An operational procedure for rapid flood risk assessment 1 in Europe 2 3

14 The development of methods for rapid flood mapping and risk assessment is a key step to increase 15 the usefulness of flood early warning systems, and is crucial for effective emergency response 16 and flood impact mitigation. Currently, flood early warning systems rarely include real–time 17 components to assess potential impacts generated by forecast flood events. To overcome this 18 limitation, this work describes the benchmarking of an operational procedure for rapid flood risk 19 assessment based on predictions issued by the European Flood Awareness System (EFAS). Daily 20 streamflow forecasts produced for major European river networks are translated into event-based 21 flood hazard maps using a large map catalogue derived from high-resolution hydrodynamic 22 simulations. Flood hazard maps are then combined with exposure and vulnerability information, 23 and the impacts of the forecast flood events are evaluated in terms of flood prone areas, economic 24 damage and affected population, infrastructures and cities. 25 An extensive testing of the operational procedure is carried out by analysing the catastrophic 26 floods of May 2014 in Bosnia-Herzegovina, Croatia and Serbia. The reliability of the flood 27 mapping methodology is tested against satellite-based and report-based flood extent data, while 28 modelled estimates of economic damage and affected population are compared against ground29 based estimations. Finally, we evaluate the skill of risk estimates derived from EFAS flood 30 forecasts with different lead times and combinations of probabilistic forecasts. Results show the 31 potential of the real-time operational procedure in helping emergency response and management. 32

increase preparedness of authorities and population, thus helping to reduce negative impacts 40 (Pappenberger et al., 2015). Early warning is particularly important for cross-border river basins 41 where cooperation between authorities of different countries may require more time in order to 42 inform and coordinate actions (Thielen et al., 2009). 43 In this context, the European Commission has developed the European Flood Awareness System 44 (EFAS) which provides operational flood predictions in major European rivers as part of the 45 Copernicus Emergency Management Services. The service has been fully operational since 2012 46 and is available to hydro-meteorological services with responsibility for flood warning, EU civil 47 protection, and their networks. 48 While EWS are routinely used to predict flood magnitude, there is still a gap in their ability to 49 translate flood forecasts into risk forecasts -in other words, to evaluate the possible consequences 50 generated by forecasted events (e.g. flood-prone areas, affected population, flood damages and 51 losses), given their probability of occurrence. Generally, flood impacts are evaluated considering 52 reference risk scenarios where a fixed return period is used for all of the area of interest, for 53 instance based on official maps issued by competent authorities (EC, 2007). However, this implies 54 some degree of interpretation to define flood impact and risk in case of a flood forecast. Some The availability of real-time operational systems for assessing potential consequences of 60 forecasted events would be a substantial advance in helping emergency response (Molinari et al.,61 2013), and indeed flood risk forecasts are increasingly being requested by end-users of early 62 warning systems (Emerton et al., 2016;Ward et al., 2015). At a local scale, the joint evaluation 63 of flood probabilities and consequences may not only increase preparedness of emergency 64 services, but also allow cost-benefit considerations for planning and prioritizing response 65 measures (e.g. strengthening flood defences, planning evacuation of people at risk). At a 66 European scale, the possibility to receive prior information on expected flood risk would help the 67 Emergency Response Coordination Centre (ERCC) in prioritizing and coordinating support to 68 national emergency services. 69 In the present paper, we describe a methodology that is designed to meet the needs of EWS users 70 and to overcome the limitations mentioned so far. The methodology translates EFAS flood 71 forecasts into event-based flood hazard maps, and combines hazard, exposure and vulnerability 72 information to produce risk estimations in near real-time. All the components are fully integrated 73 within the EFAS forecasting system, thus providing seamless risk forecasts at European scale. 74 To demonstrate the reliability of the proposed methodology, we perform a detailed assessment 75 focused on the 2014 floods in the Sava River Basin in Southeast Europe. A large dataset for the 76 evaluation of the results has been collected, consisting of observed flood magnitude, flood extent 77 derived from different satellite imagery datasets, and detailed post-event evaluation of flood 78 impacts, economic damage assessment and affected population and infrastructure. 79 The reliability of the flood mapping procedure is first assessed by assuming a "perfect" forecast, 80 where flood magnitude is taken from real observations instead of EFAS predictions. The effect 81 of the failure of flood defences is also taken into account. Subsequently, we test the performance 82 of the operational flood forecasting procedure, to evaluate the influence of different lead-times 83 and combinations of forecast members. 84

