High-Resolution 3-D Flood Information From Radar Imagery for Flood Hazard Management

 Internet

 4 views
of 11
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Description
High-Resolution 3-D Flood Information From Radar Imagery for Flood Hazard Management
Share
Tags
Transcript
  IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 45, NO. 6, JUNE 2007 1715 High-Resolution 3-D Flood Information From RadarImagery for Flood Hazard Management Guy Schumann, Student Member, IEEE  , Renaud Hostache, Christian Puech, Lucien Hoffmann,Patrick Matgen, Florian Pappenberger, and Laurent Pfister  Abstract —This paper presents a remote-sensing-based steady-state flood inundation model to improve preventive flood-management strategies and flood disaster management. TheRegression and Elevation-based Flood Information eXtraction(REFIX) model is based on regression analysis and uses a re-motely sensed flood extent and a high-resolution floodplain digitalelevation model to compute flood depths for a given flood event.The root mean squared error of the REFIX, compared to ground-surveyed high water marks, is 18 cm for the January 2003 floodevent on the River Alzette floodplain (G.D. of Luxembourg), onwhich the model is developed. Applying the same methodology ona reach of the River Mosel, France, shows that for some morecomplex river configurations (in this case, a meandering riverreach that contains a number of hydraulic structures), piecewiseregression is required to yield more accurate flood water-lineestimations. A comparison with a simulation from the HydrologicEngineeringCentersRiverAnalysisSystemhydraulicfloodmodel,calibrated on the same events, shows that, for both events, theREFIX model approximates the water line reliably.  Index Terms —Flood information mapping, light detecting andranging (lidar) digital elevation model (DEM), regression analysis,synthetic aperture radar (SAR), SAR data uncertainty, 1-D hy-draulic model. I. I NTRODUCTION W ITH THE onset of climate change implying an ampli-fied frequency of floods in many regions of the world[1] and a steadily increasing number of economic assets beinglocated within flood-prone areas, the need to understand andprotect against flooding has become increasingly importantfor our society. The spatial characterization of hazard andrisk is of paramount importance in any flood-managementplan. Hence, an appropriate, rapid, and effective response toany flood-induced disaster is essential. Only remote sensingwith its extensive spatial coverage in conjunction with field Manuscript received April 28, 2006; revised August 17, 2007. This work wassupported by the “Ministère Luxembourgeois de la Culture, de l’EnseignementSupérieur et de la Recherche.”G. Schumann is with the Public Research Centre-Gabriel Lippmann (EVADepartment), 4422 Belvaux, Luxembourg, and also with the GeographyDepartment, Dundee University, DD14HN Dundee, U.K. (e-mail: schumann@lippmann.lu).R. Hostache and C. Puech are with the Maison de la Télédétection, Cema-gref, 34093 Montpellier, France (e-mail: renaud.hostache@teledetection.fr;puech@teledetection.fr).L. Hoffmann, P. Matgen, and L. Pfister are with the Public Research Centre-Gabriel Lippmann (EVA Department), 4422 Belvaux, Luxembourg (e-mail:hoffmann@lippmann.lu; matgen@lippmann.lu; pfister@lippmann.lu).F. Pappenberger was with the Hydrology and Fluid Dynamics Group, De-partment of Environmental Science, Lancaster University, LA14YQ Lancaster,U.K. He is now with the European Centre for Medium Range Weather Fore-casts, RG2 9AX Reading, U.K. (e-mail: Florian.Pappenberger@ecmwf.int).Digital Object Identifier 10.1109/TGRS.2006.888103 measurements collected at high temporal resolution is capableof delivering the kind and quantity of information needed tomeet these objectives most satisfactory. Optical imagery hasbeen successfully used in the past to extract flood areas [2].However, given the rapid flood recession in small- to medium-sized catchments and weather conditions during events, flooddetection with visible satellite imagery seems not feasible inmany regions of the world. With its ability to acquire dataduring all meteorological conditions, day and night, and itscapability to provide information about the extent of open waterbodies, synthetic aperture radar (SAR) instruments present analternativetoopticalimagery,aerialphotography, andhydraulicmodel simulations for mapping flood extents over large areasand thus facilitate effective flood disaster management.To date, the use of radar imagery is very often neglected