Combined fluvial and pluvial urban flood hazard analysis: concept development and application to Can Tho city, Mekong Delta, Vietnam

Abstract

Many urban areas experience both fluvial and pluvial floods, because locations next to rivers are preferred settlement areas and the predominantly sealed urban surface prevents infiltration and facilitates surface inundation. The latter problem is enhanced in cities with insufficient or non-existent sewer systems. While there are a number of approaches to analyse either a fluvial or pluvial flood hazard, studies of a combined fluvial and pluvial flood hazard are hardly available. Thus this study aims to analyse a fluvial and a pluvial flood hazard individually, but also to develop a method for the analysis of a combined pluvial and fluvial flood hazard. This combined fluvial-pluvial flood hazard analysis is performed taking Can Tho city, the largest city in the Vietnamese part of the Mekong Delta, as an example. In this tropical environment the annual monsoon triggered floods of the Mekong River, which can coincide with heavy local convective precipitation events, causing both fluvial and pluvial flooding at the same time. The fluvial flood hazard was estimated with a copula-based bivariate extreme value statistic for the gauge Kratie at the upper boundary of the Mekong Delta and a large-scale hydrodynamic model of the Mekong Delta. This provided the boundaries for 2-dimensional hydrodynamic inundation simulation for Can Tho city. The pluvial hazard was estimated by a peak-over-threshold frequency estimation based on local rain gauge data and a stochastic rainstorm generator. Inundation for all flood scenarios was simulated by a 2-dimensional hydrodynamic model implemented on a Graphics Processing Unit (GPU) for time-efficient flood propagation modelling. The combined fluvial-pluvial flood scenarios were derived by adding rainstorms to the fluvial flood events during the highest fluvial water levels. The probabilities of occurrence of the combined events were determined assuming independence of the two flood types and taking the seasonality and probability of coincidence into account. All hazards - fluvial, pluvial and combined - were accompanied by an uncertainty estimation taking into account the natural variability of the flood events. This resulted in probabilistic flood hazard maps showing the maximum inundation depths for a selected set of probabilities of occurrence, with maps showing the expectation (median) and the uncertainty by percentile maps. The results are critically discussed and their usage in flood risk management are outlined.

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/ doi:10.5194/nhess-16-941-2016 Author(s) 2016. CC Attribution 3.0 License.

Heiko Apel1, Oriol Martnez Trepat1, Nguyen Nghia Hung2, Do Thi Chinh1,2, Bruno Merz1, and Nguyen Viet Dung1,2

1GFZ German Research Centre for Geosciences, Potsdam, Germany

2Southern Institute of Water Resources Research SIWRR, Ho Chi Minh City, Vietnam

Correspondence to: Heiko Apel ([email protected])

Received: 12 June 2015 Published in Nat. Hazards Earth Syst. Sci. Discuss.: 26 August 2015 Revised: 29 February 2016 Accepted: 21 March 2016 Published: 13 April 2016

Abstract. Many urban areas experience both uvial and plu-vial oods, because locations next to rivers are preferred settlement areas and the predominantly sealed urban surface prevents inltration and facilitates surface inundation. The latter problem is enhanced in cities with insufcient or non-existent sewer systems. While there are a number of approaches to analyse either a uvial or pluvial ood hazard, studies of a combined uvial and pluvial ood hazard are hardly available. Thus this study aims to analyse a uvial and a pluvial ood hazard individually, but also to develop a method for the analysis of a combined plu-vial and uvial ood hazard. This combined uvialpluvial ood hazard analysis is performed taking Can Tho city, the largest city in the Vietnamese part of the Mekong Delta, as an example. In this tropical environment the annual monsoon triggered oods of the Mekong River, which can coincide with heavy local convective precipitation events, causing both uvial and pluvial ooding at the same time. The uvial ood hazard was estimated with a copula-based bivariate extreme value statistic for the gauge Kratie at the upper boundary of the Mekong Delta and a large-scale hydrodynamic model of the Mekong Delta. This provided the boundaries for 2-dimensional hydrodynamic inundation simulation for Can Tho city. The pluvial hazard was estimated by a peak-over-threshold frequency estimation based on local rain gauge data and a stochastic rainstorm generator. Inundation for all ood scenarios was simulated by a 2-dimensional hydrodynamic model implemented on a Graphics Processing Unit (GPU) for time-efcient ood propagation modelling. The combined uvialpluvial ood scenarios were derived

