Satellite-based Mapping of Flood Extent in Denmark

Floods stand out as the most frequent and costly natural disaster in Europe. In the EU alone, there have been documented more than 1500 flood events since 1980, causing over 4300 deaths and more than €170 billion in economic damages.

Due to the compounded developments of urbanization and climate change, the frequency of floods is expected to increase with severe impacts, possibly endangering lives and leading to economic losses. Moreover, floods mobilize pollutants stored in the subsurface and urban areas. Thus, current efforts, such as coastal barriers, restoration of river courses, or resilient city and landscape planning, focus on reducing vulnerability and risks from flooding. But to implement such measures, detailed information on where and when flooding occurs is necessary. This study aims to improve and implement satellite-based mapping of flood extent under Danish conditions by presenting different methods and algorithms utilizing Sentinel-1 (S1) Synthetic Aperture Radar (SAR) imagery, digital elevation models (DEM) and river geometry. In the broader literature, various methods have been proven to successfully map flood extent, such as deep learning (DL) and change detection (CD) as employed in the Global Flood Awareness System. However, DL require extensive training and labeled data that are often not available, and CD is reliant on a comprehensive pre-processing procedure of antecedent satellite imagery or accompanied with a datacube-based algorithm that exploits the satellite orbit repetition. While these methods can provide excellent results, the steep data requirements and pre-processing procedures hinder practical usage. On the other hand, single-temporal image flood extent mapping algorithms relying on histogram analysis offering a straightforward approach potentially yielding satisfying results, especially when accompanied by techniques such as image decomposition, region-growing, active contour models or image texture algorithms. But for single-temporal image histogram analysis to work in an automated setup, the two main problems, namely class imbalance and class overlap must be addressed properly. This study proposes a novel approach for single-temporal image histogram analysis by combining automatic local histogram thresholding with two image decomposition techniques for image tiling using a quadtree and a novel combination of k-means clustering and box tiling. This study implements a bimodality test and a subsequent local-threshold selection using gaussian mixture modelling and kernel-density smoothening, followed by contextual segmentation using region-growing. Furthermore, a novel approach for improving flood extent segmentation using a combination of DEM information, geographical stream location and region-growing is presented. The proposed method is showcased for two different flood events in Denmark from 2015 to 2022 using 10 x 10 m interferometric wide swath S1 SAR imagery. Results are evaluated using Sentinel-2 optical imagery where available, and otherwise evaluated against high-precision permanent water maps. Moreover, we utilize gauged timeseries of stream water level to evaluate the temporal evolution of flood extent over the period of a flood event.