inflood disaster management, as it is most of the time limitedto a binary segmentation into flooded and nonflooded pixelsand because of the inability of SAR to record flooding inurban areas due to the corner reflection principle. Furthermore,SAR images have a medium spatial resolution, are very oftenlimited to a single frequency, are often difficult to orthorectifyaccurately, and tend to be geometrically and radiometricallydistorted. Another important issue is the revisit time associatedwith the current radar satellites, which can take up to 35 days.Having SAR imagery acquired on the day of the flood isquite fortunate indeed. However, considering the 24- to 48-himage delivery time in case of a major (flood) disaster, SARcan be considered a very valuable and welcomed source of (flood) information. Moreover, with the numerous recent andupcoming SAR satellite missions (e.g., ALOS, RADARSAT-2,TerraSAR-X, and COSMO-SkyMed), more timely image de-livery will be possible and a satellite constellation could beimagined.As high-resolution digital elevation models (DEMs) becomemore readily available, it is possible to map not only floodextents but also flood depths for a given event. Complementaryto the remote-sensing observations, a DEM can also be used toextract flooding under dense vegetation and forest cover [2], [3]and within residential areas. From a disaster managementpoint of view, water depth information for a given floodevent as well as flood visualization within urban areas is of paramount importance for obvious reasons such as propertyloss and damage, health issues related to standing waters, andthe assessment of socioeconomic damages. Thus, for opera-tional flood-management applications, the flood extent mapsthat are currently obtained with remote-sensing observationsappear to be only of limited utility. Hence, retrieving water 0196-2892/$25.00 © 2007 IEEE  1716 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 45, NO. 6, JUNE 2007 depth information by integrating the SAR imagery and high-resolution DEMs would present a significant step toward moreeffective flood management. In this context, Oberstadler et al. [4] overlay a flood extent on topographic contour lines todetermine water levels. The vertical root mean squared er-ror (RMSE) is 0.5 to 2 m with respect to field data.Brackenridge et al. [5] use the same approach and come to asimilar conclusion by reporting a 1- to 2-m precision. Puechand Raclot [6] propose a rather complex methodology thatis based on extensive fieldwork, hydraulic knowledge, andaerial photography interpretation skills. First, the floodplain issegmented into polygons on which water levels are supposedto be horizontal. Used in conjunction with topographic maps,interpretation of aerial photography followed by field measure-ment campaigns gives a vertical RMSE of 23 cm. Despite theseencouraging results, the technique seems difficult to adapt toSAR images due to their inappropriate spatial resolution andtheir relatively high level of distortions. Hostache et al. [7]apply a depth-mapping method to a RADARSAT-HH SARflood image of the 1997 flood on the Mosel River (France),which accounts for hydraulically sensitive zones and positionaluncertainty of the SAR-derived flood map. They end up withan uncertainty between maximum and minimum water depthestimation of around 30 cm, on average.II. M OTIVATION AND A VAILABLE D ATA  A. Preliminary Study and Motivation The following study presents a modified and improved ver-sion of a remote-sensing-based steady-state flood modelingapproach proposed by Matgen et al. [3]. In a steady-stateregime, a system (in this case, flood flow) is modeled assumingthat hydraulic demands and boundary conditions do not changewith respect to time. Matgen et al. [3] propose to draw crosssections perpendicular to the river channel from which toextract elevation data from a high-resolution light detectingand ranging (lidar) DEM at the boundaries of an ENVISAT-1ASAR-derived extent of the January 2, 2003 flood event on theRiver Alzette (G.D. of Luxembourg). A moving average filteris applied to the data to smooth the decreasing elevation withriver distance downstream.The SAR-based model estimates a smoothed linear trend of water levels using multiple regression analysis with X  and Y   coordinates as the independent variables H  = aX  + bY   + c (1)where H  estimated water height in m above sea level (asl); X  map coordinate X  ( m ) ; Y   map coordinate Y   ( m ) ; a , b , c regression coefficients.The final vertical RMSE accuracy is 41 cm