In this paper, an automatic near-real time (NRT) flood detection approach is presented, which combines histogram thresholding and segmentation based classification, specifically oriented to the analysis of single-polarized very high resolution Synthetic Aperture Radar (SAR) satellite data. The challenge of SAR-based flood detection is addressed in a completely unsupervised way, which assumes no training data and therefore no prior information about the class statistics to be available concerning the area of investigation. This is usually the case in NRT-disaster management, where the collection of ground truth information is not feasible due to time-constraints. A simple thresholding algorithm can be used in the most of the cases to distinguish between "flood" and "non-flood" pixels in a high resolution SAR image to detect the largest part of an inundation area. Due to the fact that local gray-level changes may not be distinguished by global thresholding techniques in large satellite scenes the thresholding algorithm is integrated into a split-based approach for the derivation of a global threshold by the analysis and combination of the split inherent information. The derived global threshold is then integrated into a multi-scale segmentation step combining the advantages of small-, medium- and large-scale per parcel segmentation. Experimental investigations performed on a TerraSAR-X Stripmap scene from southwest England during large scale flooding in the summer 2007 show high classification accuracies of the proposed split-based approach in combination with image segmentation and optional integration of digital elevation models.

Wildfires have major ecological, social and economic consequences. Information about the extent of burned areas is essential to assess these consequences and can be derived from remote sensing data. Over the last years, several methods have been developed to segment burned areas with satellite imagery. However, these methods mostly require extensive preprocessing, while deep learning techniques—which have successfully been applied to other segmentation tasks—have yet to be fully explored. In this work, we combine sensor-specific and methodological developments from the past few years and suggest an automatic processing chain, based on deep learning, for burned area segmentation using mono-temporal Sentinel-2 imagery. In particular, we created a new training and validation dataset, which is used to train a convolutional neural network based on a U-Net architecture. We performed several tests on the input data and reached optimal network performance using the spectral bands of the visual, near infrared and shortwave infrared domains. The final segmentation model achieved an overall accuracy of 0.98 and a kappa coefficient of 0.94.

In this study, an analysis of multi-temporal and multi-frequency Synthetic Aperture Radar data is performed to investigate the backscatter behavior of various semantic classes in the context of flood mapping in central Europe. The focus is mainly on partially submerged vegetation such as forests and agricultural fields. The test area is located at River Saale, Saxony-Anhalt, Germany, which is covered by a time series of 39 TerraSAR-X data acquired within the time interval December 2009 to June 2013. The data set is supplemented by ALOS PALSAR L-band and RADARSAT-2 C-band data. The time series covers two inundations in January 2011 and June 2013 which allows evaluating backscatter variations between flood periods and normal water level conditions using different radar wavelengths. According to the results, there is potential in detecting flooding beneath vegetation in all microwave wavelengths, even in X-band for sparse vegetation or leaf-off forests.

In recent years, the German Remote Sensing Data Center (DFD) of the German Aerospace Center (DLR) has gained a lot of experience in water surface extraction from synthetic aperture radar (SAR) data for various application domains. In this context, four approaches have been developed, which jointly form the so-called DFD Water Suite: The Water Mask Processor (WaMaPro) is based on a simple and high-performance algorithm that processes multi-sensor SAR data in order to provide decision-makers with information about the location of water surfaces. The Rapid Mapping of Flooding tool (RaMaFlood) has been developed for flood extent mapping using an interactive object-based classification algorithm. The TerraSAR-X Flood Service (TFS) is used for rapid mapping activities and provides satellite-derived information about the extent of floods in order to support emergency management authorities and decision-makers. It is based on a fully automated processing chain. The last approach is the TanDEM-X Water Indication Mask processor (TDX WAM). It is part of the processing chain for the generation of the seamless, accurate, and high-resolution global digital elevation model (DEM) produced based on data of the TanDEM-X mission. 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 backgrounds 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 data of the TanDEM-X mission for five test sites located in Vietnam, China, Germany, Mali, and the Netherlands.

Due to the similarity of the radar backscatter over open water and over sand surfaces a reliable near real-time flood mapping based on satellite radar sensors is usually not possible in arid areas. Within this study, an approach is presented to enhance the results of an automatic Sentinel-1 flood processing chain by removing overestimations of the water extent related to low-backscattering sand surfaces using a Sand Exclusion Layer (SEL) derived from time-series statistics of Sentinel-1 data sets. The methodology was tested and validated on a flood event in May 2016 at Webi Shabelle River, Somalia and Ethiopia, which has been covered by a time-series of 202 Sentinel-1 scenes within the period June 2014 to May 2017. The approach proved capable of significantly improving the classification accuracy of the Sentinel-1 flood service within this study site. The Overall Accuracy increased by ~5% to a value of 98.5% and the User’s Accuracy increased by 25.2% to a value of 96.0%. Experimental results have shown that the classification accuracy is influenced by several parameters such as the lengths of the time-series used for generating the SEL.