ASAR DERIVED FLOOD MAPS FOR FLOOD PROPAGATION MODEL CALIBRATION Matgen P.1, Henry J.-B.2, Pappenberger F.3, Pfister L.1, de Fraipont P.2 1

CREBS, Centre de Recherche Public - Gabriel Lippmann, 162a av. de la Faïencerie, L-1511 Luxembourg, Tel. +352 47 02 61 405, [email protected]

2

Service Régional de Traitement d’Image et de Télédétection, Pôle API, Bd Sébastien Brandt, BP 10413, F-67412 Illkirch Cedex, Tel. +33 (0)3 90 24 46 47, [email protected]

3

Dept of Environmental Science, Institute of Environmental and Natural Sciences, Lancaster University, Lancaster LA1 4YQ, UK, Tel. +44 (0)1524 65201, [email protected]

The proposed method compares the EO-derived instantaneous observation of the flood extent with the results from flood simulation models. A fuzzy rule based calibration approach within the Generalized Likelihood Uncertainty Estimation (GLUE) procedure allows evaluating uncertainties and inconsistencies in both model parameterization and flood extent extraction. 2. TEST SITE AND DATABASE A flood prone area of the river Alzette (Luxemburg), downstream of Luxembourg-City and with a length of

70

Photographs

80

Envisat Overpass

90

High Water Marks

100 ERS-2 Overpass

Hydraulic models allow the anticipation of hazardous situations and thus potentially constitute a key application of remotely sensed information. Since SAR imagery presents obvious advantages over optical instruments, especially in flood management applications, it is the aim of this paper to investigate the utility of SAR derived flood extent maps for retrieving the distributed conveyance parameters in one-dimensional flood routing models. Generally, these models are calibrated with in situ measurements. The latter are rarely available at a high density and the cost of extensive surveying campaigns often hampers robust calibrations of distributed models.

Acquired during the rising limb and at the peak discharge respectively, two ERS-2 SAR and Envisat ASAR (in Alternating Polarization mode with VV and VH polarizations) scenes cover the flooded area at two distinct stages of the event (Fig. 1).

s -1 )

1. INTRODUCTION

The evidence of a well-documented medium sized flood event in January 2003, with an estimated return period of 5 years, was used in this study. The available database comprises pre-flood and flood SAR images, continuous discharge measurements upstream and downstream of the river reach, surveyed high water marks and GPS control points of the maximum flood extent. A set of photographs taken during the flooding event is also available.

3

During flood events Earth Observation (EO) Synthetic Aperture Radar (SAR) is able to provide valuable information on the spatial extent of inundations. With its multi-polarisation capacity, the ENVISAT ASAR instrument is now able to propose reliable data for maximum flood extent retrieval. These extents can be useful in calibrating and/or evaluating flood propagation models and their uncertainties. Models are currently calibrated with in situ measurements such as high water marks, photographs and flow velocities. Since these are point measurements, it may sometimes be difficult to relate this information to a highly spatially dependent phenomenon: flooding. It is shown that flood propagation model calibration and uncertainties evaluation can greatly benefit from EO derived information.

approximately 10 km, is selected as test site for this research. The floodplain, with its surrounding villages, was exposed to severe flooding in the past.

Discharge (m

ABSTRACT

60 50 40 30 20 10 0 0

12

24

36

48

60

72

Tim e (hours )

Fig. 1 – Upstream hydrograph for the 2003 flood event and corresponding satellite overpasses and field surveys. 3. FLOOD MAPPING WITH ASAR Currently, flood mapping using space-borne instruments is one of the most relevant methods to get quickly a map of the event. Several tools are available depending on

