Robust automated thresholding of SAR imagery for open-water detection

Most thresholding algorithms reported in the literature for water detection in SAR imagery are optimized for a single sensor, limited collection parameters, or scene content. This study uses a diverse library of data (106 independent collects) to develop a robust thresholding algorithm for automated water detection in SAR imagery. The library includes RadarSat-2, TerraSAR-X, and COSMO-SkyMed collects for a diverse set of ground sites (18 sites; littoral and riverine). The study of water features extracted from this library provides a platform from which to investigate the relationships between scene content, collection parameters, and the ability to extract water features in an automated way. This is the first work to the authors´ knowledge that rigorously compares the automated detection of water features using a library of data with an approximation of ground truth. Two key observations are: 1) a positive correlation between the fraction of water in a scene and the classification accuracy; 2) cross-pol (i.e. HV or VH) often provides better contrast for segmentation of open-water and non-open-water than the HH polarization.