An automatic and operational system for near-real time flood extent and duration mapping is presented. The system is based on three fully automatic processing chains for the derivation of the flood extent based on Sentinel-1 and TerraSAR-X Synthetic Aperture Radar (SAR) as well as multi-spectral Sentinel-2 satellite data.
All processing chains contain the following generic steps: automatic data ingestion, pre-processing of the Earth Observation data, computation and adaption of global auxiliary data (digital elevation models, topographic slope information and topographic indices, as well as reference water masks), unsupervised initialization of the classification, post-classification refinement and dissemination of the crisis information via a web-client.
While the systematically acquired Sentinel-1 and Sentinel-2 data allow a continuous monitoring of inundated areas according to the defined observation scenario of the European Space Agency (ESA), the TerraSAR-X flood service needs to be activated on-demand in case of emergency situations. For this purpose dedicated satellite acquisitions are tasked over flood-affected regions (e.g. in the frame of disaster response mechanisms such as the International Charter “Space and Major Disasters”).
The radar-based processing chains make use of an automatic hierarchical tile-based thresholding approach in combination with a fuzzy-logic-based post-processing step for the unsupervised extraction of the flood extent. The thematic Sentinel-2 processing chain performs cloud and shadow segmentation with a globally trained convolutional neural network and extracts flooded areas with a semi-supervised stochastic gradient descent classifier. Training seeds are automatically identified for each image scene by sampling from a Normalized Difference Water Index (NDWI) image.
Further, an additional flood duration layer, which shows the stability of an inundation for each image element in days, is generated in near-real time by using the crisis information derived from the three satellite sources. This is performed by checking the flood status of each pixel along the time series starting from the most recent flood product. The flood duration layer is affected by uncertainties which are a result on information gaps between consecutive data acquisitions. Thus, an uncertainty layer is additionally provided. It shows the total number of days without flood information within a time series of satellite acquisitions where each pixel has been consecutively identified as being flooded. The combination of different satellite products increases the revisit frequencies over the flood-affected regions and, therefore, reduces the uncertainty of the flood duration product.
The flood monitoring system is used to support emergency response activities and provides valuable information for the development of an index-based flood insurance, which aims at securing the livelihood of vulnerable people in rural areas of developing and emerging countries. The system is demonstrated in the Indian states Bihar and Tamil Nadu, which are seasonally affected by flooding due to monsoon rain.

We propose an automatic algorithm for mapping floods affecting urban areas by making use of Sentinel-1 Interferometric SAR coherence. The approach is composed of two steps, it first uses the SAR data to identify buildings and then takes advantage of the InSAR coherence feature to detect the presence of floodwater in urbanized areas. The rationale is that an urban settlement is a very steady target characterized by high multi-temporal InSAR coherence values, even at large temporal baselines. Indeed, previous studies tackling this issue (Chini et al., 2012; Pulvirenti eta l., 2016), take advantage of the fact that the appearance of floodwater implies a significant decrease of the coherence between pairs of images. They have used very high-resolution X-band SAR data and investigated the decrease of the InSAR coherence for double bounce features.

The SAR spatial resolution plays an important role on the paradigm definition of how buildings are represented or how urban areas look like or how changes can be detected. Indeed, in very high-resolution SAR images the double bounce can be identified as the only contribution in a pixel, while in medium resolution images it can be mixed with contributions from other classes. In this latter case, the other present scattering mechanisms can attenuate the main double bounce characteristics, i.e. coherent in time and high intensity value. The enhanced observational capabilities of the C-band Sentinel-1 mission reduce the drawbacks of previous moderate resolution SARs and potentially enable a more accurate mapping of buildings and also the detection of floodwater in urban areas. Indeed, the high repeat cycle (i.e. small temporal baseline) and the relatively narrow orbit tube (i.e. small perpendicular interferometric baseline) are all characteristics that provide new momentum for the development of innovative automatic flood mapping algorithms in urban areas.