when validatingthe SAR model with field data. In contrast, the previouslycalibrated unsteady 1-D Hydrologic Engineering Centers RiverAnalysis System (HEC-RAS) model [9] leads to an RMSE of 13 cm. This demonstrates that SAR data can be used to estimatea reliable flood water line but indicates that the hydraulic modeloutperforms the SAR-based modeling approach. The higherperformance of the HEC-RAS model is to be expected as ittakes account of processes such as backwater effects, whichdominate flow behavior locally and are neglected in the SAR-based modeling approach.The objectives of this paper are adapted from the conclusionsin [3].1) Combine a triangular irregular network (TIN)-basedmethod and the multiple regression model.2) Address the challenges of using X  and Y   coordinates asindependent inputs to the multiple regression analysis inthe case of a meandering river reach, for example.3) Investigate the pretreatment of the data, which may im-prove the results.This paper presents an improvement of the methodology,thereby increasing the accuracy of the results.In order to demonstrate this most objectively, a Regressionand Elevation-based Flood Information eXtraction model(REFIX) is developed for a well-documented flood of theRiver Alzette and subsequently transferred to a reach of theRiver Mosel which experienced a medium-sized flood onFebruary 28, 1997 recorded by RADARSAT in HH mode.III. S TUDY A REA AND A VAILABLE D ATA  A. River Alzette Study Site The study site at the River Alzette is located downstreamof Luxembourg City between the gauging stations at Steinseland Mersch (Fig. 1). It is characterized by a relatively largeand flat floodplain. The villages along this river stretch havebeen subject to frequent flooding in the past two decades. Theinvestigated flood event of January 2, 2003 has a peak dischargeof approximately 69 m 3 · s − 1 . The reach is 10-km long and hasan average floodplain width of approximately 300 m. It hasa basin size of approximately 1175 km 2 , an average channeldepth of around 4 m, and an average slope of 0.08%.  B. River Mosel Validation Site The River Mosel has a long history of severe flooding. Themodel validation site at the River Mosel is located betweenThionville, France and Perl, Germany (Fig. 1). It is charac-terized by a very flat floodplain with its villages situated ata slightly higher altitude than covered by the 1997 flood forwhich a peak discharge of around 1450 m 3 · s − 1 was recorded.The reach is 16-km long and has an average width of approx-imately 3 km. It has a basin size of around 35500 km 2 , anaverage channel depth of around 8.5 m, and an average slopeof 0.05%.On both test sites, a water line of the flood events is simulatedby the HEC-RAS model and compared to the REFIX-modeledwater line. Although 1-D models are often criticized for theirinability to modeling dynamic floodplain flows, such a modelmay often perform as well as a more complex 2-D model, andthus, in such a case, no justification based on data exists for theuse of more complex models [10]. Moreover, the initial SAR-derived flood extents are compared with the REFIX-generatedextents.  SCHUMANN et al. : HIGH-RESOLUTION 3-D FLOOD INFORMATION FROM RADAR IMAGERY 1717 Fig. 1. River Alzette study site and River Mosel validation site. The study database comprises the following.For the Alzette reach the following hold.1) A single dual-polarized ENVISAT-1 ASAR image (VV,VH) of the January 2, 2003 event acquired with C-bandat 5.3 GHz and with an incidence angle of 35 ◦ . The imagewas recorded at the time of flood peak, and it has a spatialresolution scaled from an initial 25–30 to 12.5 m usingmultilook filtering.2) Continuous upstream (Steinsel) and downstream(Mersch) discharge measurements.3) In situ measurements of maximum water height(cf. Fig. 6).4) Digital photographs of the event.5) Ninety one GPS reference marks of the maximum floodextent.6) A lidar DEM of the floodplain at a spatial resolution of 2 m and a vertical accuracy of  ± 15 cm, which is used toextract the water heights at the maximum flood extent.7) Seventy four measured cross sections to determinechannel geometry.For the Mosel reach:8) A single RADARSAT-HH SAR image of theFebruary 28, 1997 event acquired with C-band andwith an incidence angle of 28 ◦ . The image wasrecorded close to flood peak at a discharge of around1360 m 3 · s − 1 , and it has a spatial resolution scaled froman initial 30 to 12.5 m.9) A nonflooded aerial photography of 1999.10) Continuous upstream (Uckange) and downstream (Perl)discharge