FloodPULSEO
Year: 2026
Type: dataset
Abstract: Accurate and timely flood inundation mapping from Earth observation (EO) data remains challenging because of the incomplete spatial and temporal coverage of EO observations, particularly during rapidly evolving flood events. Existing global flood datasets are predominantly designed for flood segmentation from individual optical or synthetic aperture radar (SAR) observations, which limits their applicability to flood inundation mapping and continuous flood monitoring. FloodPULSEO is an original global satellite remote sensing dataset, created to support the development of deep learning (DL) models for flood inundation prediction through multimodal fusion of dynamic hydroclimatic variables and static flood conditioning factors. Such models predict the flood extent on a given day from the precipitation and soil moisture of the preceding days and the land surface characteristics of the area, rather than by segmenting a satellite image of the flood. The dataset comprises 190 flood events, one per Copernicus Emergency Management Service (CEMS) rapid-mapping activation between April 2017 and December 2025, across six continents. These events cover 845 areas of interest (AOIs) and 1,565 mapping products, resulting in 566,669 image patches, each covering a 2.56 km × 2.56 km ground footprint. Each event includes manually annotated flood labels from CEMS, together with 30 days of precipitation and soil moisture from the Global Precipitation Measurement (GPM) and Soil Moisture Active Passive (SMAP) missions before the mapping date, and flood conditioning factors: median composites of Sentinel-1 backscatter and Sentinel-2 NDVI and NDBI describing the recent land surface, permanent water from ESA WorldCover, terrain features from MERIT Hydro, and soil texture from SoilGrids. The precipitation and soil moisture of the mapping date itself are added as two optional bands. These data are provided as co-registered multi-resolution layers. Optional train, validation and test splits are provided, exclusive by river basin, mapping product and activation, which avoids data leakage between hydrologically connected mapping products, and each flood label carries a confidence class based on the resolution of the reference image used for manual annotation. FloodPULSEO supports the development of DL models for flood inundation prediction and research on multimodal EO data fusion and on the influence of flood label quality on model performance. This deposit holds the patches as GeoTIFFs in 442 tar archives (about 649 GiB, folder patches/), the original CEMS AOI and flood-extent shapefiles of all 1,565 mapping products (folder annotations/), and the metadata tables, the train, validation and test splits (381,694, 100,223 and 84,752 patches), and the archive index and MD5 checksums (metadata/metadata.tar). The MERIT Hydro terrain layers (input_80m) are not redistributed because of the MERIT Hydro licence (CC BY-NC 4.0 / ODbL 1.0); they can be rebuilt locally with the scripts supplied with the data generation pipeline.
Source: Harvard Dataverse
DOI: 10.7910/dvn/a7x9jd
Authors: Kumar, Ravi, Surojaya, Ali, Dasgupta, Antara
Institutions: Sapienza University of Rome, RWTH Aachen University
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7225204486