Earth Observation · isp-uv-es
WorldFloodsv2
A global remote sensing dataset pairing optical satellite imagery with flood extent segmentation masks.
What WorldFloodsv2 is
WorldFloodsv2 is an Earth observation dataset developed by researchers at the Image and Signal Processing (ISP) group at the Universitat de València (isp-uv-es). Created to advance automated water delineation in earth-science and climate applications, the collection provides optical satellite imagery alongside paired ground-truth masks covering flood events worldwide.
The dataset addresses the challenge of accurately mapping flood extents from space by offering standardised data for training and evaluating machine learning segmentation models across diverse global landscapes.
How it was made
WorldFloodsv2 contains 509 pairs of Sentinel-2 optical satellite scenes and corresponding flood segmentation masks, with coverage linked to Landsat and Sentinel observations across hundreds of global flood events. The data are structured into pre-partitioned training, validation, and test sets.
The collection is formatted as an image folder structure and is designed to integrate with the ml4floods library for dataset inspection, plotting, and loading into PyTorch pipelines.
Key facts
- Imagery sources
- Sentinel-2 and Landsat
- Dataset size
- 509 image and mask pairs
- Storage requirement
- Approximately 76 GB
- Data format
- ImageFolder
- Modalities
- Geospatial imagery and segmentation masks
- Data splits
- Train, validation, and test sets
What you can use it for
- Flood extent segmentation
- Training computer vision and deep learning models to identify and segment surface water boundaries from optical satellite observations.
- Algorithm benchmarking
- Evaluating the performance of remote sensing segmentation algorithms using fixed training, validation, and test splits across diverse flood events.
- Climate and hazard monitoring
- Supporting earth-science research into spatial flood distribution and environmental impacts using verified satellite-derived flood masks.
Explore the data
The schema, real rows — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from isp-uv-es/WorldFloodsv2 at the moment you asked. Nothing is cached or stored — every row above came from that request.
Strengths and limits
Good at
- Provides verified global flood segmentation masks paired directly with high-resolution optical satellite imagery.
- Includes designated train, validation, and test splits to support standardised model development and comparison.
- Supported by ml4floods utilities for visualisation and PyTorch data loading workflows.
Watch out for
- Focuses primarily on optical imagery such as Sentinel-2, which is susceptible to cloud cover obscuring flood extents.
- Requires substantial local disk storage, occupying approximately 76 GB.
Questions
Can I look inside WorldFloodsv2 without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser. Nothing is downloaded to your machine.
What licence is WorldFloodsv2 under?
Its authors declare cc-by-nc-4.0. Read the terms on the original page before reusing the data.
What satellite data does the dataset use?
The dataset contains 509 pairs of Sentinel-2 images and flood segmentation masks, curated alongside Landsat and Sentinel observations.
How much disk storage is required?
The full dataset requires approximately 76 GB of local hard-disk storage space.
Are pre-defined training and evaluation splits provided?
Yes, the 509 image and mask pairs are divided into separate training, validation, and test sets.
Can WorldFloodsv2 be loaded directly into PyTorch?
Yes, the ml4floods library provides tutorials and data loaders to load WorldFloodsv2 directly into PyTorch.
How to cite it
Portalés-Julià, E., Mateo-García, G., Purcell, C. & Gómez-Chova, L. Global flood extent segmentation in optical satellite images. Scientific Reports 13, 20316 (2023). DOI: 10.1038/s41598-023-47595-7.
Written from WorldFloodsv2's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.