the data used, and the ASAR instrument on-board of ENVISAT brings new pertinent solutions [1]. 3.1. Mapping principle Due to specular reflection on water surfaces and the resulting low signal return, flood mapping using SAR data may appear quite straightforward. However, in the transient shallow water zone between the flooded and the non-flooded part of the floodplain (with protruding vegetation producing increased signal returns), the radar signal only gradually increases, thus making the accurate delineation of the flood boundary more difficult. Hence, a rather arbitrary choice of a backscattering threshold value is made during the automatic binary segmentation of flooded/unflooded areas. Wind roughening is also a well known effect that eventually has to be taken into account. Despite these limitations, the use of SAR imagery compares favourably with other remote sensing systems [2]. In a case study, reference [3] found that, despite the aforementioned limitations, the SAR segmentation algorithm classified 70 % of the SAR shoreline within 20 m of the shoreline as derived from aerial photographic data. Using ASAR data, ref [4] shows that alternative polarization (AP) mode is more suitable for flood mapping, especially by using like- and cross-polarization (Fig. 2). This leads to an accordance of 85% between Landsat TM derived maps and ASAR AP (HH/HV) derived maps. And again, the high sensitivity of VV polarization to wind-induced water surface conditions is strongly highlighted. 3.2. Introducing uncertainty A rather simplistic threshold approach was used in this study to obtain the inundation maps from the two avail-

able radar scenes. Therefore, profiles of pixel values at several cross sections of the river floodplain are drawn and confronted with the GPS control points of the maximum lateral flood extent. Thus, threshold values of radar backscattering are determined and used to classify the radar image as “flooded” and “non-flooded” respectively. As this raw classification method does not provide entirely satisfying results, three probability classes were defined reflecting our lack of knowledge about the “real” flood extent in the whole area. At each cross section the coordinates of the extension with high, intermediate and low probability of flooding are defined. Therefore, a fuzzy performance measure is defined that at each cross section reflects the uncertainties in the maximum flood extent derived from the radar data. The performance measures, their derivation and theory are documented in detail in ref [5]. The resulting classified image (Fig. 3) shows several flooded areas that are not directly connected to the riverbed. These patterns can be explained by surface runoff, groundwater resurgence or misclassified pixels of the SAR imagery. The latter can be partially removed after confronting the classified image with photographs taken during the event. A further evaluation of the flood map can be achieved by combining the resulting image with a high resolution digital elevation model in order to derive the water elevation at each cross section. By comparing the profile of the water line with the elevation of the river bed, doubtful values can be outlined and eventually be removed from the reference data used during calibration. 4. MODEL CALIBRATION It is commonly accepted that river flow becomes twodimensional once the riverbank bursts its banks. As 2D flood propagation models become more readily avail-

(a) VV polarization (b) VH polarization Fig. 2 – ASAR alternating polarization data acquired during the Alzette river 2003 flood

reach, the friction slope is set to the average channel slope. Assuming normal flow, the Manning equation then allows calculating for each time step the water height which is used as downstream boundary condition.

450 Pixel on Radar Scene

400 350

Pixel Value

300 250

Lo w

150

Int ermed iat e Hig h

100

High Medium Low

Nowadays, connections between 1D hydraulic models and GIS allow for the accurate 2D mapping of simulated inundations. Hence, the comparison of these modelled flood extents with remotely sensed flood areas has become straightforward. However, the needed export of the model results into GIS makes that this remains a very time consuming task.

200

50 0 0

200

400

600

800

Dis tance (m )

Fig. 3 – Spectral profile across a river transect able, the calibration based on flood maps derived from earth observation data become more popular. Despite the obvious limitations in 1D model formulation, it has been shown that in some river reaches simple 1D models and the more sophisticated 2D models performed equally well [6]. Computational advantages therefore suggest that 1D model should be used whenever the topography of the floodplain allows considering the 1D river flow hypotheses. 4.1. Flood propagation modelling The widely used 1D HEC-RAS model is used in this study. This model requires a minimal amount of input data and computer resources and is thus very easy to use. The required input data includes • the topographical description of adjacent river and floodplain cross sections, • the dimensions of hydraulic structures, • the boundary conditions at the upstream and downstream end of the river reach. The description of the flood propagation is based on the Manning-Strickler formula. Thus, three roughness values (one for the channel and two for the floodplains) have to be calibrated to minimise the difference between simulations and observations. This equation is used by nearly all 1D and quasi 2D flood propagation models to estimate empirically the friction slope [7]. Downstream of Luxembourg-city, the HEC-RAS model was set up using 74 cross-sections that describe the channel and floodplain geometry. These data are extracted in a Geographical Information System (GIS) based on a high resolution, high accuracy DEM. The latter was obtained by combining the data describing the ground surveyed cross-sections and the floodplain data obtained by airborne laser altimetry. The inflow hydrograph of the January 2003 flood event constitutes the upstream boundary. At the downstream end of the river