In the presented approach, the preliminary detection of buildings step is a pre-requisite for focusing the analysis on double bounce features that also allow to detect the presence of water. To extract the double bounce objects, we use an approach developed for mapping buildings at global scale by means of Sentinel-1 data proposed by Chini et al. (2018). The method is based on the fact that in the co-polarization channel buildings are characterized by high backscattering values as well as by a high degree of temporal correlation between images. To increase the potential for detecting buildings when the line of sight is not orthogonal to the building facade, the cross-polarization backscattering is used as well, as it rises when rotated dihedrals are present. However, high backscattering values can also occur in the presence of vegetation. To cope with false alarms generated by vegetated areas, the algorithm exploits the InSAR coherence. Indeed, the InSAR coherence has a contrasting behavior in vegetated and built-up areas, respectively, as buildings are stable structures unlike vegetation. Another risk of false alarms is caused by the foreshortening effect in mountainous areas. By means of a DEM and the geometric characteristics of the SAR acquisition, the local incidence angle (LIA) is computed and the foreshortening areas are masked out. Thus, the building detection algorithm is composed of two steps: i) discrimination of bright pixels in co- and cross-polarization channels, making use of a statistical modelling-based adaptive thresholding approach (Chini et al., 2017); ii) removal of false alarms by using InSAR coherence and the LIA.

Once the double bounce feature is detected, the detection of the InSAR coherence drop-off on the previously detected double bounce map is used as an indicator of the presence of water. The method is based on a multi-pass approach exploiting a stack of three interferometric SAR acquisitions. The coherence map between each consecutive pair of images is extracted using a kernel sliding window. Given t0, i.e. the date of the image acquired during the flood event, we denote CCco (co-event InSAR coherence) the coherence of the image pair acquired on t0 and t-1, and CCpre (co-event InSAR coherence) the one with images acquired on t-1 and t-2. Finally, we classify as inundated those buildings where CCpre – CCco is positive and higher than a certain value.

For this study, data provided by the Sentinel-1 mission acquired in both Strip Map and Interferometric Wide Swath modes have been used, with a geometric resolution of 5m and 20m, respectively. The algorithm has been tested for two different test cases: the flood that affected the city of Houston (Texas) following the landfall of the Harvey hurricane in 2017 and the flood event that hit the Somalia on May 2018. The results have been evaluated through a comparison with very high-resolution optical images acquired almost simultaneously with the SAR images .To this end, we have employed both crowdsourcing points derived by photointerpretation and inundation models over the area. The comparison with such independent data has shown that the proposed approach has been able to detect flooded urban areas with high accuracy.

M. Chini, R. Pelich, R. Hostache, P. Matgen, C. Lopez-Martinez, “Towards a 20 m Global Building Map from Sentinel-1 SAR Data”, Remote Sensing, 2018.

M. Chini, R. Hostache, L. Giustarini, P. Matgen, “A Hierarchical Split-Based Approach for parametric thresholding of SAR images: flood inundation as a test case”, IEEE Transactions on Geoscience and Remote Sensing, 55, 6975-6988, 2017.

L. Pulvirenti, M. Chini, N. Pierdicca, G. Boni, “Use of SAR Data for Detecting Floodwater in Urban and Agricultural Areas: The Role of the Interferometric Coherence”, IEEE Transactions on Geoscience and Remote Sensing, 54, 1532-1544, 2016.

M. Chini, L. Pulvirenti, N. Pierdicca, “Analysis and interpretation of the COSMO-SkyMed observations of the 2011 Japan tsunami”, IEEE Geosciences and Remote Sensing Letters, 9, 467-471, 2012.

