Floodwater Dataset

A dataset of oblique UAV video from flood-affected areas in Belgium with binary, frame-level water pseudo-labels and a separately packaged manual evaluation set. It accompanies the IEEE JSTARS article Efficient On-Board Processing of Oblique UAV Video for Rapid Flood Extent Mapping.

The code and dataset are released under the GNU General Public License v3.0. The video corpus contains SAM2-assisted pseudo-labels; the manual evaluation set contains directly annotated masks.

Dataset overview

	Train	Validation	Test	Total
Source videos	22	4	5	31
Video chunks	68	13	17	98
Labeled frames	159,848	30,017	41,159	231,024
Manual image/mask pairs	—	—	—	700
Water-pixel fraction	19.32%	11.31%	19.17%	18.29%

• Binary semantic segmentation: 0 = background/non-water, 255 = water.
• 93 chunks at 1280×720 and five chunks at 1920×1080, all at 25 fps.
• 9,541 seconds of video, approximately 2 h 39 min.
• Splits are source-video-disjoint: all chunks from a source remain in one split.

The separate manual evaluation set contains 700 manually annotated image/mask pairs: 466 at 1280×720 and 234 at 1920×1080.

Download and visualize

Video dataset download

The dataset payload is distributed as one floodwater-dataset-v1.0.0.zip archive containing videos, masks, manifests, and split definitions. Documentation and utilities are maintained in this repository.

Download the Floodwater Dataset

SHARE_TOKEN="YqqrGKwZPwwBSqK"
DATASET_URL="https://cloud.ilabt.imec.be/public.php/dav/files/${SHARE_TOKEN}/?accept=zip"
curl -fL --user "${SHARE_TOKEN}:" "$DATASET_URL" -o floodwater-dataset-v1.0.0.zip
mkdir -p data
unzip -q floodwater-dataset-v1.0.0.zip -d data
mv data/floodwater-dataset-v1.0.0 data/floodwater_dataset

Manual evaluation dataset

The 700 manually annotated image/mask pairs are distributed separately as floodwater-manual-v1.0.0.zip.

Download the manual evaluation dataset

MANUAL_SHARE_TOKEN="8ZsKAa7yRtznRjY"
MANUAL_DATASET_URL="https://cloud.ilabt.imec.be/public.php/dav/files/${MANUAL_SHARE_TOKEN}/?accept=zip"
curl -fL --user "${MANUAL_SHARE_TOKEN}:" "$MANUAL_DATASET_URL" -o floodwater-manual-v1.0.0.zip
mkdir -p data
unzip -q floodwater-manual-v1.0.0.zip -d data

The extracted datasets are stored inside the repository-local data/ directory:

floodwater-dataset/
├── data/
│   ├── floodwater_dataset/      # extracted videos, masks, and manifests
│   └── floodwater_manual/       # extracted manual image/mask pairs
├── tools/
└── visualisations/

Alternative locations are supported through FLOODWATER_DATA for the video corpus, FLOODWATER_MANUAL_DATA for the manual set, or each command’s --data-root option.

Install

cd floodwater-dataset
pip install -r requirements.txt

Visualize

Video corpus:

python -m tools.visualize_dataset
python -m tools.visualize_dataset --split val
python -m tools.visualize_dataset --split val --chunk-id video_00008_chunk_001
python -m tools.visualize_dataset --split val --chunk-id video_00008_chunk_001 --frame 100

Controls: Space pauses, N/P changes chunk, A steps backward, and Q quits.

Manual evaluation set:

python -m tools.visualize_manual_dataset
python -m tools.visualize_manual_dataset --sample-id sample_000001

Controls: N or any unassigned key advances, P goes back, and Q quits.

Validate

python -m tools.validate_dataset
python -m tools.validate_dataset --check-assets
python -m tools.validate_manual_dataset

The first command checks manifests, counts, and split isolation. The second also confirms every released video and mask path. The third validates all manual image/mask pairs, formats, dimensions, IDs, and binary mask values.

Repository layout

floodwater-dataset/
├── README.md
├── DATASET_CARD.md
├── LICENSE
├── data/
│   ├── floodwater_dataset/      # downloaded separately; ignored by Git
│   └── floodwater_manual/       # downloaded separately; ignored by Git
├── tools/
│   ├── __init__.py
│   ├── config.py
│   ├── dataset_utils.py
│   ├── manual_dataset_utils.py
│   ├── validate_dataset.py
│   ├── validate_manual_dataset.py
│   ├── visualize_dataset.py
│   └── visualize_manual_dataset.py
└── visualisations/
    ├── README.md
    ├── create_showcase.py
    ├── floodwater_showcase.gif
    └── floodwater_showcase.jpg

The large media files are hosted separately and are not committed to GitHub.

Dataset layout

data/floodwater_dataset/
├── data/
│   ├── videos/<chunk_id>.mp4
│   └── masks/<chunk_id>/<frame_index>.png
├── metadata/
│   ├── chunks.csv
│   └── samples.csv
├── splits/
│   ├── train_videos.txt, val_videos.txt, test_videos.txt
│   └── train_chunks.txt, val_chunks.txt, test_chunks.txt
└── metadata.json

data/floodwater_manual/
├── images/sample_<sample_id>.jpg
└── masks/sample_<sample_id>.png

The manual image and mask directories contain 700 pairs with matching sample identifiers.

Filename naming scheme

Video chunks and their mask directories share the same descriptive, zero-padded identifier:

video_{source_video_id:05d}_chunk_{chunk_index:03d}

Example:

data/videos/video_00008_chunk_001.mp4
data/masks/video_00008_chunk_001/00084.png

Token	Example	Meaning
source_video_id	00008	Stable source-video index
chunk_index	001	Chunk number within that source video
mask filename	00084.png	Zero-based frame index within the chunk

Resolution, duration, frame count, split, and label provenance are recorded in metadata/chunks.csv and metadata/samples.csv.

Manual pairs use matching identifiers such as images/sample_000001.jpg and masks/sample_000001.png.

Metadata format

metadata/chunks.csv contains one row per video chunk, including the source-video identity, original filename, split, video and mask paths, frame and mask counts, and video digest.

metadata/samples.csv contains one row per released labeled frame, including its stable sample ID, chunk, source video, split, zero-based frame index, chunk-relative time, video path, mask path, and label type. Column-name-based access is recommended because column order is not part of the interface.

Python use

from tools.dataset_utils import decode, iter_samples

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

Manual evaluation samples:

from tools.manual_dataset_utils import decode_manual, iter_manual_samples

sample = next(iter_manual_samples())
image_bgr, mask = decode_manual(sample)
water = mask == 255

Citation

Publications using the Floodwater Dataset should cite the following work:

@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}
}