As a matter of fact, in this study, where a large number of computational runs were undertaken, the distributed 2D inundation data are therefore treated to become compatible with 1D model calibration and evaluation. Therefore, the x and y coordinates at the intersection of the digitised flood boundary and each of the river cross sections considered in the model formulation, are extracted in the GIS environment. Next, for each model run and at each cross section, the distance between the simulated and the radar derived flood extent is computed and used to determine the likelihood of the underlying parameter set. Point measurements of stage and travel time do not need to be transformed prior to 1D model calibration and evaluation. 4.2. Calibration method Flood propagation model calibration can be realised with water depths recorded during the event or the maximum flood extent derived from aerial photography or satellite images. This evidence is compared to modelled state variables (e.g. extent or water depths) through an objective function. This approach can be refined by taking into account the uncertainties associated to both model parameters and flood extent map. This requires a model performance analysis using the GLUE (Generalized Likelihood Uncertainty Estimation) methodology [8]. The procedure is a Bayesian Monte Carlo based technique and is recommended for inundation modelling, because it rejects the concept of optimal models in favour of multiple behavioural models (principle of parameter equifinality). It enables an accurate evaluation, quantified with a performance measure (PM), of the most pertinent parameters sets according to all available data. In this study, the GLUE prediction limits are conditional probabilities, related to the simulated flood extent at each river cross section. Their definitions are conditioned by the choice of the model and the errors in both radar and ground based data.

5. RESULTS AND DISCUSSION The model is run 22000 times with randomly chosen roughness coefficients (from a uniform distribution between 0.001 and 0.2) Numerical instabilities caused by almost half of these parameter sets lead to their rejection by the model itself. These instabilities may have many possible origins [7]. Finally, 11608 parameter sets remain for further analysis. For each run, different PMs are calculated. The dotty plots in Fig. 4 represent projections of the parameter space into 1 dimension (roughness parameter on x-axis and performance measure on y-axis). Each dot stands for a single simulation, i.e for the performance achieved by a single parameter set. Each row is associated with one of the 3 parameters considered in the hydraulic model: channel roughness, left and right floodplain roughness. These plots are presented for the three performance measures that were considered in this study: high water marks (HW), flood boundaries derived from ERS SAR and ENVISAT ASAR. Obviously, a large number of parameter sets almost perform equally well. The model performance mainly depends on the channel roughness coefficients, whereas the sensitivity is limited considering the two floodplain friction parameters. Depending on the choice of the channel roughness, good fits can be achieved in the Envisat 02/01/2003 22h18

HW & Envisat

95 % quantile

Envisat

ERS-2

Channel

Right Floodplain Left Floodplain

Fig. 4 – Dotty plots of parameter distributions for the different PMs whole range of the sampled parameter values. In particular, a maximum likelihood value of 1 could be obtained for many different roughness parameters, with the measured high water marks. The dotty plots also show that a considerable range of performance measures are produced inside the sampled parameter range. Some of these parameter sets produce output that has to be considered to be non-behavioural, i.e. the response deviates so far from the observations that the model cannot be considered as an adequate representation of the system. The parameter spaces of the simulations that meet the user-defined behaviourability criteria have ERS-2 02/01/2003 11h00