by adding rainstorms to the uvial ood events during the highest uvial water levels. The probabilities of occurrence of the combined events were determined assuming independence of the two ood types and taking the seasonality and probability of coincidence into account. All hazards uvial, pluvial and combined were accompanied by an uncertainty estimation taking into account the natural variability of the ood events. This resulted in probabilistic ood hazard maps showing the maximum inundation depths for a selected set of probabilities of occurrence, with maps showing the expectation (median) and the uncertainty by percentile maps. The results are critically discussed and their usage in ood risk management are outlined.

1 Introduction

Floods are among the most damaging natural disasters, as statistics of the insurance industry regularly show (MunichRe, 2015). A large share of the damages caused by oods occurs in urban areas, where most of the assets and population are concentrated. Rapid urban development increases the exposure to oods on the one hand, while climate change including sea level rise increases the hazard, especially in coastal and estuarine regions (Merz et al., 2010). Assessing ood hazards and risks, preparing effective ood mitigation measures, and utilizing ood benets at the same time have thus become an even more vital task in water resource planning and ood management.

Published by Copernicus Publications on behalf of the European Geosciences Union.

Combined uvial and pluvial urban ood hazard analysis: concept development and application to Can Tho city, Mekong Delta, Vietnam

942 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

In urban areas different ood pathways typically exist. Because locations next to rivers are preferred settlement locations and because the predominantly sealed urban surface prevents inltration into the ground, both uvial and pluvial oods typically occur in urban areas. This situation is exacerbated in coastal cities or cities in river deltas, where the low topographical gradient limits the drainage capacity and tidal inuences and storm surges can cause additional uvial inundation. Thus oods of different origins are a main hazard for the worlds river deltas (Renaud et al., 2013; Syvitski and Higgins, 2012; Syvitski, 2008). Particularly prone to this combination of ood types are Asian megacities with their extraordinary urban growth, their location at major rivers under a pronounced monsoonal ood regime, and heavy precipitation caused by cyclones (Chan et al., 2012). Another important aspect, particularly in Asia, is the low coordination of water resources and ood management of riparian countries in a river basin, where changes in ood hazards can be caused by upstream water management measures, as shown for the Mekong by Kuenzer et al. (2013a).

Despite the large impact of urban inundation by different pathways, ood hazard and risk assessments are typically restricted to a single ood type. A growing number of u-vial ood hazard analyses can be found in literature, particularly in the last decade with growing computational capabilities and a particular shift towards probabilistic hazard and risk assessments. Examples for thematic and methodological advances in uvial hazard or risk assessments are Apel et al. (2004, 2008), Merz and Thieken (2005), Hall et al. (2005), McMillan and Brasington (2008), Vorogushyn et al. (2010), Arrighi et al. (2013), de Bruijn et al. (2014), and Falter et al. (2015).

Pluvial ood hazard analyses are less frequently found in the literature. This can be explained by the inherent complexity of dening and quantifying these events and by the large spatial heterogeneity of the inundation causing rainfall events. Note that this statement refers to hazard analyses, i.e. quantitatively dening rainfall events and their ood impact in a spatially explicit manner including an estimation of the probability of occurrence of the rainfall events.Traditionally, the probability of rainfall events is quantied by intensity-duration-frequency (IDF) curves, which are often established for meteorologically similar regions. However, stochastic, spatially explicit rainfall simulators are also used in order to fully describe possible rainfall intensities and their spatial coverage (Burton et al., 2008; Hundecha et al., 2009; Willems, 2001; Wheater et al., 2005). Studies using synthetic rainfall events associated with probabilities of occurrence for ood inundation modelling for the derivation of pluvial ood hazard maps are, however, not very frequent.Examples of full pluvial ood hazard analyses are Nuswantoro et al. (2014) or Blanc et al. (2012).

