The Floodwater Dataset
Aerial scenes from the Floodwater Dataset. The animated sequences illustrate water pseudo-labels over time, alongside examples from the manual evaluation set.
Aerial Observation for Disaster Management
During a flood, response teams need timely information about affected areas, damaged infrastructure, and access routes. Drones can provide aerial observations that support this assessment, including in locations that are difficult to inspect from the ground. Their use in disaster response is already practical: during the response to Cyclone Ana, UNICEF reported using drone imagery to identify damaged road sections and disrupted access.
In a pilot-operated survey, the pilot directs the aircraft while operators and response personnel can review the footage to identify areas requiring attention. This interpretation supports decisions about where to inspect next and where assistance may be needed. Automated analysis can support that process by identifying visible water throughout a video stream and presenting consistent segmentation overlays for human review.
Developing these capabilities requires data that capture both the visual complexity of flood-affected environments and the temporal characteristics of aerial observation. We introduce the Floodwater Dataset, a collection of oblique aerial video from Belgium with dense water pseudo-labels and a separate manual evaluation set.
Representing Realistic Aerial Surveys
Flood surveys involve observing and scanning an area rather than collecting only isolated images. Slow panning movements keep substantial parts of the landscape visible across successive frames, while faster movements and changes in viewpoint introduce new content. Oblique views capture water, vegetation, buildings, and infrastructure at different distances and apparent scales.
The Floodwater video sequences contain both slow-panning and faster-changing imagery, making them relevant to these observation conditions. Reflections, varying illumination, and water boundaries near vegetation or infrastructure add further segmentation challenges. Aerial video was provided by the Flanders Environment Agency (VMM) and Flanders Hydraulics Research.
The following excerpts illustrate the contrast. Both show 12 seconds at the original playback speed, with the resolution reduced for display.
Slow scan. Water, tree lines, and roads remain in nearly the same positions as the view changes gradually. Much of the scene persists across successive frames. Source: video_00001_chunk_001, 00:10–00:22.
Fast aerial sweep. The view moves quickly across woodland, water, and nearby roads, with landscape features continually entering and leaving the frame. Source: video_00031_chunk_001, 00:27–00:39.
Continuous video also preserves information that independent images cannot: how much a scene changes over time. During a slow scan, large regions may remain visually similar for many frames. This provides a setting for investigating whether a segmentation model can reuse earlier computations while responding to newly visible or changing regions.
Dataset Composition
The released video corpus contains 31 source videos, divided into 98 chunks, with 231,024 labelled frames. It spans approximately 2 hours and 39 minutes, with all chunks recorded at 25 frames per second. Of the 98 chunks, 93 have a resolution of 1280 × 720 pixels and five have a resolution of 1920 × 1080 pixels.
The training, validation, and test partitions contain 22, four, and five source videos, respectively. Every chunk from a given source remains in the same partition. A separately distributed evaluation set contains 700 manually annotated image/mask pairs, providing directly annotated examples alongside the larger pseudo-labelled video corpus.
The task is binary segmentation of visible water. Masks distinguish water from background; they do not distinguish floodwater from permanent water. Detailed split counts and download links are available on the dataset overview page.
Producing Dense Video Annotations
Annotating every frame of a long video manually is labour-intensive. The video corpus therefore uses a SAM2-assisted workflow. An annotator supplies positive and negative point prompts to identify water, and SAM2 propagates the resulting masks through the sequence. The propagated masks are inspected, and visible tracking failures are corrected through additional prompts and re-propagation.
The resulting video annotations are pseudo-labels, supported by human guidance and correction. They are distinct from the directly annotated masks in the 700-image manual evaluation set. This distinction matters when selecting data for model development and interpreting evaluation results.
From Continuous Video to Efficient Onboard Processing
The dataset supports research on water segmentation, consistency between successive predictions, and adaptive computation. Our accompanying IEEE JSTARS paper investigates the last of these through Temporal Token Reuse (TTR).
TTR compares image patches over time and reuses previously computed features for regions that remain sufficiently similar. Regions requiring an update are processed again. This addresses a characteristic of slow-panning aerial video: processing every region from scratch can repeat substantial computation even when much of the scene has changed little.
The table below summarises the paper’s Floodwater benchmark results for two backbones on the NVIDIA Jetson Orin Nano. Mean intersection over union (mIoU) measures the average overlap between predicted and reference regions across classes; higher values indicate better segmentation. Frames per second (FPS) measures processing throughput.
| Model | mIoU (%) | FPS on Orin Nano | Speedup |
|---|---|---|---|
| EfficientNet-B4 | 78.0 → 77.6 | 15 ± 1 → 25 ± 3 | 1.67× |
| ResNet-50 | 74.0 → 73.8 | 11 ± 1 → 18 ± 3 | 1.64× |
FPS values are means ± standard deviations over five inference runs; speedups are calculated from the reported means. TTR increases throughput for both models, with mIoU reductions of 0.4 and 0.2 percentage points, respectively. The paper describes the experimental protocol underlying these benchmark results.
Water segmentation is one component of a broader disaster-management workflow. Translating image-space masks into georeferenced flood extents or operational access information requires additional processing and validation. The dataset provides a basis for studying the visual analysis and computational efficiency needed within that workflow.
Access the Dataset
The video corpus and manual evaluation set are available through the Floodwater dataset page. The GitHub repository provides download instructions, visualisation tools, and validation utilities. Code and data are released under the GNU General Public License v3.0. Please cite Efficient On-Board Processing of Oblique UAV Video for Rapid Flood Extent Mapping when using the dataset in research.