Observed Flood Envisat

a

High Water

HW & ERS

HW

HW & ERS & Envisat

Observed Flood ERS-2

b

HW & Envisat

HW & ERS

HW

HW & ERS & Envisat

5 % quantile

Fig. 5 – Uncertainty maps computed on Envisat (a) and ERS (b) acquisition times, based on each available EO and ground data

their channel roughness parameter constrained to its lower range from the initial sampling limits. The total number of behavioural simulations for each objective is given in tab. 1. The dotty plots associated to each one of these objectives show the consistencies of the results obtained with flood areas derived from Envisat ASAR and the surveyed high water marks i.e. the plausible parameter range with the best PMs almost remains the same using these two objectives. However, owing to changing roughness values with increasing flow velocities, a slight shift of behavioural parameters can be observed with the ERS data obtained at an earlier stage of the flood event. With only 286 behavioural parameter sets remaining, the most important reduction of parameter sets is achieved with the HW criterion. As these data are the most reliable, a relatively high threshold PM value can be chosen to discriminate between behavioural and non-behavioural models. The fuzziness of the radar data hampers the use of higher threshold PM values. Data used HW Envisat ERS HW & Envisat HW & ERS HW & ERS &Envisat

Total 286 3926 4542 212 143 140

Tab. 1 – Behavioural simulations for different combinations of control data (EO and ground based) The effect of combining different datasets is shown in tab. 1. Only the parameter sets that meet the behaviourability criteria are retained in the final sample. Clearly, the number of behavioural simulations is considerably reduced. More than half of the 286 selected parameter sets are rejected based on the additional criteria, and finally, only 140 among the initial 22000 runs are considered as behavioural. The ERS-derived extent constrains the model response most. This is not surprising as the ERS image was acquired during the rising limb, preceding by several hours the peak discharge, whereas the Envisat image was acquired almost at peak discharge. Hence, the cor-

b.

responding flood map is somewhat redundant to the high water marks and the resulting parameter constrain is not noteworthy. The progressive constraining of the model predictions by incorporating the additional radar data is also illustrated by fig. 6, with the distance between the boundary limits of the 5% and 95% percentile flood maps continuously narrowing (Fig. 5). Simulations constrained with all data show smaller behaviour variability than with the only ground data. Major uncertainties remain upstream. 6. SENSITIVITY ANALYSIS In this study, the Envisat ASAR amplitude image allows delineating the flood boundaries at peak discharge. Therefore, the threshold value, that was determined with the GPS control points, provides the reference flood map (high probability class of flooding). By increasing and decreasing this threshold value by 20 %, the effect of choice of the threshold on the selection of optimal model calibration parameter sets will be investigated. Recent studies have already outlined the need for a careful choice of measures of fit to be optimized during the calibration process. The scope of this study is to assess the calibration sensitivity toward the threshold values applied during a binary segmentation of flooded/unflooded areas. Practitioners in the field of SAR image treatment know that this task is quite far from being straightforward, especially when no ground control points are available. During this sensitivity analysis an optimum model parameter set is found that allows producing the inundation map with a minimal percentage of misclassified pixels in the alluvial plain (area in error). The resulting error surfaces clearly show that the calibrated channel roughness parameter gradually shifts toward higher values when the threshold is increased (Fig. 6). The increase or decrease of the effective roughness parameter compensates the over- and underestimation of the real flood extent. Meanwhile, the model has also been calibrated with a set of surveyed high water marks. In the HEC-RAS model formulation, a channel roughness of 0.065 s/m1/3 gives the best match with the recorded

c.

a.