2) Methodology
In this Section we describe the three components which comprise the rapid risk assessment 87 procedure: 1) streamflow and flood forecasting; 2) event-based rapid flood hazard mapping 3) 88 impact assessment. Figure 1 shows a conceptual scheme of the steps comprising the methodology. 89 90 91 Figure 1: Conceptual scheme of the rapid risk assessment procedure 92 93 The basic workflow of the procedure is outlined as follows: 94  Every time a new forecast is available, the procedure defines the river sections potentially 95 affected and local flood magnitude, expressed as the return period of the peak discharge; 96  Areas at risk of flooding are identified using a map catalogue, which defines all the flood-97 prone areas for each river section and flood magnitude; these local flood maps are then 98 compared against local flood protection levels and merged to derive event-based hazard maps; 99  Event hazard maps are combined with exposure and vulnerability information to assess 100 affected population, infrastructures and urban areas, and economic damage. 101 102 The described procedure is fully integrated within the existing EFAS forecast analysis chain and 103 operates in near real-time. When a new EFAS hydrological forecast becomes available (Step 1), 104 the risk assessment procedure is activated for those locations where predicted peak discharges 105 exceed the flood protection levels (Step 2). When activated, the execution time depends on the 106 extent and spatial spread of the affected areas over the full forecasting domain. Even in the case  107  of flood events occurring simultaneously in different European countries, the results of the  108  analysis are delivered within one hour after the EFAS forecast runs are finished.  109 The Linking streamflow forecast with inundation mapping is complex because inundation modelling 141 tools are computationally much more demanding than hydrological models used in EWS, which 142 currently prevent a real-time integration of these two components. To overcome this limitation, 143 in this study we have created a catalogue of flood inundation maps, covering all of the EFAS river 144 network and linked to EFAS streamflow forecasts. 145 The hydrological input for creating the map catalogue is derived from the streamflow dataset of 146 the EFAS reference simulation, described in Section 2.1. The information is available for the 147 EFAS river network at 5 km grid spacing for rivers with upstream drainage areas larger than 500 148 km 2 . Since hydrographs simulated in the EFAS reference simulation do not refer to specific return 149 periods, we use a statistical analysis of extreme values to derive peak discharges for every cell of 150 the river network for reference return periods of 10, 20, 50, 100, 200 and 500 years. In addition, 151 we extract flow duration curves from the reference simulation, which are used together with peak 152 discharges to calculate synthetic flood hydrographs (see Alfieri et al., 2014b for a detailed 153 description). 154 The streamflow data are then downscaled to a high-resolution river network (100 m), where 155 reference sections are identified at regular spacing along the stream-wise direction every 5 km. 156 100 m sections are then linked to a section of the 0.1° river network, in order to assign to each 157 section a synthetic discharge hydrograph. Where the coarse-and high-resolution river networks 158 do not overlap, flood points are linked with the closest 0.1° pixel in the upstream direction. Note 159 that there is not a one-to-one correspondence between 5 km and 100 m river sections. In particular, 160 some 5 km sections have no related sections in the 100 m river network, while others can have 161 more than one. Figure  The flood maps related to the same EFAS river section (i.e. pixel of the 5 km river network) are 168 merged together, to identify the areas at risk of flooding due to overflowing from a specific EFAS 169 river section, and archived in the flood map catalogue. The merging is performed separately for 170 each return period, in order to relate flooded areas with the magnitude of the flood event. This step of the procedure provides a rapid estimation of the expected flood hazard, using the 180 database of flood maps described in Section 2.2.1 to translate EFAS discharge forecasts into 181 event-based flood mapping. 182 At each grid cell, we first identify the median of the ensemble forecast given by the latest EFAS 183 prediction, and then select the maximum discharge of the median over the full forecasting period 184 (10 days). This value is compared with the reference long-term climatology to calculate the return 185 period. In this way, the range of ensemble forecasts is taken as a measure of the probability of 186 occurrence, while forecast return periods allow estimation of the magnitude of predicted flood 187 events. Then, predicted streamflow is compared with the local flood protection level, and river 188 grid cells where the protection level is exceeded are considered to activate the impact assessment The land use layer also provides the exposure information to compute direct economic losses in 215 combination with flood hazard variables and flood damage functions, following the approach 216 developed by Huizinga et al. (2007). More specifically, we use a set of normalized damage 217 functions to calculate the damage ratio as a function of water depth, ranging from 0 (no damage) 218 to 1 (maximum damage  Morava and Mlava, in Serbia. 