Christophe Fatras (Collecte Localisation Satellites, Ramonville-Saint-Agnes, France); Santiago Pena Luque (CNES, Toulouse, France); Pierre Dubois (Collecte Localisation Satellites, Ramonville-Saint-Agnes, France); Ludovic Cassan (CECI, Université de Toulouse, CERFACS/CNRS/IRD, Toulouse, France); Andrea Piacentini (CERFACS, Toulouse, France); Sophie Ricci (CECI, Université de Toulouse, CERFACS/CNRS/IRD, Toulouse, France); Nguyen Thanh Huy (Luxembourg Institute of Science and Technology (LIST), Luxembourg)

The Copernicus Emergency Management Services (EMS), specifically the Emergency Rapid Mapping (ERM) service, requires the rapid generation of flood maps within hours of receiving remote sensing data. This data may be derived from optical or Synthetic Aperture Radar (SAR) sensors, each with distinct properties, such as wavelength, resolution, and band availability, which influence flood detection accuracy. Currently, the process is semi-automatic to reduce misclassification, but large-scale coverage necessitates an accurate automated first guess that can be manually refined. Machine learning (ML) approaches are being developed to assist in flood detection for both optical and SAR data, although challenges remain in regions with low sensor visibility, particularly forests and urban environments. SAR signals in flooded forests are affected by factors such as tree height and canopy density, complicating ML efforts.

In December 2022, the Surface Water and Ocean Topography (SWOT) satellite, equipped with the Ka-band Radar Interferometer (KaRIn), began providing high-quality radar data, offering at least one high-resolution pixel-cloud (PIXC) observation per 21-day repeat cycle over 97 % of the Earth's surface. Along with 2D water level estimation, SWOT provides backscattering, coherent power coefficients, interferometric coherence, and Signal-to-Noise Ratio (SNR) values—key variables for flood detection. Initially, SWOT’s Ka-band data was not designed for high-resolution observations under vegetation and let alone in urban areas. However, this study explores SWOT’s performance during four significant flood events worldwide: Farkadona (Greece, September 15, 2023), Chinon (France, March 31, 2024), Porto Alegre (Brazil, May 5, 2024), and Owensboro (USA, February 20, 2025). Each event is paired with Sentinel-1 or Sentinel-2 imagery within a 3-hour timeframe, providing a valuable opportunity to compare and evaluate SWOT’s flood detection capabilities.

The study demonstrates that SWOT-like satellites show great promise in detecting flooded vegetation and urban areas, using the combined information from the aforementioned radar variables. Additionally, SWOT can detect floods even in regions with high snow cover. However, limitations are observed during the flood recession phase, as wet soils can cause signal saturation, leading to less reliable flood extent estimation. These findings highlight the potential of SWOT-like satellites for improving global flood mapping, as well as the need for further exploration to address current limitations and enhance flood monitoring capabilities in the near future.