A local thresholding approach to flood water delineation using Sentinel-1 SAR imagery

Abstract

Abstract Emergency response to floods requires timely information on water extents, which can be produced by satellite-based remote sensing. As the synthetic aperture radar (SAR) can emit and receive signal in nighttime or cloudy conditions, it is particularly suitable to delineate water extent during flood events. Thresholding SAR imagery is one of the most widely used approaches to delineate water extent for its effectiveness and efficiency. However, most thresholding methods rely on a single threshold to separate water and land without considering the complexity and variability of different land surface types in an image. To account for the heterogeneous surface characteristics, this paper proposes a new local thresholding method to water delineation with SAR images. Specifically, our method follows four major steps. First, a global threshold is applied to the SAR imagery to delineate initial water pixels, from which non-water pixels are further clustered into several land surface types. This divides the SAR imagery into one water cluster and several land clusters. Second, local thresholds are estimated at each subset of land cluster paired with water cluster by fitting Gamma distributions to the backscatter intensities of the combined water/land pixels in each subset. Third, local water extents are delineated from each subset and then merged as the union of all subsets. The results are combined across multiple polarizations by taking an intersection operation to generate the global inundation extent. Finally, the flood water extent is further improved by imposing basic hydrologic constraints. This approach is fast and fully automated for flood detection. Our experiments using Sentinel-1 SAR imagery show that the proposed local thresholding approach could distinguish water from non-water with significantly higher accuracy (4–13% improvement in the harmonic mean of user’s and producer’s accuracy of water) than conventional global-thresholding methods.