Floods

The German Remote Sensing Data Center (DFD) of the German Aerospace Center (DLR) has gained a lot of experience in surface water extraction from Synthetic Aperture Radar (SAR) data for various application domains within the last years. In this context four operational approaches have been developed which jointly form the so called DFD Water Suite: The Water Mask Processor (WaMaPro) (Huth et al. 2015, Kuenzer et al. 2013a, 2013b) bases on a simple and high-performance algorithm integrated in a license-free tool, which processes multi-sensor SAR data in order to provide decision makers with information about the location and extent of water surfaces. The Rapid Mapping of Flooding tool (RaMaFlood) (Martinis et al. 2009, 2011) is developed for semi-automatic flood extent mapping using an interactive object-based classification algorithm which is used within flood rapid mapping activities of DLR`s Center for Satellite Based Crisis Information (ZKI). The TerraSAR-X Flood Service (TFS) (Martinis et al. 2015) is used for near-real time rapid mapping activities and provides satellite-derived information about the flood extent in order to support emergency management authorities and decision makers without interaction of an active image interpreter. It bases on a fully-automated processing chain. The last approach is the TanDEM-X Water Indication Mask processor (TDX WAM) (Wendleder et al. 2013). It is part of the processing chain for the generation of the seamless, accurate and high resolution global Digital Elevation Model (DEM) and bases on data of the TanDEM-X mission by combining SAR amplitude and bistatic coherence information. Its purpose is to support the subsequent DEM editing process by the generation of a global reference water mask. In this study, the design of the four approaches and their methodological background are explained in detail, while simultaneously elaborating on the preferred application domains for the different algorithms. The advantages and disadvantages of the four approaches are identified by qualitatively as well as quantitatively evaluating the water masks derived from co-registered single-look slant range complex (CoSSC) data of the TanDEM-X mission for five test sites located in Vietnam, China, Germany, Mali, and the Netherlands (figures 1-5).

SAR based flood detection enables the mapping of large areas regardless of the prevalent cloud and weather conditions. In open (non-forested) areas floods have a considerably lower backscatter compared to dry areas due to the specular reflection on the water surface, whereas in flooded forests the backscatter is typically higher because of corner reflection between water surface and tree trunks. Floods in semi-forested areas typical to the northern boreal forest zone have not yet been investigated deeply. They are not expected to cause backscatter low or high enough to be classified as flooded. The scope of this research is 1) to study how canopy cover (CC) and tree height (TH) affect the backscatter in flooded regions and 2) to create a semi-automatic process which enables near real-time mapping of floods in all forest conditions.

Four Cosmo Sky-Med acquisitions taken during flood events from four different locations in Finland are analyzed against LiDAR based high accuracy forest data. Field measurements for validation purposes were also conducted.  The flood detection process is fully automated, apart from the the selection of flooded and dry training areas from the SAR image, which are used to calculate the threshold values separating flooded and dry areas. A high resolution DEM is used to calculate the flood depth and to estimate whether low CC and low TH areas adjacent to detected floods are flooded or not. To further understand the SAR behavior in flooded forests, the total backscatter is modeled as a function of CC and TH, based on the HUT (Helsinki University of Technology) semi-empirical forest backscattering model.

According to the results, the threshold values separating flood and dry areas were negatively correlated with the incidence angle i.e. higher threshold values when smaller incidence angles. In Kittilä, 85.3 %, Kolari, 89.1 %, Evo, 73.9 % and Pudasjärvi, 82.7 % of the total floods were detected based only on the SAR image. The undetected floods were mostly located in low TH or low CC areas. TH was found to have more influence on the capability of SAR to detect the forest floods than CC. In low TH forests, when the average TH was between 1.5 and 4 m, 25-45 % of floods in Kittilä and 30-65 % of floods in Pudasjärvi were undetected by the SAR. On the contrary, when TH was higher than 6 m, more than 90 % of the floods were detected in both study areas. The assimilation of a high resolution DEM to the SAR based flood recognition significantly improved the flood detection accuracy. The model was well able to explain the SAR backscattering mechanisms of flooded forests.