Flooding in urban areas is a serious threat to human lives and social property and up-to-date information about flood extent and evolution are critical for emergency response. Synthetic Aperture Radar (SAR) remote sensing has been widely used for flood mapping and monitoring, but the detection of flooded areas in urban environments still proves to be particularly challenging to SAR due to the complex landscapes. Deep learning showed high capability in visual pattern recognition and more recently has also shown promising results in remote sensing applications. In this study, we assess the roles of SAR intensity and coherence in urban flood detection based on a deep convolution neural network (CNN) and develop an active self-learning CNN framework to alleviate the effect of a limited annotated training dataset. The study area is located in Houston, Texas, which was affected by flooding related to heavy rainfalls accompanying hurricane Harvey in August 2017. Four TerraSAR-X HH-polarized Stripmap data sets are used as data basis for mapping this flood event. The following five scenarios are considered: scenario I : only co-event coherence is used; scenario II: only co-event intensity is used; scenario III: both co- and post-event coherence are used; scenario IV: both pre- and co-event intensity are used; scenario V: the combination of multi-temporal intensity and coherence are used. Our experiments indicate that multi-temporal intensity (scenario IV) is the most important information source for urban flood detection, and the additional information of multi-temporal coherence increases the reliability of the results (scenario V). A temporal-ensembling active self-learning CNN framework was developed to leverage both labeled and unlabeled data. The framework integrates active learning and self-learning at the conceptual level based on a temporal-ensembling model. It works in an iterative fashion. The informative unlabeled data are pseudo-labeled with a multi-level spatial constraint and updated to the existing training dataset. To penalize the incorrect label caused by pseudo-labeling, a consistency-regularization is introduced. The proposed framework is in a machine-machine interaction process, thus no extra annotation effort by an expert is needed. The experiments were evaluated by very high-resolution optical image, which show encouraging results by the proposed active self-learning framework: an overall 11.6% (from 0.6144 to 0.6856) improvement in terms of к statistic is achieved in comparison to its supervised counterpart.

Flooding due to ice jams is a major and frequent hazard for communities located near rivers in cold and northern regions of the world. Ice jams might be formed during the development of ice covers in early winter or during ice cover break up in early spring. A related issue is the management of river flows on regulated rivers, especially the routing of early spring runoff during ice conditions while minimizing the risk of initiating ice jams. Flood forecasting for such rivers is challenging due to the complexities of modelling ice formation and breakup in real-time in an operational context. It is further complicated by the effect of climate change, which has changed historically observed ice cover formation behavior.

This paper describes the creation and expansion of an operational river ice monitoring program based on satellite RADAR imagery to address the challenges of flood forecasting, river flow management and climate change adaptation. Initially developed and implemented for the Town of Badger on the island of Newfoundland, Canada, following a severe river ice induced flood in 2003, the satellite-based service is provided to communities at risk across Canada. Satellite-based river ice monitoring initially relied on the use of satellite RADAR imagery from the ENVISAT, RADARSAT 1 and RADARSAT-2. The lack of available SAR imagery due to conflict during critical periods of the monitoring season made it necessary to integrate routinely acquired LANDSAT and MODIS data into the monitoring process, although the utility of optical imagery is severely limited by the prevalence of cloud cover.

The availability of Sentinel-1 imagery represented a step change in terms of reduced acquisition conflicts and increased spatial and temporal coverage capabilities. Today, C-CORE’s river ice monitoring services rely on a variety of operational EO missions, including Sentinel-1, RADARSAT-2, TerraSAR-X, COSMO-SkyMed, Sentinel-2, Sentinel-3, LANDSAT-8 and MODIS. This investigation will present lessons learned from 15 years of EO-based for river ice monitoring, discuss issues of operational integration, highlight the contribution to emergency response, emphasize the importance for climate change adaptation and present results from the synergistic use of satellite monitoring and model-based forecasting for flood risk mitigation.

