Sentinel-1 Flood Maps Using Exponential Filter as No-Flood Reference

Background

The TU Wien flood mapping algorithm is a Sentinel-1-based workflow using Bayes Inference at the pixel level. The algorithm is currently deployed in global operations under the Copernicus GFM project and have been shown to work generally well. However, the current approach has overestimation issues related to imperfect no-flood probability modeling. In a recent study, we proposed and compared an Exponential Filter derived from no-flood references versus the original Harmonic Model. We have conducted experiments on seven study sites for flooded and no-flood scenarios. A full description and discussion are found in the paper: Assessment of Time-Series-Derived No-Flood Reference for SAR-based Bayesian Flood Mapping.

Methodology

• We generated no-flood references using the Exponential Filter at various T-parameter values and the original Harmonic Model as a baseline.
• Flood maps were generated using the Bayes Inference-based SAR Flood mapping algorithm implemented in Python using the Yeoda software package. Flood maps using the various no-flood references for all available Sentinel-1 image acquisitions for a selected relative orbit per study site.
• Each flood map is compared with the reference CEMS Rapid Mapping or Sentinel Asia reference dataset to generate validation/confusion maps.

Technical details

• Datasets are stored in GeoTiff format using LZW Compression.
• Files are compressed in two bundles: 1) flood maps, 2) false positive count maps, and 3) validation results.
• Files are organized and tiled following the T3 Equi7Grid tilling system at 20m x 20m resolution.
  • Folder structure: dataset/map product>(continental)subgrid>tile>files.
  • The study covers the following study sites:
    • EU E039N027T: Scotland
    • AS E054N015T3: Vietnam
    • EU E054N006T3: Greece
    • EU E051N012T3: Slovenia
    • AS E024N027T3: India
    • OC E057N117T3: Philippines
    • EU E057N024T3: Latvia
• Files are named following the Yeoda file naming convention.
• Summary Accuracy Assessment Metrics are in CSV format.

Datasets:

• Flood: flood maps generated using different parameterizations of no-flood reference.
• FP_Count: false positive count maps.
• Validation results include:
  • Confusion maps were generated from the difference between the flood maps and the rasterized CEMS Rapid Mapping reference or Sentinel Asia datasets. Summary Accuracy Assessment Metrics in CSV format.
  • ERA5-LAND daily aggregates in CSV format.
  • Root Mean Square Error time-series analysis in CSV format.
  • False Positive Rate time-series analysis in CSV format.
• *Due to storage constraints, no flood reference is available upon request.