The results from Evo were weaker due to a geometric inaccuracy of the SAR detected forest floods. This inaccuracy occurred most likely due to the combination of a large incidence angle and high trees. This caused the double bouncing effect to produce larger geometric error, because a radar beam returning to the SAR sensor from a tree trunk was in fact mostly reflected to the tree trunk from the flood surface which can be many meters away. This study confirms that X-band SAR flood mapping in the northern boreal forest zone is possible but can be problematic in low TH and low CC forests, where other spatial data are needed to complement the SAR data. In the southern boreal forest zone more research should be done concerning the geometric shift of the detected forest floods, which is supposedly related to the incidence angle and the tree height.

Floods poses frequent challenges in Norway and the trend is expected to be increasing. During large flood events, it is challenging to get an overview of the situation, and the weather is often too poor for acquisitions of aerial photographs. In this study we investigate the possibilities and challenges in using radar satellite data to monitor flood situations in near real-time.

The proposed service is based on Sentinel-1 Interferometric Wide swath (IW) mode SAR data, and potential floods are detected by comparing an event image with a reference image. Flood-affected areas appear as bright areas in the difference (reference minus event) image. The flood-detection algorithm consists of the following steps: (i) Radiometric calibration to sigma-naught (dB); (ii) Multi-looking; (iii) Geocoding of the SAR images to the desired map projection using a digital elevation model (DEM); (iv) Calculation of the pixel-wise difference image; and (v) Detection of flooded areas. A pixel is masked as flood if the difference minus a local mean value, normalized with respect to the standard deviation, is larger than a given threshold. In order to integrate spatial context, we apply a Markov random field in order to smooth the boundaries. There are potentially many sources that may result in dark areas in the difference image, e.g. wet snow. To reduce the number of false alarms we apply national geographic map data about water bodies (lakes and rivers). In order for a dark area to be classified as flood, it need to be connected a water body.

The algorithm is demonstrated on a set of Sentinel-1 images covering several locations in southern Norway. Preliminary results show that the algorithm is able to estimate the changing area of water bodies, both during the melting season and during floods.

Initially, such surveys are useful for the mapping of affected areas. Eventually, when retrieval quality is improved and ordering procedures are in place, such data may also give valuable input to Emergency Authorities and the Early-warning service of floods in Norway.

Presently, the most common approach for using EO data for flood mapping is based on SAR images. Working in the microwave range of the electromagnetic spectrum (1-10 GHz), SAR is characterized by a good sensitivity to water and is able to provide data day and night, regardless of cloud cover. Several past studies demonstrated that SAR systems are suitable tools for flood mapping so that the use of SAR data is presently well-established in operational services for flood management. For the purpose of a timely and effective flood management, a number of algorithms published in the literature are available to produce near real time (NRT) flood maps. In spite of this progress in the development of NRT flood mapping procedures, the detection of inundation in vegetated and urban areas still represents a critical issue. This is due to the fact that radar signatures of such targets are often ambiguous.

As a matter of fact, there is clearly a necessity to make use of the enhanced observational capabilities of Sentinel-1 in order to find better ways for detecting floodwater in vegetated and urban areas. Indeed the dual-pol acquisitions mode, the medium spatial resolution (20m) with a huge swath width (250 Km), the high repeat cycle (small temporal baseline) and relative narrow orbit tube (small perpendicular baseline in case of interferometric acquisitions) are all characteristics that make momentum for the development of new automatic flood mapping algorithms in more complex land cover environment.

In the present work, the high sensitivity of the interferometric coherence to detect small changes has been exploited to detect the presence of water in urban area, but also to reduce possible false alarms caused by soil moisture effects in case of base soil. Instead, the polarimetric characteristics of Sentinel-1 have been used to detect flood in vegetated area, since the cross- and co-polarized backscattering behave differently respect to the double-bounce between vertical structures and the water on surface. All these effects have been taken into account making use of a change detection approach where changes of the backscattering at the different polarizations, co- and cross-, and the interferometric coherence respect their dry-state have driven the detection of flooded pixels in a statistical based framework.