and stage measurements.11) A high-resolution DEM of the floodplain, obtained viaa TIN generation from contour lines [7], at a spatialresolution of 7 m, and a vertical accuracy of  ± 25 cm.12) Sixty six measured cross sections to determine channelgeometry.It should be noted that both the HEC-RAS-simulated waterline and the field-based maximum water heights only serveas validation sets and are not used to derive the regression-estimated water line. In other words, only the SAR image(s)and the DEM(s) are used to derive the water heights neededto estimate the flood water line and to subsequently generate aflood-depth map. For each flood event, the HEC-RAS modelis set up using terrain geometry for cross-section profiling,flood hydrographs, and water stages as upstream and down-stream boundary conditions [9]. For the Alzette River reach,the hydraulic model is calibrated using the ground-surveyedwater levels. For the Mosel River, the calibration is performedusing upstream, intermediate, and downstream gauging station-measuredwaterstages,therebytestingtheHEC-RAScapabilityof routing the flood wave accurately.IV. M ODEL D EVELOPMENT ON THE 2003R IVER A LZETTE F LOOD E VENT  A. Methodology Starting with the ASAR image, the REFIX-based flood-depth mapping approach follows the procedure illustratedin Fig. 2. 1) Image Preprocessing: The ASAR image acquired in dualpolarization mode (VV, VH) is first orthorectified with anRMSE of below two pixels (25 m). Then, image speckle isreduced with a Frost filter (5 × 5) that uses image statistics toremove high-frequency noise while preserving the edges. Thesuperiority of VH polarization over VV polarization for floodmappingisevident[11],giventhefactthatinVVmode,verticalobjects(e.g.,emerginglowvegetation)arehighlighted,whereasin VH mode, the horizontal nature of the smooth floodwatersurface is being reflected.For the REFIX to perform well, it is essential to smoothout as much SAR data uncertainty as possible while keepingenough information on the flood boundaries on which to allowwater height extraction. After applying the edge-preservingfilter, fluctuations between the left and right flood extent, whichare assumed to be at level, remain. It is believed that theregression model smoothes out most of this uncertainty. 2) Flood Boundary Delineation and Water-Line Modeling: For a binary segmentation of the image, a simple but nonethe-less effective and widely used thresholding [12] is applied.The threshold value is determined using GPS marks of themaximum flood extent. After flood boundary delineation, thelidar DEM is used to extract water heights for both riverbanksat each HEC-RAS cross section. If a systematic altitudinaldifference between the left and right riverbank is present,the orthorectification accuracy and/or the flood area extractionmethodneedtobereviewed.Ifallseemsadequate,afloodwater  1718 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 45, NO. 6, JUNE 2007 Fig. 2. REFIX methodology. line can generally be estimated using a regression model (2)with the dependent variable being the water height extractedfrom the DEM and the independent variable the downstreamriver distance H  = a · d + b (2)where H  estimated water height ( m asl); a slope of the regression (m/m), which of course changeswith stream (bed) characteristics; d downstream river distance ( m ); b intercept ( m ).It should be noted that for special cases under consideration,the model given by (2) may require certain adaptation (i.e.,using piecewise linear or nonlinear regression analysis) to com-plex flow processes and interactions with local infrastructure inorder to generate a more accurate flood water line, as will beillustrated by the River Mosel case study. 3) Integration of the Regression Model With a TIN:a) Simple regression model: The advantage of simple lin-earregressionasopposedtomultiplelinearregression[3]isthatthe former, by using the downstream distance instead of the X  and Y   image coordinates,accounts forandadaptstoallchangesin river orientation. However, when there is a significant changein riverbed slope or/and the presence of hydraulic structures,piecewise linear or nonlinear regression analysis is necessaryto obtain a more accurate flood water line. b) Water height TIN: As in this paper, the analysis is donewithout the ASAR pixel coordinates (cf. [3]), the generationof a 3-D water height map (or rather 2.5-D [13] with heightbeing a function of  x and y coordinates) from the estimatedflood water line requires the creation of a TIN mesh. For eachriver cross section, the water height is calculated through (1),andthecrosssectionsareusedasbreaklines,withthewaterlinebeing linearly interpolated between successive cross sections.The advantage of the TIN is that in this study approach, thewater level is kept horizontal on each cross section, while atthe same time, changes in river flow directions are respected,if needed.Finally, to obtain the actual flood depths, the lidar DEMneeds to be subtracted from the water height TIN.  