 m per pixel.
PERMANENTWATERJRC: A reference mask derived from the JRC (Joint Research Centre) permanent water layer.
GT: The curated flood extent mask, distinguishing water vs. land, as labeled by Copernicus EMS or other authoritative sources.

All three samples are co-registered over the same bounding box, enabling direct pixel-wise comparisons and model training. Each tortilla also includes STAC-like metadata describing the coordinate reference system (CRS), affine transform, time stamps, and other key descriptors.

🔄 Reproducible Example

Load this dataset using the tacoreader library.

import tacoreader
import rasterio as rio
import matplotlib.pyplot as plt

# Load the dataset
dataset = tacoreader.load("tacofoundation:worldfloods")

# Read a sample row
idx = 273
row = dataset.read(idx)
row_id = dataset.iloc[idx]["tortilla:id"]

# Retrieve the data
s2_path, per_path, gt_path  = row.read(0), row.read(1), row.read(2)
with rio.open(s2_path) as src_s2, rio.open(per_path) as src_per, rio.open(gt_path) as src_gt:
    window_s2 = rio.windows.Window(10300, 2900, 512, 512)
    window_per = rio.windows.Window(10300, 2900, 512, 512)
    window_gt = rio.windows.Window(10300, 2900, 512, 512)
    
    s2 = src_s2.read([2, 3, 4], window=window_s2)
    per = src_per.read(1, window=window_per)
    gt = src_gt.read(2, window=window_gt)

# Display
fig, axs = plt.subplots(1, 3, figsize=(15, 5))
axs[0].imshow(s2.transpose(1, 2, 0) / 2000)
axs[0].set_title('Sentinel-2')
axs[0].axis('off')
axs[1].imshow(gt, cmap='viridis') 
axs[1].set_title('Ground Truth')
axs[1].axis('off')
axs[2].imshow(per, cmap='coolwarm') 
axs[2].set_title('Permanent Water')
axs[2].axis('off')

plt.tight_layout()
plt.show()

🛰️ Sensor information

The sensor related to the dataset: sentinel2msi.

🎯 Task

The task associated with this dataset: semantic-segmentatio

💬 Discussion

Insights or clarifications about the dataset: https://huggingface.co/datasets/tacofoundation/worldfloods/discussions

🔀 Split Strategy

Train, test and validation.

📚 Scientific Publications

Publications that reference or describe the dataset.

Publication 01

DOI: 10.1038/s41598-023-47595-7
Summary:
This paper presents a novel machine learning framework for operational flood mapping from optical satellite images, addressing issues like cloud coverage and accuracy in flood detection. The model was trained on an extended version of the WorldFloods dataset and validated on globally distributed flood events. It shows superior performance to current operational methods, especially in areas covered by thin clouds. The paper also demonstrates the applicability of this model to map significant recent flood events, such as those in Pakistan and Australia.
BibTeX Citation:

@article{portales-julia_global_2023,
  title = {Global flood extent segmentation in optical satellite images},
  volume = {13},
  issn = {2045-2322},
  doi = {10.1038/s41598-023-47595-7},
  number = {1},
  urldate = {2023-11-30},
  journal = {Scientific Reports},
  author = {Portalés-Julià, Enrique and Mateo-García, Gonzalo and Purcell, Cormac and Gómez-Chova, Luis},
  month = nov,
  year = {2023},
  pages = {20316},
}

Publication 02

DOI: 10.1038/s41598-021-86650-z
Summary:
This paper describes the use of machine learning for global flood mapping on low-cost satellites, focusing on the WorldFloods dataset from 2016 to 2019. The dataset was created semi-automatically with a few manually curated flood events. The paper demonstrates the potential of using this dataset for improving flood mapping accuracy using Sentinel-2 and Landsat imagery.
BibTeX Citation:

@article{mateo-garcia_towards_2021,
  title = {Towards global flood mapping onboard low cost satellites with machine learning},
  volume = {11},
  issn = {2045-2322},
  doi = {10.1038/s41598-021-86650-z},
  number = {1},
  urldate = {2021-04-01},
  journal = {Scientific Reports},
  author = {Mateo-García, Gonzalo and Veitch-Michaelis, Joshua and Smith, Lewis and Oprea, Silviu Vlad and Schumann, Guy and Gal, Yarin and Baydin, Atılım Güneş and Backes, Dietmar},
  month = mar,
  year = {2021},
  pages = {7249},
}

🤝 Data providers

Organizations or individuals responsible for the dataset.

Name	Role	URL
Universitat de València - Image & Signal Processing Group	producer	https://isp.uv.es/
European Space Agency (ESA)	producer	https://earth.esa.int/eogateway/missions/sentinel-2
Copernicus Emergency Management Service (EMS)	contributor	https://emergency.copernicus.eu/
United Nations Operational Satellite Applications Team (UNOSAT)	contributor	https://unitar.org/maps
TACO Foundation	curator	https://github.com/tacofoundations/

🧑‍🔬 Curators

Responsible for structuring the dataset in the TACO format.

Name	Organization	URL
Julio Contreras	Image & Signal Processing	https://juliocontrerash.github.io/

Datasets:

tacofoundation
/

worldfloods