The test case is the Evros River, the second largest river in Eastern Europe after Danube River, flowing through Bulgaria, Greece and Turkey discharging into the north Aegean Sea. The river has a total length of about 515 km of which 218 flowing along the Greece-Turkey border and the rest in Bulgaria. The Evros river catchment, of 52,900 km² is one of the most intensively cultivated areas in the Balkans with a high socioeconomic and environmental (Evros Delta is protected under Ramsar Convention) importance. Although the last fifty years significant hydraulic works have been constructed several severe floods events struck the catchment and particularly the southern part the most severe of them happen during 2004, 2005, 2006, 2009 and during the current year causing severe damages to agriculture, transport and water supply networks. In particular, in the period between October 2014 and August 2015 have been collected more than twenty dual-pol Sentinel-1 images (VV/VH) making possible to monitor the flood evolution. The flood is present in almost all images, except two, which have been taken as reference. Landsat-8 images, acquired close in time to the Sentinel-1 acquisitions, have been considered as well as a cross validation of the obtained SAR-based flood map.

In the past, the revisit time of SAR satellites was a critical issue for their operational use for flood monitoring. Nowadays, satellite missions such as Sentinel-1 and COSMO-SkyMed or indeed the combination of data from different missions offer the opportunity to obtain a large amount of SAR images that can be used to sample different inundation phases, with metric and decametric spatial resolutions. As a matter of fact, there is a need for fast, reliable and automatic algorithms that are sufficiently flexible to deal with different and often inversely related sensor characteristics. From an image processing point of view, it is notable that a flood event appears differently depending on the sensor resolution. As a result there will be significant differences in terms of percentage of flood pixels in the scene. This directly influences any thresholding-based classification algorithms, because a sufficient number of flood pixels is typically required to estimate a reliable threshold value. It is further evident that the capability to detect changes cannot only rely on spectral signatures but also on the amount of changed pixels w.r.t. the image size. Split-based approaches (SBA) have already been successfully applied to overcome this drawback. They consist in tiling the original image in sub-images and in extracting an adequate threshold value by looking at the statistics of each tile. An alternative is to consider all tiles containing a sufficient number of changed pixels for the estimation of a global threshold. The SBA has also been used for detecting flood extent by considering the backscatter statistics inferred from a single “flood” image in order to separate the “water” class from all other classes. It uses tiles of fixed size that were defined based on the sensor resolution, the image size and the expected extension of the occurred changes.

Here, we introduce a Hierarchical Split-Based Approach (HSBA) where the tile size is not fixed a priori; instead, a hierarchical tiling of the scene is realized, with 4⁰ tiles (the image itself) in the first level, down to the lower one L where the number of tiles is 4^L. Descending from the upper level (full image) to the lower one (quarters of image, sixteenths of image, etc.), only the tiles that fulfil a bimodality criterion both in the flood and in the change image are selected. Next, based on the statistics of the selected area using the HSBA, a hybrid methodology, which combines backscatter thresholding, region growing, and change detection, is used for the automatic extraction of the flood extent in the entire scene.

The algorithm is quantitatively evaluated on the comprehensive data set acquired during the flood event of the Severn River (UK) on July 2007. The image database includes a moderate resolution Envisat-Wide Swath and a high resolution TerraSAR-X Strip Map image, respectively. Evaluation of the performances is carried out by comparing the generated flood maps with a validation map obtained through the manual photointerpretation of aerial photographs taken during the event. The classification obtained accuracies indicate that the algorithm provides reliable classification maps, regardless of spatial resolution and coverage of the input images.  

Moreover, the algorithm has been successfully tested with Sentinel-1 images acquired during the flood events occurred all around world in this last one year and half, such as Po River (Italy) 2014, Malawi 2015,  Bangladesh 2015 and Texas 2015.