278 Table 1  In our analysis we considered the river network of the Sava River basin, where some of the most 293 affected areas are located and for which detailed information is available from various reports. 294 To evaluate the skill of the flood hazard mapping procedure, we used observed flood magnitudes 295 ( Figure 3) to identify the return period of peak discharges and thus select the appropriate flood 296 maps. In addition, we used the information on flood protection level and dyke failures to select 297 only those river sections where flooding actually occurred, either due to defence failures or 298 exceeding discharge. The resulting flood hazard map is referred to, for the remainder of this paper, 299 as the "reference simulation". Such a procedure excludes the uncertainty due to the hydrological 300 input from the analysis, focusing on the evaluation of the flood hazard mapping approach alone. 301 In other words, the test can be seen as an application of the procedure in the case of a single, 302 deterministic and "perfect" forecast. The resulting inundation map is displayed in Figure 4. 303 It is important to note that a margin of uncertainty remains because of the emergency measures 304 which were taken during the event. In several river sections of the Sava River, the flood defences 305 were actually able to withstand discharges well above their design value, thanks to timely Despite these numerous available data sources, the evaluation of the simulated flood extent is not 319 straightforward. All of the available images were acquired when the flood was receding (from 19 320 May onwards), while flood peaks were observed between 15-17 May. Therefore, several areas 321 which have been reported as flooded in the available documentation are not included in the 322 detected flood footprints, which results in a significant difference between satellite-detected and 323 reported flood extent from ground surveys (see Table 1). On the other hand, EMS satellite maps 324 are designed to produce a low rate of false positive errors, and can therefore be considered as a 325 "lower limit" for the real flood extent. Finally, it must be considered that, for each country, the 326 available information sources report different extents of flooded area, as can be seen in Table 1. 327 In order to take these issues into account, we first compare the total simulated and reported flood 328 extent at country level, calculating over-estimation or under-estimation rates against all the 329 available reported data.

Evaluation of forecast-based flood hazard maps
To evaluate the overall performance of forecast-based flood hazard mapping, we considered the 349 EFAS forecasts issued on 12 and 13 May for the Sava river basin, i.e. immediately before the first 350 flood events occurred on 14 May. We first applied the standard procedure described in Section 2 351 above, to derive peak discharges, estimated return periods and flood maps using the median of 352 the EFAS ensemble forecasts. To provide a more complete overview of risk scenarios, we also 353 applied the procedure considering the 25 and 75 percentiles of discharge in the ensemble 354 forecasts. As a first step, we evaluate EFAS forecasts by comparing forecast and observed return 355 periods. Then, forecast-based flood hazard maps are evaluated against the reference simulation, 356 comparing the river sectors and the urban areas (or municipalities) at risk of flooding. Note that 357 we selected the reference simulation as the benchmark because it represents the best result 358 achievable in case of a perfect forecast. Conversely, we did not carry out a comparison against 359 observation-based flood maps, because they incorporate the effect of defence failures or 360 strengthening, which could only be considered as hypothetical scenarios in forecast-based maps. 361

363
Inundation maps derived from the reference simulation and flood forecasts have been used to 364 compute flood impacts in terms of number of affected people, affected major towns and cities, 365 and economic damage. 366 The results are compared with the available impact estimations both at national and local level. 367 For Serbia and Bosnia-Herzegovina, the national figures reported in Table 1 refer to the total  368 impact given by river floods, landslides and pluvial floods, and so cannot be directly compared 369 with methodology results. Therefore, the comparison has been done only for Croatia and for a 370 number of municipalities (e.g. Obrenovac in Serbia) where impacts can be attributed to river 371 flooding alone. 372 The figures for affected population computed with the reference simulation, are also useful to test 373 the reliability of the population map used as the exposure dataset. Similarly, damage estimations 374 provide an indication of the reliability of depth-damage curves for the study area. 375 As was done for the flood hazard maps, forecast-based risk estimations are evaluated against the 376 results from the reference simulation, comparing both population and damage figures. Note that 377 other variables produced by the operational procedure (e.g. roads affected, extent of flooded urban 378 and agricultural areas) could not be tested due to lack of observed data, and therefore are not 379 discussed here. To add a further term of comparison, affected population has been computed using 380 Copernicus EMS flood footprints. 381

