Flood mapping using Sentinel-1 imagery with topographical and hydrological contextualization: Case study from Ribe, Denmark

Abstract

Advancements in Synthetic Aperture Radar (SAR) imagery have made it the standard datasource for large-scale operational flood mapping. SAR’s applicability under all-weather conditions and at night is a major advantage. However, challenges remain in mapping low-contrast surface water due to emergent vegetation and heterogenous flood extent variability. To address these issues, we propose a framework applicable for fully automatic flood mapping. The proposed framework was tested using Sentinel-1 SAR imagery in Ribe, Denmark, a site with frequent inundation with highly variable magnitudes. The framework features several novel methods for refining surface water extents with topographical and hydrological contextualization. A bimodal mask is generated from quadtree decomposition and gaussian mixture modelling, in combination with a bimodality test, which enables straightforward determination of local thresholds separating water and background. Mapped flood extents are contextually refined with ancillary topographical and hydrological datasets, using region-growing and linear regression. A nuanced surface water likelihood output is created from a fuzzy logic procedure using image specific backscatter coefficient statistics, topographic position index and height above nearest drainage. Results were verified through comprehensive spatial- and temporal validation, using Sentinel-2 optical imagery, a permanent water dataset, and timeseries of gauged stream water elevation. A satisfying result was achieved with an average overall accuracy of 98.5 %, a temporal correlation with gauged stream elevations of 0.92, and a total of 82.4 % of permanent water surfaces mapped correctly during peak flooding.