Studies considering the combined effects of uvial and pluvial ooding are very rare. In fact only three studies were found. Chen et al. (2010) performed a scenario-based inunda-

tion modelling study considering the combination of uvial and pluvial oods for a small city area in the UK. A uvial ood with a return period of 200 years was with a dyke over-topping scenario and a synthetic dyke breach scenario. These two uvial scenarios were combined with rainfall scenarios of different probabilities of occurrence. The rainfall intensity was assumed to be uniform over the simulation area. However, due to the limited number of uvial scenarios and the absence of an estimation of the joint occurrence of the u-vial and pluvial events, the study is not a hazard analysis, but rather a sensitivity and feasibility study, as the authors acknowledge. Another study in this direction was published by Thompson and Frazier (2014), who analysed the joint occurrence of hurricane-induced coastal ooding with heavy precipitation. But in this study the probabilities of occurrence of coastal and pluvial oods were also not fully dened, as well as the probability of their joint occurrence. Just recently Breinl et al. (2015) published a study where a framework for estimating the probabilities of joint occurrence of uvial and pluvial oods was developed. In their study the authors focus on a catchment-scale analysis, including a multi-site weather generator and a hydrological model to estimate the joint occurrence of pluvial and uvial ood days in the city of Salzburg. The framework explicitly accounts for dependency of the triggering rainfall of the uvial and pluvial oods. Using the stochastic weather generator, long time series of peak discharges and precipitation within the city were generated.By this the joint occurrence of uvial and pluvial ood days could be quantied. However, only the magnitudes in terms of peak discharge and precipitation were quantied, not the consequences. This means that no inundation modelling was included, and thus hazard maps suitable for a risk assessment were not derived. As the core element of the framework of Breinl et al. (2015) is a multi-site weather generator, it requires a solid precipitation database and considerable effort for the construction of the actual weather generator, which is likely to put some restriction on the applicability of the method for data-sparse regions or very large catchments.

Based on these observations this study aims to develop a concept for joint and spatially explicit uvial and pluvial hazard analyses, by which the probabilities of occurrences and the magnitudes for both ood types as well as their combinations are quantied. Additionally, spatially explicit probabilistic hazard maps are derived, depicting the individual as well as the combined hazard. The study also contains an analysis of some natural uncertainty sources, in order to consider the natural variability of the ood triggering events.

The study was developed in a tropical setting, and takes Can Tho city in the Mekong Delta as an example. Can Tho experiences both uvial and pluvial oods, which occur during the monsoon season. Thus both ood types may coincide, although they are independent from each other in the given meteo-hydrological setting. A core motivation for taking Can Tho as an example is that a ood hazard or risk analysis has never been performed for the city. Flood mitigation planning

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 943

Mekong delta

Hau river

Can Tho city

C

a

Hau river

r

e

v

i

r

o

h

n

T

Mekong basin

Figure 1. Sketch of the geographical setting of the study: the whole Mekong basin, the Mekong Delta with the main gauging stations, and Can Tho city. The area shaded in grey in the Can Tho city image shows the modelling domain and the blue dots indicate the uvial boundaries of the hydraulic model.

is, if existing, based on experiences of past oods only. The presented study might therefore be used as a basis for ood mitigation planning for Can Tho city, if the discussed limitations are taken into account. Can Tho city can also be taken as a role model for growing cities in delta regions in the tropics, where uvial oods and heavy convective rains typically occur in the same period. Therefore, the developed method might also serve as a blueprint for ood hazard analyses in the tropics.

2 Study area and data

2.1 Study area

Can Tho city is the largest city and the economic hub of the

Vietnamese part of the Mekong Delta with about 1.2 million inhabitants (in 2011). It is located at the junction of the Hau River (Bassac) and the Can Tho River in the Vietnamese part of the Mekong Delta (Fig. 1). The centre of the city is the heart of the Ninh Kieu district, located directly north of the junction, bordering both Hau and Can Tho Rivers. The study is restricted to this part of Ninh Kieu, because a digital elevation model (DEM) of sufcient resolution was available for this part of the city only. No large-scale ood defences exist in Ninh Kieu. However, within the Vietnam Urban Upgrading Project Can Tho city sub-component (20042014) and Mekong Delta Region Urban Upgrading Project Can Tho city sub-component (20122017) road curbs and house entrances were and are elevated in order to reduce inundation of urban areas surrounding major roads. The distance to the coast is about 80 km, but due to the low topography and river