Fig. 6 – Error surfaces for 3 different threshold values: high probability–20% (a), high probability (b) and high probability+20% (c)

water levels. Thus, the optimisation process with flood maps derived with a threshold value determined with the GPS control points gives the same pair of optimised model parameters (error surface in the middle). However, the error surfaces also show that different parameters were calibrated when the threshold value was slightly changed (±20%). In these cases, the calibrated models largely over- and underestimated the observed water levels. These results prove that model calibration with flood maps obtained with the available SAR imagery is a valuable approach. On the other hand, if the SAR data are to be used as stand-alone utility in model calibration, this research showed that the robustness of this method still has to be improved and, to date, the flood extraction method has to be chosen very carefully in order not to considerably bias the modelling results. It can also be shown that whatever the chosen threshold value is, the inundation maps obtained with the best performing parameter set are very close to Envisat’s observation. Thus, it is interesting to note that a close match can be achieved even if evidence exists that the threshold value has been badly chosen. A small error committed at the moment of preparing the reference data propagates down the whole chain of the modelling process and cannot be compensated at a later stage. 7. CONCLUSIONS This study shows that pertinent information can be derived from EO data for flood propagation model calibration. The same set of parameters was retained when calibrating the model with the SAR data and the ground data respectively. The main difference between both calibration methodologies can be related to the increased fuzziness of earth observation data that leads to larger prediction uncertainties. SAR observations of the flood maximum do not allow any constrain of the model parameters, due to the redundancy with high water marks. However, in the case of sparse ground data, the improvement could become significant. Because of the high density of sample measurements within the Alzette floodplain, the most effective constrain is achieved with EO derived flood extent obtained several hours before peak flow. This observation stresses the ad equation of the time sample of the different control data (EO and ground based) with the schedule of the flood. More generally, the obtained parameter sets have to be validated on different flood intensities before being used in risk studies, which necessarily need extrapolation. Therefore, it will be interesting to investigate the effects

of additional datasets on other flood events: constraining the parameter sets or rejecting all simulations which would prove that a calibrated parameter set is only valuable for a particular order of magnitude of inflow. On large study sites with sparse point data sets, uncertainty reduction with EO radar derived information appears to be a very promising approach. It may, however, be necessary to account for the dataset’ fuzziness. This uncertainty analysis is far from being completed and future research will focus on calibrating distributed parameter sets with radar data. 8. ACKNOWLEDGEMENTS This study is supported by the ‘Ministère Luxembourgeois de la Culture, de l’Enseignement Supérieur et de la Recherche’. The authors would like to thank Mr. Gilbert Schleich at the ‘Service National de la Protection Civile’ of the Grand-Duchy of Luxembourg for providing some of the data used in this study and Mr. Jean-Paul Abadie at the French Space Agency (CNES) for supporting this research project. 9. REFERENCES 1. Henry J.-B., 2004, Systèmes d’information spatiaux pour la gestion du risque d’inondation de plaine, Ph. D. Thesis, Louis Pasteur University, Strasbourg, 192 p. (in french) 2. Biggin D.S., Blyth K., 1996, A comparison of ERS-1 satellite radar and aerial photography for river flood mapping, Journal of the Chartered Institute of Water Engineers and Managers, 10, pp. 59-64. 3. Horritt M.S., Mason D.C., Luckman A.J., 2001, Flood boundary delineation from Synthetic Aperture Radar imagery using a statistical active contour model, International Journal of Remote Sensing, 22(13), pp. 2489-2507. 4. Henry J.-B., Chastanet P., Fellah K., Desnos Y.-L., 2003, ENVISAT Multi-Polarised ASAR Data for flood mapping, Proceedings of IGARSS'03, Toulouse, France. 5. Matgen P., Henry J.-B., Pappenberger F., Pfister L., de Fraipont P. and Hoffmann L., 2004, Uncertainty in calibrating flood propagation models with flood boundaries derived from synthetic aperture radar imagery, Proceedings of the XXth ISPRS Congress, Istanbul (Turkey), 12-23 July 6. Horritt M.S., Bates P.D., 2002. Evaluation of 1D and 2D numerical models for predicting river flood inundation,. Journal of Hydrology, 268, pp. 87-99. 7. Pappenberger F., Beven K.J., Blazkova S., Uncertainty in the calibration of effective roughness parameters in HECRAS using inundation data, Journal of Hydrology, in press. 8. Beven K.J., Binley A.M., 1992, The future of distributed models: model calibration and predictive uncertainty, Hydrological Processes, Vol. 6, pp. 279-298.