4) Results and discussions
The results of the evaluation exercise are shown and discussed separately for each component of 384 the procedure. 385 386   387  Table 3 reports the observed flood extent data from available sources, and the simulated extent 388 derived from the reference simulation (i.e. the mapping procedure applied to discharge 389 observations). The ratios between simulations and observations are also included. As expected, the simulated flood extent is significantly larger, in all cases, than the satellite extent 406 (see Table 3), given the delay between the times of flood peak and image acquisition, as 407 mentioned in Section 3.2. Flood extent indicated in the ICPDR and ISRBC Report is also 408 consistently lower than values from both simulated and satellite maps. from GeoSerbia satellite maps is smaller than the simulation, but it has to be considered that these 413 maps have the same problem of delayed image acquisition as mentioned for the Copernicus maps. 414

Flood hazard mapping
For Croatia, the flood mapping methodology is largely over-estimating both the satellite-based 415 and reported flood extents. The main reason is that flooding on the left side of Sava was limited 416 due to the reinforcing of river dykes in the area close to the city of Zupanja, which could withstand 417 the reported 500-year return period discharge, despite having been designed for a one in one 418 hundred years event. In fact, all of the left bank of Sava in this area was reported as an area at risk 419 in case of a flood defence failure, and only the emergency measures taken prevented more severe 420 flooding (ICPDR and ISRBC, 2015). Therefore we performed an additional flood simulation 421 excluding any failure on the river's left bank between the Bosna confluence and Zupanja, and in 422 this case we found a total flood extent of 319 km 2 . Even if this estimate still exceeds the reported 423 flood extent (Wikipedia, 2016), it has to be considered that this figure refers only to the Vukovar-424 Srijem county, which was the most affected area, therefore the total affected area in the whole 425 country was probably larger. 426 Regarding Table 4, the scores of the hit ratio (H) indicate that the mapping procedure correctly 427 detected most of the flooded areas, with the partial exception of the lower Sava area. In particular, 428 the vast majority of towns reported to have been flooded are correctly detected by the simulations, 429 with only a few false alarms (e.g. the already mentioned Zupanja). 430 When looking at the results it is important to bear in mind the limitations of the procedure. As 431 mentioned in Section 2.3, the mapping is able to reproduce only maximum flood depths, while 432 the dynamics of the flood event are not taken into account. This means that processes like flood-433 wave attenuation due to inundation occurring upstream, cannot be simulated, and possible flood 434 mitigation measures taken during the event are also not considered. Furthermore, due to the coarse 435 resolution (100 m) of the DEM used, flood simulations do not include small-scale topographic 436 features like minor river channels, dykes and road embankments. performance of impact assessment, we consider only Table 6, because national estimates in Table  444 5 include also people displaced by landslides and pluvial floods not simulated in EFAS. 445 Note that in both Tables we compare simulated impacts with figures for evacuated population  446 because reported estimates of affected population include also people affected by indirect effects 447 such as energy shortage and road blockage. Also, the figures for evacuated population are not 448 equivalent to directly affected population (i.e. whose houses were actually flooded). In some 449 areas, evacuation was taken as a precautionary measure, even if flooding did not eventually occur. As can be seen, differences between results and reported figures are in the order of hundreds, 462 suggesting that the procedure is able to provide a general indication of the impact on population, 463 but with a limited precision where impacts are small, as in the case of Osjek-Baranja county. 464 However, differences are larger for Vukovar-Srijem county in Croatia, and Obrenovac 465 municipality in Serbia. For the former, this is due to over-estimation of flooded areas as discussed 466 in The observed under-estimation should be evaluated considering the limitations of both observed 480 data and damage assessment methodology. On the one hand, available damage functions for 481 Croatia are not specifically designed for the country, as discussed in Section 2.  Table 7 illustrates return periods of peak discharge derived from 12 and 13 May forecasts for the 493 main rivers of the Sava basin, visible in Figure 3. Simulations are compared against values  494 reported by ICPDR and ISRBC (2015). 495 496 River 12/5 25p.