gradient the tidal signal is still strong in Can Tho, even during high ood stages. Flood events are thus typically a combination of high water levels during the annual oods and a high tide. This results in water levels exceeding the river banks for the time of the semi-diurnal high tides, causing inundations of short duration in the vicinity of the river banks. High water levels in Can Tho causing inundation typically occur late in the ood season from September to November.

Next to the two large rivers a number of channels can be found in and around the city. Inundation can also occur by bank overtopping of these channels. The water levels in these channels at the model domain boundary differ from the main rivers and need to be considered explicitly in a uvial hazard analysis. However, gauge data for these channels do not exist and cannot be derived from the gauge in Can Tho. Thus the whole hydraulic system surrounding Can Tho, i.e. the whole Mekong Delta, has to be considered in order to derive consistent boundary conditions for the hydraulic model. The topography of the city is very low ranging, between 1 m and 2.5 m a.s.l. (Fig. 2). Most parts of Ninh Kieu are urban areas with built-up houses and infrastructure. Housing typically extends to the very river banks and in some cases even exceeds them, with houses partially on stilts in the rivers or channels.

2.2 Data

For Ninh Kieu a DEM with a 15 m resolution derived from topographical maps exists and has been used for hydraulic modelling by Huong and Pathirana (2013).The same DEM was used in this study (Fig. 2, top panel). The simulation do-

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

944 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

Water level records were collected for the main river gauge of Can Tho in the Hau River and a gauging station was installed in the Can Tho River within the Water-related Information System for a Sustainable Development of the Mekong Delta (WISDOM) project (http://www.wisdom.eoc.dlr.de

Web End =http://www.wisdom.eoc. http://www.wisdom.eoc.dlr.de

Web End =dlr.de ) which was datum-referenced to the main gauge. The locations of the gauges are indicated in Fig. 2 (top panel).A comparison of the recorded and modelled water levels for the ood event 2011 with the bank elevation revealed that the bank elevation as determined by the DEM was too high compared to the gauge records. Even for extraordinary high water levels as experienced during the ood event in 2011 the banks as given in the DEM were not overtopped and thus no inundation would occur (Fig. 3, top panel). This indicates a datum error or at least a discrepancy between gauge data and DEM. Thus the DEM datum had to be corrected. This was performed by comparing the water levels recorded for the ood event in 2011 with the elevation of a river bank stretch along the Hau River (orange circle in Fig. 2, top panel) and information about inundation depths from household surveys near the river (Chinh et al., 2016). This comparison suggested that the DEM datum should be corrected by 0.5 m

in order to provide inundation depths comparable to the surveyed depths. This correction war corroborated by surveyed water levels of Takagi et al. (2015). They surveyed water levels relative to the river bank at the same point as above in March 2012. Following their records the bank elevation of the DEM had to be reduced by 0.5 m as well in order to produce a similar picture as in Takagi et al. (2015) compared to the gauge records (Fig. 3, bottom panel). Thus we corrected the datum of the DEM by 0.5 m.

For the pluvial ood hazard analysis, hourly rainfall records from the rainfall gauge at Can Tho airport were obtained. These span a time series of 30 years from 1982 to 2012. Figure 4 shows the time series along with the annual maxima. The highest recorded hourly rainfall in this period was 87 mm h1. The gure shows that the rainfall is distinctively seasonal with high rainfall amounts during the monsoon season (MayOctober) and little rainfall during the remaining time of the year. Next to the annual maximum a large number of additional high rainfall events are obvious in the data series. This calls for a peak-over-threshold (POT) approach for the frequency analysis of heavy rainfall events, which will be explained further in Sect. 3.2.