B. Model Development Results This part summarizes the application of the REFIX modelto the Alzette flood event while illustrating the results andhighlighting strengths and shortcomings. A general discussionwill follow after validation of the model on the Mosel flood.After orthorectification and filtering of the ASAR image, theflood area is extracted using the thresholding method (Fig. 3).The boundary of the ASAR-derived flood map matches theposition of the GPS marks and agrees generally well withdigital photography of the event.Binary flood classification uncertainty of the SAR flood mapcan result either in over- or underdetection of flood. Overdetec-tion is due to remaining speckle and relief shadow classified asflooded, whereas underdetection is the result of local roughnesscaused by, for example, strong winds, or water masking factorssuch as emerging or strong vegetation and buildings. Before theREFIX model is applied, areas of shadow may be reclassified asnonflooded according to slope and azimuth angle [8]. Accord-ing to the study in [14], underdetection due to individual trees,  SCHUMANN et al. : HIGH-RESOLUTION 3-D FLOOD INFORMATION FROM RADAR IMAGERY 1719 Fig. 3. 2003 Alzette river flood extracted from the ASAR-VH image. forests, and buildings could be identified using a nonfloodedaerial photograph.After extraction of the water heights at the left and rightflood extent, a scatter plot shows in general good agreementbetween the left and right riverbank heights (Fig. 4). It shouldbe noted that a positional adjustment was performed on thesrcinal PRI image to eliminate most distortions. The georefer-encing adjustment method used, which shifts the image until anacceptable agreement between left and right flood extent waterlevel is found, is described in detail by Schumann et al. [15].However, even after position fine tuning, there is still a consid-erable disagreement between the left and right riverbanks (upto 1.7 to 2.9 m) at two locations (around 4000 and 7500 m).At these locations, there is a sudden change in terrain slopeon the right floodplain, which results from a positional errorof the ASAR-derived flood extent. It should be noted that thegeoreferencing error of the ASAR image is one to two pixels.This initial georectification error (i.e., a positional error)combined with the rather coarse spatial resolution of the ASARimage (scaled to 12.5 m) will worsen this effect.Due to the coarse spatial resolution of the ASAR image,it is thought impossible to fine tune image georeferencing tothe extent that the disagreement in heights between riverbanks Fig. 4. Scatter plot of the DEM-derived water levels for both riverbanks(Alzette River).Fig. 5. Comparison between in situ water levels (in centimeters asl), theHEC-RAS-simulated water line, and the REFIX-modeled water line. N.B.: Theposition of each in situ high water mark is shown in Fig. 6. would disappear. Moreover, it is expected that the linear re-gression model, which is applied to the mean altitude of theleft and right riverbanks, is able to smooth out these heightdifferences.The model equation used to estimate the water heights thatenable the generation of the water height TIN is of the form H  = − 0 . 00089 · d + 225 . 74 . (3)Validating the estimated water line (dashed line on graphin Fig. 4) with well-distributed field data gives an RMSE of 18 cm. The REFIX water line when compared to that simulatedwith the HEC-RAS model has an RMSE of 16 cm (Fig. 5).Thus, it can be said that the proposed REFIX model is capableof generating a reliable water line for a given flood event andthat the linear model chosen is appropriate.As can be seen in Fig. 5, the HEC-RAS model is able tocompute specific energy losses in the vicinity of hydraulicstructures such as bridges that may occur along the river reach(e.g., the sudden nonlinear fall in water height at around 6000 mat the location of a hydraulic structure). Where such differencesoccur, the HEC-RAS model, of course, outperforms the REFIX
Related Search
Similar documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks