Wagner et al. (2025) The fully-automatic Sentinel-1 Global Flood Monitoring service: Scientific challenges and future directions

Identification

• Journal: Remote Sensing of Environment
• Year: 2025
• Date: 2025-11-04
• Authors: Wolfgang Wagner, Bernhard Bauer-Marschallinger, Florian Roth, Tobias Raiger-Stachl, Christoph Reimer, Niall McCormick, Patrick Matgen, Marco Chini, Yu Li, Sandro Martinis, Marc Wieland, Franziska Kraft, Davide Festa, Muhammed Hassaan, Mark Edwin Tupas, Jie Zhao, Michaela Seewald, Michael Riffler, Luca Molini, Richard Kidd, Christian Briese, Peter Salamon
• DOI: 10.1016/j.rse.2025.115108

Research Groups

• Technische Universität Wien, Department of Geodesy and Geoinformation, Vienna, Austria
• EODC Earth Observation Data Centre, Vienna, Austria
• European Commission, Joint Research Centre, Ispra, Italy
• Luxembourg Institute of Science and Technology, Esch-Belval Belvaux Sanem, Luxembourg
• Deutsches Zentrum für Luft- und Raumfahrt, Weßling, Germany
• University of the Philippines, Department of Geodetic Engineering, Quezon City, Philippines
• Technical University of Munich, Munich, Germany
• GeoVille Information Systems and Data Processing, Innsbruck, Austria
• Centro Internazionale in Monitoraggio Ambientale, Savona, Italy

Short Summary

The Global Flood Monitoring (GFM) service, launched in 2021 as part of the Copernicus Emergency Management Service (CEMS), provides fully-automatic, near-real-time global flood maps using Sentinel-1 SAR imagery. This paper presents a comprehensive analysis of GFM's scientific achievements and challenges, demonstrating its rapid delivery and good accuracy for larger-scale floods while identifying limitations in coverage and detection in specific environments.

Objective

• To present the first comprehensive analysis of the Global Flood Monitoring (GFM) service's scientific achievements and challenges during its initial years of operation.
• To discuss how the service was set up to benefit from novel scientific algorithms and Big Data solutions in a cloud platform environment.
• To outline directions for scientific research and system development to enhance the GFM service.

Study Configuration

• Spatial Scale: Global (all continental land areas except Antarctica), 20 meter pixel spacing.
• Temporal Scale: Sentinel-1 observation period (2015 to present) for the archive; near-real-time processing for new acquisitions (typically within 5 hours of image acquisition). Average revisit time of 1 to 12 days depending on region and satellite constellation.

Methodology and Data

• Models used:
  • Ensemble approach combining three complementary flood-mapping algorithms:
    • Single-image classifier (DLR): Hierarchical tile-based thresholding, fuzzy logic post-classification, Height Above Nearest Drainage (HAND) index.
    • Dual-image classifier (LIST): Hierarchical split-based approach, statistical backscatter modeling, region growing, change detection (compares with last image from same orbit).
    • Time-series classifier (TU Wien): Bayesian inference based on a harmonic backscatter model (compares with simulated reference image).
  • Majority voting mechanism for binary flood map.
  • Arithmetic mean for flood likelihood layer.
• Data sources:
  • Sentinel-1 Synthetic Aperture Radar (SAR) Interferometric Wide (IW) swath mode, Ground Range Detected (GRDH) images, VV polarisation.
  • Global Sentinel-1 backscatter datacube (20 meter resolution, Equi7Grid).
  • Ancillary datasets: Monthly reference water maps (derived from Sentinel-1 median backscatter), exclusion mask (derived from Sentinel-1 datacube, Global Forest Change, World Settlement Footprint, Copernicus DEM Water Body Mask, HAND index), Copernicus Water Body Mask, Global Human Settlement Layer (GHS-POP) for population, Copernicus Global Land Cover dataset.
  • Global Disaster Alert and Coordination System (GDACS) for flood event validation.
  • Human-interpreted Sentinel-1 images for accuracy assessment.

Main Results

• The GFM service typically delivers flood maps within 5 hours of image acquisition, with best-case scenarios under 90 minutes from sensing to dissemination.
• A complete global Sentinel-1 flood archive from 2015 to present has been created and is continuously updated.
• The GFM exclusion mask covers 69.9% of the global land surface, primarily due to topographic distortions (54.2%), non-sensitive areas (32.0% for forests and urban areas), and low-backscatter areas (11.8% for arid regions).
• Detection performance for 104 flood events (2022-2024, with only Sentinel-1A operational) showed 70.2% detected, 20.2% missed, and 9.6% with no Sentinel-1 data acquired. Europe had the highest detection rate (95%), while Oceania had the lowest.
• Accuracy (GFM version v2.1.0) showed Overall Accuracy (OA) values of 95.9% for permanent water, 74.4% for seasonal water, and 72.0% for flood. Critical Success Index (CSI) values were 64.1% for permanent water, 55.2% for seasonal water, and 43.7% for flood. The target CSI of 70% was met for 7 out of 12 selected flood events.
• Best accuracies were achieved in temperate and tropical zones, with lower accuracies in arid environments.
• Challenges include overdetection in agricultural/grassland areas, arid environments (dry soil, sand movement), and areas with frozen/snow-covered land, as well as underdetection in dense vegetation and urban areas due to limitations of VV polarisation and algorithmic filtering.

Contributions

• Established the first fully-automatic, global, near-real-time flood monitoring service based on Synthetic Aperture Radar (SAR) data (Sentinel-1).
• Developed and implemented an innovative ensemble approach that integrates three complementary flood mapping algorithms, enhancing the robustness and accuracy of flood detection.
• Created a global 20 meter Sentinel-1 backscatter datacube, enabling continuous monitoring and the generation of a comprehensive historical flood archive from 2015 onwards.
• Generated novel contextual information layers, including monthly Sentinel-1 derived reference water maps, a global exclusion mask, and advisory flags, all tailored to the physical characteristics of SAR data.
• Shifted the paradigm of flood mapping from a scene-specific classification task to a geophysical variable retrieval problem, incorporating uncertainty quantification and evaluating performance across both flood and non-flood scenarios.
• Provided a comprehensive analysis of operational challenges and performance metrics (timeliness, coverage, and accuracy) for a global SAR-based flood monitoring service.

Funding

• European Commission (framework contract 939866-IPR-2020) for the CEMS GFM service.
• Austrian Research Promotion Agency (FFG) through the ScaleFloodS project (FO999900598) for TU Wien.
• German Federal Ministry for Economic Affairs and Energy within the framework of the "National Center of Excellence ML4Earth" (grant number: 50EE2201C) for Jie Zhao.
• Vienna Scientific Cluster (VSC) for computational results.

Citation

@article{Wagner2025fullyautomatic,
  author = {Wagner, Wolfgang and Bauer-Marschallinger, Bernhard and Roth, Florian and Raiger-Stachl, Tobias and Reimer, Christoph and McCormick, Niall and Matgen, Patrick and Chini, Marco and Li, Yu and Martinis, Sandro and Wieland, Marc and Kraft, Franziska and Festa, Davide and Hassaan, Muhammed and Tupas, Mark Edwin and Zhao, Jie and Seewald, Michaela and Riffler, Michael and Molini, Luca and Kidd, Richard and Briese, Christian and Salamon, Peter},
  title = {The fully-automatic Sentinel-1 Global Flood Monitoring service: Scientific challenges and future directions},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.115108},
  url = {https://doi.org/10.1016/j.rse.2025.115108}
}