3 Methods

3.1 Fluvial hazard analysis

The uvial hazard analysis comprises two basic components:

1. a 2-D hydraulic model for the simulation of the ood propagation in the study area

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

Figure 2. Top panel: central part of Can Tho city with DEM overlay. The yellow circles indicate gauge locations, the orange circle the location of the river bank, where Takagi et al. (2015) measured water levels which were used for correcting the datum of the DEM. Bottom panel: central part of Can Tho with hydraulic roughness map (Mannings n values) derived from land use classication.

main consists of 243 264 grid cells. In Ninh Kieu a sewer

system exists, consisting of open channels and underground pipes. However, the sewer system is reportedly not capable of draining ood waters, neither pluvial nor uvial, due to its limited capacity. Additionally, the sewer system is not well maintained and malfunctioning in many parts (Huong and Pathirana, 2013), reducing the ood sewer capacity even further. Thus in this study the sewer system is neglected in the hydraulic ood simulations.

A land use classication based on high resolution Rapid-Eye satellite data with 5 m resolution was derived by the German Aerospace Center (DLR) using the algorithms presented in Huth et al. (2012). The 5 m resolution data was resampled to 15 m to be in line with the DEM. Based on the land use classes, ve roughness classes were derived (Fig. 2). For this study we adopt an urban porosity approach of assigning high roughness values to dense built-up areas. This aims at reproducing the hydraulic resistance of buildings to urban inundation ows.

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 945

Water levels flood event 2011

2.5

Xuanh Khan

Gauge Can Tho Model Mean bank elev.

Bank elev. 0.5m

2

Elevation [m a.s.l.]

1.5

1

0.5

29/09/11 30/09/11

DEM correction validation: peaks should be just below bank, cf. figure 6 in Takagi (2015)

2

Xuanh Khan Gauge Can Tho

Mean bank elevation 0.5m

Elevation [m a.s.l.]

1

0

1

10/03/12 11/03/12 12/03/12 13/03/12

Figure 3. Correction of datum error of DEM. Top panel: comparison of bank elevation Ninh Kieu extracted from DEM and monitored and modelled water levels for the ood event 2011. Bottom panel: comparison of bank elevation extracted from corrected DEM with measured water levels and eld campaign data published in Takagi et al. (2015).

@hi,j

qi1,jx qi,jx + qi1,jy qi,jy[Delta1]xy , (2)

where t is time; [Delta1]t is the time step; q is the specic ow per unit width; i, j are cell indices; how is the ow depth between cells, i.e. the difference between the maximum water elevation (surface elevation + water depth) and the maximum

surface elevation between two adjacent cells; g is the acceleration of gravity; n is Mannings roughness coefcient; Sf is the friction slope; h is the water depth, and [Delta1]xy is the size of the square cells. The ow is represented by qx in x direction(horizontal in grid space) and qy in y direction (vertical in grid space). Cell index i enumerates the cells in x direction, j in y direction. The continuity Eq. (2) for each grid cell i, j is thus a mass balance of ows in and out of the cell in both x and y direction.

The internal model time step is determined analogously to the CourantFriedrichsLewy criterion:

[Delta1]tmax =

@t =

Figure 4. Hourly rainfall Can Tho gauge 19822012 with annual maxima marked by red circles.

2. an extreme value statistic for the estimation of the probability of occurrence of ood events of certain magnitudes.

These two components are described in detail in the following sections.

3.1.1 Hydraulic model

The hydraulic model for Ninh Kieu was developed on the basis of the 2-dimensional formulation of the shallow water equations solving the momentum (Eq. 1) and continuity (Eq. 2) equations numerically on a grid as described in Bates et al. (2010):

qt+[Delta1]t =

qt ghow[Delta1]tSf

1 + ghow[Delta1]t + n2 |q|

h

10

3 ow

[Delta1]xy

pghow , (3)

where denotes an empirical coefcient reducing the time step determined by the CourantFriedrichsLewy criterion to ensure model stability and is set to 0.8.

