The Floodwater Dataset

Mapping visible water in aerial video is one step towards understanding the extent of a flood. For UAVs surveying affected areas, processing these images on board can help turn a continuous video stream into useful information while data are still being collected.

The Floodwater Dataset brings together oblique UAV video of flood-affected areas in Belgium to support research on water segmentation and temporally adaptive inference. It contains 231,024 labelled frames from 31 source videos, alongside a separate evaluation set of 700 manually annotated image/mask pairs.

The dataset accompanies our article, Efficient On-Board Processing of Oblique UAV Video for Rapid Flood Extent Mapping, published in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

For the motivation behind the dataset, its relevance to aerial flood surveys, and an example of efficient onboard processing, see our introductory blog post.

Downloads & Resources

• Video dataset: Download videos, pseudo-labels, manifests, and split definitions. Video and masks occupy approximately 5.30 GB before archive overhead.
• Manual evaluation dataset: Download 700 manually annotated image/mask pairs.
• GitHub repository: download instructions, Python utilities, visualisation tools, and validation scripts.
• Research paper.

Data Collection & Annotation

The video corpus captures flood-affected areas in Belgium from oblique UAV viewpoints. Aerial video was provided by the Flanders Environment Agency (VMM) and Flanders Hydraulics Research.

The 31 source videos are divided into 98 chunks, totalling 9,541 seconds (approximately 2 hours and 39 minutes). All chunks run at 25 frames per second: 93 have a resolution of 1280 × 720 pixels and five have a resolution of 1920 × 1080 pixels.

Video masks were produced using a SAM2-assisted annotation workflow. An annotator supplied positive and negative water prompts, SAM2 propagated masks through the video, and visible tracking failures were corrected with additional prompts and re-propagation. These masks are pseudo-labels, rather than independent manual annotations of every frame.

The separate manual evaluation set provides 700 directly annotated image/mask pairs: 466 at 1280 × 720 pixels and 234 at 1920 × 1080 pixels. It is distributed in its own archive and has no train, validation, or test subdivisions.

Dataset Summary

The video corpus is split at the source-video level: every chunk from a given source belongs to the same split. This keeps neighbouring chunks from the same source together.

Working with the Dataset

The task is binary semantic segmentation of visible water. PNG masks encode background/non-water as 0 and water as 255, with no ignore-label value. The labels do not distinguish floodwater from permanent water. Muddy water, vegetation, wet soil, and reflections can make the water boundary ambiguous.

The manual archive contains JPEG images and binary PNG masks with matching sample identifiers. After downloading the data and installing the repository’s requirements, individual samples can be loaded with the provided utilities:

from tools.dataset_utils import decode, iter_samples

sample = next(iter_samples(split="train"))
frame_bgr, mask = decode(sample)
water = mask == 255

For setup instructions, interactive visualisation, and checks of the downloaded assets, see the repository guide.

Citation & Licence

The code and dataset are released under the GNU General Public License v3.0. Publications using the dataset should cite:

@ARTICLE{11673226,
  author={Sharma, Vishisht and Leroux, Sam and Landuyt, Lisa and Witvrouwen, Nick and Simoens, Pieter},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  title={Efficient On-Board Processing of Oblique UAV Video for Rapid Flood Extent Mapping},
  year={2026},
  pages={1-17},
  doi={10.1109/JSTARS.2026.3729310},
  ISSN={2151-1535}
}