This paper presents an automated Synthetic Aperture Radar (SAR) multi-temporal and statistical analysis process designed for false alarm reduction in flood detection and monitoring at complex transboundary river system based on Sentinel-1 SAR data.
In the specific study, the methodology is based on a 3-step approach. First, the preprocessing of the Sentinel-1 data is conducted using an automatic approach based on the Sentinel Application Platform (SNAP) toolbox. Then an automatic dynamic thresholding approach is used for first identification of areas with low backscatter (generally flooded areas) which are later cleared out using a statistical analysis of the flooded images with respect to the reference ones.
Similar methodology has been applied in Twele et al. (2016) and Li et al. (2018) where results provided sufficient classification results with overall accuracies over 90%. In this paper, the assessment of Sentinel-1 data combined with the above-mentioned methodological tool is validated on the latest severe flood event occurred on January till April 2018 at the transboundary river Evros, an area shared between Bulgaria, Greece and Turkey. The specific area has been acknowledged as a complex river system environment in the relevant literature (Ireland, Volpi and Petropoulos, 2015; Chini et al., 2016; Twele et al., 2016; Mentzafou, Markogianni and Dimitriou, 2017; Li et al., 2018) where the identification of the real flood is difficult because of the different types of land cover (bare and poorly vegetated soil, vegetated areas and urban areas) and especially the presence of rice crops. Except for the technical issues that may occur when studying the specific area, the cooperation in terms of exchange information is not always feasible between the three countries. The situation is even more challenging for researchers when it comes to the cross-border area between Greece and Turkey due to political matters.
The preliminary analysis revealed some significant findings (figure 1). It was identified that the real flood period significantly differs from the official dates given by the Greek Civil Protection authority (Region of East Macedonia and Thrace). Specifically, according to the Greek authorities the flood event in the Greek territory lasted for almost a two-month period (March until April 2018) which does not correspond to the real flood event in the transboundary scale.
Moreover, using a sensitivity analysis for the threshold value various prediction models were developed to allow more detailed analysis than the proportion of correct classifications (accuracy). Specifically, the aim was set on maximizing the F1 score which is the harmonic mean of sensitivity and specificity. Results show that while the methodology may provide reduced metrics in terms of accuracy, the predictive power is increased in terms of eliminating both false positives (false alarms) and false negatives.
The results of this study come as a supportive argument in the context of establishing an integrated transboundary flood monitoring and management approach in such river systems. Satellite Earth Observations (EO) proved that they are not only an unambiguous asset that provide useful information on the flood evolution that is unhindered by geopolitical boundaries but also have access to remote and unreachable areas providing frequent and reliable data. This study can set the framework for the development of a dynamic transboundary flood hazard monitoring system for the whole river basin for an efficient and effective flood risk management. For our future research attempts, the proposed approach, which is entirely based on open access remotely sensed data and processing tools, could be implemented in the same area for past flood events to produce dynamic archive retrospective data and use advanced machine learning techniques to further improve classification metrics.

Keywords: Sentinel-1, SAR multi-temporal, statistical analysis, flood extent mapping, false alarm reduction, transboundary river

References
Chini, M. et al. (2016) ‘SAR coherence and polarimetric information for improving flood mapping’, in 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). IEEE, pp. 7577–7580. doi: 10.1109/IGARSS.2016.7730976.
Ireland, G., Volpi, M. and Petropoulos, G. P. (2015) ‘Examining the capability of supervised machine learning classifiers in extracting flooded areas from landsat tm imagery: A case study from a mediterranean flood’, Remote Sensing, 7(3), pp. 3372–3399. doi: 10.3390/rs70303372.
Li, Y. et al. (2018) ‘An automatic change detection approach for rapid flood mapping in Sentinel-1 SAR data’, International Journal of Applied Earth Observation and Geoinformation, 73(May), pp. 123–135. doi: 10.1016/j.jag.2018.05.023.
Mentzafou, A., Markogianni, V. and Dimitriou, E. (2017) ‘The Use of Geospatial Technologies in Flood Hazard Mapping and Assessment: Case Study from River Evros’, Pure and Applied Geophysics, 174(2), pp. 679–700. doi: 10.1007/s00024-016-1433-6.
Twele, A. et al. (2016) ‘Sentinel-1-based flood mapping: a fully automated processing chain’, International Journal of Remote Sensing, 37(13). doi: 10.1080/01431161.2016.1192304.