The implemented model is grid-based and thus able to simulate ood propagation directly on the basis of the provided DEM. Rivers and channels are also modelled in 2-D in contrast to the LISFLOOD-FP implementation of Bates et al. (2010). Flows are calculated between cells, based on the water depths and surface elevations of the neighbouring cells. Thus routing precipitation causing overland ow is quite simple and straightforward: at each time step the spatially distributed precipitation (given in mm depth) is added to the inundation depths of the affected grid cells. By this the

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

(1)

[parenrightBigg]

946 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

inundation depths of the grid cells are updated each time step and consequently the hydraulic gradients change, resulting in ow. This simple and effective way of routing rainfall is another benet of grid-based inundation models, particularly for the present study.

Hydraulic models based on simplied shallow water equations have been proven to be an adequate choice for the simulation of ood propagation and inundation in many ood hazard studies, for example Apel et al. (2009), de Almeida and Bates (2013), Dimitriadis et al. (2016), Fewtrell et al. (2011).For the typical shallow inundation depths in uvial and plu-vial oods (except ash oods) with subcritical ow conditions, the simplied models yield comparable results to models based on the full shallow water equations, but with significantly lower computation cost. Thus the use of such a model in the presented study is well justied, particularly considering the slow inundation propagation and low surface gradients.

In order to speed up simulation runtime, the model code was implemented on a Graphics Processing Unit (NVIDIA Tesla C1060 GPU) using Portland CUDA Fortran. The NVIDIA Tesla C1060 GPU card contains 240 processor cores enabling a highly parallelized execution of the model code in the spatial domain. The parallelization of the model is handled internally and automatically on a single GPU card and does not require user intervention for domain decomposition and synchronization. This makes the use of GPU for general purpose computing highly attractive. However, code adaptation and reimplementation in CUDA Fortran requires investing time and effort.

Hydraulic models need to be calibrated, mainly by adjusting the hydraulic roughness. This was attempted by simulating the large inundation events from 2011. Two large u-vial events occurred between 26 September and 2 October, and between 24 and 31 October 2011, with peak water levels at 29 September and 29 October. In the aftermath of the events a household survey was conducted in which the inundation depth was estimated by the house owners (Chinh et al., 2016). Additionally, maps showing the inundation extent based on TerraSAR-X Stripmap satellite images with 2.75 m spatial resolution covering particular days in the ood season were provided by the German Aerospace Center (DLR).Unfortunately the satellite maps do not cover the peak water levels exactly. For the September event, an image one day after the peak water level was available, while for the event in October no image was taken during the entire event. A closer inspection of the mapped inundation extent for 30 September revealed that only a few inundated areas are classied within the city limits, in contrast to the actual observed ood.The detected inundation areas are concentrated on the main roads and wider channels. This behaviour is caused by reection artefacts of the radar signals from building walls and the close distance of the buildings in streets causing over-radiation of pixels and thus erroneous classication. These intrinsic errors of radar images used for ood mapping in ur-

Figure 5. Plausibility check of the hydraulic model: maximum inundation depth of the simulated ood event 2011 and surveyed water depth inside inundated houses (dots).

ban areas are reported by Kuenzer et al. (2013b), taking Can Tho city and province as an example. Due to the temporal mismatch and the underestimation of the inundation extent in densely settled areas, the satellite-based ood-extent maps could not be used for the calibration/validation of the hydraulic model, unlike in previous large-scale hydrodynamic modelling studies in the Mekong Delta (Dung et al., 2011;Manh et al., 2014).

In the household interviews the owners provided information of the inundation depths inside their houses, i.e. water depth relative to the ground oor level. However as the ground oor levels vary greatly in relation to the street level, depending on construction details and precautionary measures of the owners, the recorded inundation depths in the house cannot be compared to simulated inundation depths based on the DEM at hand. It was only possible to check the model simulations for plausibility using the information from the survey as indicators for being inundated or not, i.e. as a rough and incomplete estimate for the inundation extent. Figure 5 shows the result of the simulation of the 2011 ood in terms of maximum inundation depths and the reported in-house water depths. The colour coding of the simulation and survey data can serve as an indication of a plausible model performance in terms of inundation depths, but given the conceptual mismatch between the simulated inundation depths and the surveyed water depths inside the houses, this is just an indication. However, it can denitively be stated that at every location of inundated houses the model also simulated an inundation, which is the most robust statement regarding model performance that can be made. But as by far not all ooded households were surveyed, the model performance in terms of overall inundation extent cannot be assessed conclusively with the present data. The described plausibility check does not provide any quantitative measure of model performance, but rather a positive qualitative feedback about the applicability of the model for the given purpose. Given the

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 947

quality of the DEM and validation data at hand, the performance of the 2-D hydraulic model was assessed as plausible and suitable for the purpose of the study.

