Bonassies et al. (2025) A comprehensive study of Surface Water and Ocean Topography (SWOT) Pixel Cloud data for flood extent extraction

Identification

Journal: Remote Sensing of Environment
Year: 2025
Date: 2025-11-05
Authors: Quentin Bonassies, Christophe Fatras, Santiago Peña-Luque, Pierre Dubois, Andrea Piacentini, Ludovic Cassan, Sophie Ricci, Thanh Huy Nguyen
DOI: 10.1016/j.rse.2025.115101

Research Groups

Univ Toulouse, CNRS/Cerfacs/IRD, CECI, Toulouse, France
Collecte Localisation Satellites, Ramonville-Saint-Agne, France
CNES, Toulouse, France
CERFACS, Toulouse, France
Luxembourg Institute of Science and Technology, Esch-sur-Alzette, Luxembourg

Short Summary

This study comprehensively evaluates the capabilities and limitations of the Surface Water and Ocean Topography (SWOT) satellite's Ka-band Radar Interferometer (KaRIn) Pixel Cloud products for flood extent extraction across four major flood events, comparing its performance against Sentinel-1/2 data and its built-in classification. It demonstrates SWOT's potential for detecting floods in vegetated and urban areas while identifying sensitivities to high soil moisture and incidence angle.

Objective

To comprehensively evaluate the capabilities and limitations of the Surface Water and Ocean Topography (SWOT) satellite's Ka-band Radar Interferometer (KaRIn) Pixel Cloud (PIXC) products—specifically backscattering coefficient (𝜎0), coherent power, and interferometric coherence—for accurate flood extent extraction across diverse environments during major flood events, comparing them against Sentinel-1/2 derived flood masks and SWOT's built-in classification.

Study Configuration

Spatial Scale: Four major flood events in Chinon (France), Porto Alegre (Brazil), Owensboro (USA), and Farkadona (Greece). SWOT Pixel Cloud native resolution is approximately 10–70 meters (cross-track) by 5–10 meters (along-track), rasterized to 10 meter × 10 meter grids. Reference data from Sentinel-1 (20 meters × 22 meters) and Sentinel-2 (10, 20, or 60 meters). Digital Elevation Models (DEMs) used include French RGE ALTI (1 meter and 5 meters) and FABDEM v1.2 (30 meters).
Temporal Scale: Flood events occurred between 2023 and 2025. SWOT acquisitions were paired with Sentinel-1 or Sentinel-2 imagery within a 3-hour timeframe. Dry reference data were computed from four SWOT acquisitions for each case, selected from the scientific orbit data (from 2023-07-21 onward) with the same satellite pass as the flood event.

Methodology and Data

Models used:
- FloodML (Random Forest-based classifier for flood detection).
- Histogram thresholding method for SWOT variables.
- Height Above Nearest Drainage (HAND) for manual delineation refinement.
Data sources:
- Satellite:
  - Surface Water and Ocean Topography (SWOT) satellite: Ka-band Radar Interferometer (KaRIn) Pixel Cloud (PIXC) products (L2HRPIXC) including backscattering coefficient (𝜎0), coherent power, and interferometric coherence.
  - Sentinel-1 (C-band Synthetic Aperture Radar).
  - Sentinel-2 (Multi-Spectral Instrument).
- Ancillary/Reference:
  - ESA WorldCover (WC) dataset (10 meter resolution, 2021 version) for land cover classification.
  - Global Surface Water Occurrence (GSWO) dataset (Pekel et al., 2016) used by SWOT's built-in classification.
  - Digital Elevation Models (DEMs): French RGE ALTI (1 meter, 5 meters) and FABDEM v1.2 (30 meters).