548
As mentioned in Section 1, the availability of a risk forecasting procedure able to transform hazard 549 warning information into effective emergency management (i.e. risk reduction) (Molinari et al.,550 2013), opens the door to a wide number of new applications in emergency management and 551 response. However, to better understand the limitations of such a procedure, as well as its potential 552 for future applications, some considerations have to be made. 553 Firstly, it is important to remember that EFAS is a continental-scale system which is mainly 554 designed to provide additional information and to support the activity of national flood emergency 555 managers. Therefore, the practical use of risk forecasts to activate emergency measures would 556 need to be discussed and coordinated with services and policy-makers at local level. 557 Secondly, the new procedure needs to undergo an accurate uncertainty analysis before risk 558 forecasts can effectively be used for emergency management. While a detailed analysis is beyond 559 the scope of this paper, to this end we have recently begun to evaluate the performance of the 560 procedure for the flood events recorded in the EFAS and Copernicus EMS databases. 561 Another point to consider is the approach chosen to assess flood risk. In the current version of the 562 procedure, we produce a single evaluation based on the ensemble forecast median, to provide a 563 straightforward measure of the flood risk resulting from the overall forecast. A more rigorous 564 approach would require analysis of all relevant flood scenarios resulting from EFAS forecasts, 565 and estimation of their consequences together with the conditional probability of occurrence, 566 given the range of ensemble forecast members and the forecast uncertainty (Apel et al., 2004). 567 While such a framework would enable a cost-benefit analysis of response measures in an explicit 568 manner, it would also require evaluation of the consequences of wrong forecasts, such as missing 569 or under-estimating impending events, or issuing false alarms ( This paper presents the first application of a risk forecasting procedure which is fully integrated 602 within a continental-scale flood early warning system. The procedure has been thoroughly tested 603 in all its components to reproduce the Sava River basin floods in May 2014, and the results 604 highlight the potential of the proposed approach. 605 The rapid flood hazard mapping procedure applied using observed river discharges, was able to 606 identify flood extent and flooded urban areas, while simulated impacts were comparable with 607 observed figures of affected population and economic damage. The evaluation was complicated 608 on the one hand by the scarcity of reported data at local scale, and on the other hand by the 609 considerable differences in impacts reported by different sources, especially regarding flood 610 extent. This is a well-recognised problem in flood risk literature, due to the fact that existing 611 standards for impact data collection and reporting are still rarely applied (Thieken et al., 2016). 612 Therefore, further improvements of impact models will require the availability of impact data 613 complying with international standards (Corbane et al., 2015;IRDR, 2015). 614 The use of EFAS ensemble forecasts enabled the identification of areas at risk with a lead-time 615 ranging from one to four days, and the correct evaluation of the magnitude of flood impacts, 616 although with some inevitable limitations, due to difference between simulated and observed 617 streamflow. When evaluating the outcomes, it is important to remember that, even in case of a 618 risk assessment based on "perfect" forecasts and modelling, simulated impacts will always be 619 different from actual impacts. As we have shown in the test case of the floods in the Sava River 620 basin, unexpected defence failures can occur for flow magnitudes lower than the design-level, 621 thus increasing flood impacts. On the other hand, flood defences might be able to withstand 622 greater discharges than their design-level, and emergency measures can improve the strength of 623 flood defences or create new temporary structures. Therefore, forecast-based risk assessment may 624 be regarded as plausible risk scenarios that can provide valuable information for local, national 625 and international authorities, complementing standard flood warnings. In particular, the explicit 626 quantification of impacts opens the way to more effective use of early warning information in 627 emergency management, enabling the evaluation of costs and benefits of response measures. 628 After a testing phase that started in September 2016, the procedure described in this paper has 629 been fully operational within the EFAS modelling chain, since March 2017. For the immediate 630 future, we plan to test a number of modifications and alternative approaches for the hazard 631 mapping and risk assessment components. For instance, flood hazard maps are now computed 632 using only the median of EFAS ensemble forecasts, but in principle the methodology can also be 633 applied to more ensemble members, in order to take account of (for example) flood scenarios that 634 are less probable but potentially more severe, and to provide a more complete risk evaluation 635 (such as the application described this paper). Furthermore, additional risk scenarios can be 636 produced, by considering the failure of local flood defences, or replacing EFAS flood hazard 637 maps with official hazard maps developed by national authorities, where available. The influence 638 of lead-time on flood predictions may also be assessed, for example by setting a criterion which 639 is based on forecast persistence over a period, to trigger the release of impact forecasts. All of 640 these alternatives will be tested in collaboration with the community of EFAS users, in order to 641 maximize the value of the information provided, and to avoid information overload, which can 642 be difficult to manage in emergency situations. 643 A further promising application which is being tested is the use of inundation forecasting to 644 activate rapid flood mapping from satellites, exploiting the European Commission's Copernicus 645 Emergency Mapping Service. 646 Finally, the proposed procedure will also be incorporated into the Global Flood Awareness 647 System (GloFAS), thereby enabling a near real-time flood risk alert system at a global scale. 648 649 been updated, their geographic location (in some cases, protection values have been modified 663 only at specific locations along the river), previous and updated values, and the source of 664 information. Protection values are expressed in terms of years of an event's return period. 665 In addition to the modifications in Table S1, further updates of the EFAS database are planned,  666 using the global flood protection layer FloPROS (Scussolini et al., 2016