In recent decades, many factors have contributed to the considerable increased of natural disasters related to flood events. A rapid response in the face of these hazards is crucial since flooding is considered one of the most costly type of natural disaster in terms of property damage and mortality. In this framework, Sentinel-1 (S1) imagery provides useful data for the monitoring of damaged areas from local to regional scale with a lower cost compared with conventional techniques. Being S1 a Synthetic Aperture Radar (SAR) sensor, it provides repeated observations regardless of weather conditions and day-night cycles, and allows a frequent update of information that is extremely useful in front of disaster situations.
This paper presents a semi-unsupervised flood monitoring approach developed by the company DARES TECHNOLOGY using SAR images that can improve the disaster risk management and the coordination of the response activities in front of flood crisis events. The core of this flood mapping algorithm consists in analysing the SAR differences in slant-range coordinates supported by multi-temporal interferometric parameters such as the coherence and change detection techniques.
The first step of the approach proposed is based on reducing the speckle of SAR images which is crucial to reduce the number of outliers due to the roughness of water. Among the wide range of speckle filters used in the SAR community (Median, Lee, Frost, Kuan, Gamma filter, etc.), the use of the NLSAR filter is proposed in this paper [1]. This filter allows the consideration of pixels that are far apart rather than restricting to similar pixels that belong to the local neighborhood, such as boxcar filtering. One pixel is assumed to come from the same statistical population as the given pixel if the patches that surround both pixels present a high similarity measure. The weights are computed using the maximum likelihood estimator (WMLE) and are iteratively refined by comparing patches of previous estimates and patches of the noisy image.
Then, SAR images are divided in spatial blocks to have a suitable sampling of water and non-water classes and estimate a suited threshold to separate both classes. The following step consists of identifying those blocks with bi-modal behaviour using the statistical Hartigan’s Dip test over all the spatial blocks of the image. After this point, Kittler-Illingworth (KI) thresholding algorithm [2,3] is applied adaptively in order to classify water and non-water classes. This first rough water mask may be improved by means of segmentation based on classification approaches. Different strategies have been investigated. On the one hand, algorithms that minimize cost functions such as extended K-Means, agglomerative hierarchical clustering or Fuzzy K-means have been investigated. On the other hand, more advanced but time consuming algorithms, such as Gaussian Mixture or Spatial Clustering, have been also evaluated.
At this point, flood maps can be computed through the differences between all the water masks and a mask of reference selected before the flood event. Flood maps are then improved using multi-temporal SAR-based change detection techniques exploiting the so-called log-ratio and, additionally, interferometric parameters such as the coherence.
The challenge of this detection approach is working towards an unsupervised one, i.e., training data is no required and, therefore, ground-truth about the class statistics within the area of interest are not necessary.
Results corresponding to the flood analysis of two locations are provided to illustrate the benefits of the algorithm proposed: the Ebro basin, Spain (on April 15th, 2018), and the Attapeu/Champasak area, Laos (on July 23rd, 2018). The former refers to floods affecting Zaragoza province caused by heavy rainfall and snowmelt in the Pyrenees Mountains. The latter refers to severe floods caused by the collapse of part of a hydroelectric dam system that led to the massive destruction and damages in the nearby areas.

REFERENCES

[1] Deledalle, C.-A., Denis, L., and Tupin, F., “NL-InSAR: Nonlocal Interferogram Estimation”, IEEE Transactions on Geoscience and Remote Sensing, 49 (4), 1441-1452, Apr.2011.
[2] S. Martinis, A Twele and S. Voigt, “Towards operational near real-time flood detection using a split-based automatic thresholding procedure on high resolution TerraSAR-X data”, Natural Hazards and Earth System Sciences, 9, 303-314, Mar. 2009.
[3] Kittler, J. and Illingworth, J.: Minimum error thresholding, Pattern Recogn., 19, 41–47, 1986