Main Results

SWOT KaRIn effectively detects flooded vegetation and urban areas, and performs well even with snow cover (Owensboro event). Small rivers (<10 meters wide) under vegetation were distinguishable using 𝜎0 or coherent power.
Limitations include: (1) signal saturation due to high soil moisture, leading to overdetection (e.g., Farkadona during flood recession); (2) sensitivity to incidence angle, particularly for 𝜎0 and coherent power, impacting flood extent estimation near nadir or at high incidence angles; and (3) low coherence and "dark water" phenomena, which can result in missed flood areas (e.g., Porto Alegre).
Interferometric coherence (𝛾𝑡𝑜𝑡𝑎𝑙) generally exhibits a more stable response to incidence angle compared to 𝜎0 and coherent power, with flood-related values clustering near 1.
A histogram thresholding method, using WorldCover-class-specific thresholds for 𝛾𝑡𝑜𝑡𝑎𝑙, 𝜎0, and coherent power, effectively delineates flood extents, often aligning well with SWOT's built-in classification and FloodML products.
Critical Success Index (CSI) scores (Jaccard Index) varied by event and variable:
- Chinon: 𝛾𝑡𝑜𝑡𝑎𝑙 mask (64.92%), 𝑃𝑐𝑜ℎ𝑒𝑟𝑒𝑛𝑡 mask (69.98%), 𝜎0 mask (62.64%), Classification (59.55%), FloodML (30.91%). SWOT masks outperformed FloodML in this vegetated area.
- Porto Alegre: 𝛾𝑡𝑜𝑡𝑎𝑙 mask (42.54%), 𝑃𝑐𝑜ℎ𝑒𝑟𝑒𝑛𝑡 mask (29.38%), 𝜎0 mask (28.22%), Classification (22.95%), FloodML (70.55%). FloodML performed better due to poor SWOT image quality.
- Owensboro: 𝛾𝑡𝑜𝑡𝑎𝑙 mask (48.42%), 𝑃𝑐𝑜ℎ𝑒𝑟𝑒𝑛𝑡 mask (47.57%), 𝜎0 mask (48.95%), Classification (40.14%), FloodML (34.9%).
- Farkadona: 𝛾𝑡𝑜𝑡𝑎𝑙 mask (20.36%), 𝑃𝑐𝑜ℎ𝑒𝑟𝑒𝑛𝑡 mask (24.51%), 𝜎0 mask (26.61%), Classification (24.44%), FloodML (42.4%). SWOT masks were impacted by high soil moisture.
The computed signal-to-noise ratio coherence (𝛾𝑆𝑁𝑅) provides valuable uncertainty information, enabling the exclusion of low-confidence data from the flood masks.

Contributions

This is the first comprehensive study to integrate SWOT Pixel Cloud data with other satellite observations (Sentinel-1/2) for flood detection.
It provides a detailed assessment of SWOT KaRIn's behavior during flood events, specifically for accurate flood extent extraction, addressing its performance in densely vegetated, urban, and snow-covered areas, and under high soil moisture conditions.
The study develops and evaluates a streamlined, operational method for flood extent extraction using SWOT's backscattering coefficient (𝜎0), coherent power, and interferometric coherence, which enhances the built-in PIXC classification.
It identifies the complementary strengths of the three SWOT variables, noting that 𝜎0 is less sensitive to soil moisture and snow, coherent power is effective for detecting small structures, and interferometric coherence is less sensitive to incidence angle.
The research offers crucial insights into the capabilities and limitations of the novel KaRIn sensor for global flood mapping, which are vital for future research and operational applications, including integration into data assimilation systems.

Funding

Centre National d’Etudes Spatiales (CNES) [grant Number CNES 5100020268]
Collecte Localisation Satellites (CLS) [grant Number CNES 5100020268]
Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique (CERFACS) [grant Number CNES 5100020268]

Citation

@article{Bonassies2025comprehensive,
  author = {Bonassies, Quentin and Fatras, Christophe and Peña-Luque, Santiago and Dubois, Pierre and Piacentini, Andrea and Cassan, Ludovic and Ricci, Sophie and Nguyen, Thanh Huy},
  title = {A comprehensive study of Surface Water and Ocean Topography (SWOT) Pixel Cloud data for flood extent extraction},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.115101},
  url = {https://doi.org/10.1016/j.rse.2025.115101}
}

Original Source: https://doi.org/10.1016/j.rse.2025.115101