Recently

Recently, there has been a significant improvement in the development of SAR-based algorithms for detecting FV (Cazals et al. 2016; Mleczko & Mróz 2018; Muro et al. 2016; Plank et al. 2017; Twele et al. 2016; White et al. 2014). These algorithms have in common that they make use of thresholding for the initialization of the classification process. The accuracy of thresholding varies drastically of the land cover (LC) characteristics prevalent in the scene (e.g. rough soil surface, vegetation). Theoretical electromagnetic backscattering models have traditionally been used to define the threshold for mapping FV (Pulvirenti et al. 2013). However, such approaches require detailed soil, vegetation and LC maps, to accurately estimate the threshold, which are often unavailable. Hence the need to shift to approaches that determine automatically the optimal threshold value for a scene/landscape, which take into account double-bouncing vegetation and the diffuse backscattering of dry areas. Martinis, Twele & Voigt (2009) applied the Kittler and Illingworth’s (KI) global parametric thresholding algorithm, which uses a minimum error approach to group the sets of pixels of grey-scale images into object and background classes for near-real time flood detection using a split-based automatic thresholding procedure on TerraSAR-X. Matgen et al. (2011) performed thresholding by modeling the flood class using a non-linear fitting algorithm under the gamma distribution assumption. Schumann et al. (2010) and Pulvirenti et al. (2012) compute a threshold value from the global grey level histograms of SAR data using the widely used Otsu’s method (Otsu 1975). The Otsu method finds a threshold that minimizes the between-class variance of water and non-water areas.
Automatic processing chains based on automatic thresholding offer rapid flood mapping activities that have helped to improve the delivery time of the emergency information to the users. Twele et al. (2016) proposed an automated Sentinel-1 based processing chain for detection and monitoring of floods in near-real time (NRT). However, the algorithm has been applied on single SAR images for open flood detection. Although a single SAR image can provide a reasonable estimate of the flood extent in an area, the approach of setting a threshold for the probability of the flood presence is inflexible, especially in vegetated flood plains. Changes in local LC are difficult to deduce from a single image, but from a series of SAR images additional information such as temporary FV can be extracted by considering a trend in time series of backscatter images.
This study presents a semi-automated technique for mapping the entire flooded area including temporary open water and flooded vegetation (FV), based on the time series of Sentinel-1 datasets and with methods that are both quick and easy to apply over large areas as new data becomes available.