3.1.2 Boundary conditions of the hydraulic model

In order to simulate meaningful ood propagation, boundary conditions have to be provided for the simulation domain along the Hau and Can Tho Rivers as well as for the channels entering the simulation domain on the western border. These boundary conditions have to be associated with probabilities of occurrence in order to obtain a hazard analysis. As gauge data sufcient for an extreme value statistic were not available, particularly for the channels, the boundary conditions for the Can Tho model were derived from a large-scale hydraulic model for the whole Mekong Delta (Dung et al., 2011). The boundary conditions for the Mekong Delta model were derived for the gauge Kratie in Cambodia denoting the upper boundary of the Mekong Delta. A copula-based bivariate extreme value statistic is performed, using annual maximum discharge Qmax and ood volume V and standardized hydrographs, which are scaled to the maximum discharge and ood volume for given annual probabilities of non-exceedance. The method is extensively described in Dung et al. (2015), thus it is not repeated here. From the different statistical models tested in Dung et al. (2015), the GumbelHougaard copula with the marginals of Qmax and V described by log-normal distributions is used. Synthetic scenarios were derived for annual probabilities of nonexceedance of p = 0.5, 0.8, 0.9, 0.95, 0.98, 0.99, whereas for

each probability level, 140 combinations of annual maximum discharge Qmax and ood volume V were drawn in a Monte-

Carlo process from the curves shown in Fig. 6 in Dung et al. (2015). This resulted in 6 140 = 840 ood scenarios al

together, which were simulated by the large-scale hydraulic model.

From these scenarios the simulated water levels along the Hau River, Can Tho River and the channel were taken as boundary conditions for the 2-D hydraulic model of the study area. To save computational time, the 2-D simulation period was set to six days around the maximum water level of each ood scenario at the Hau River. This period is equivalent to the length of the spring tides in the Mekong Delta (Wass-mann et al., 2004). This limitation can also be justied from a technical point of view, because sensitivity runs showed that the maximum inundation depth and extent do not change with longer simulation periods. This is due to the distinct tidal inuence on the water levels, which exceed the bank levels during maximum water levels, typically during spring tides and for short periods only. During low tides the water can drain back into the river. Figure 7 shows the Monte Carlo derived boundary conditions for all probability levels for both river and channel.

The inundation simulations resulting from these boundary conditions resulted in 140 inundation maps for Can Tho

POT frequency analysis for hourly rainfall events for

Can Tho: 1. Threshold determination2. GP t to POT series

140 MC simulation of synthetic rainstorms with random location of storm centre for each p-level

Temporal disaggregation of hourly storm events to minute time scale

2-D pluvial inundation simulation for Can Tho

Derivation of pluvial ood hazard maps

Rainfall intensity for six probability levels of exceedance (0.5, 0.2, 0.1, 0.05, 0.02, 0.01)

Bivariate frequency analysis of Qams and V for Kratie

Six probability levels of simultaneous exceedance (0.5, 0.2, 0.1, 0.05, 0.02, 0.01)

140 MC samples of Qams and V pairs for each probability level

Derivation of synthetic ood hydrographs from Qams and

V pairs and standard ood hydrographs

Large-scale inundation simulation for the whole Mekong Delta for 6x140 synthetic ood events

Boundary conditions for 2-D Can Tho model from large scale model

2-D uvial inundation simulation for Can Tho

Derivation of uvial ood hazard maps

Combined 2-D uvial & pluvial inundation simulation for Can Tho

Derivation of uvialpluvial ood hazard maps

Figure 6. Scheme of the uvial, pluvial and combined hazard analysis.

showing maximum inundation depths for every six probability levels. For each probability level, the maximum inundation maps were evaluated per grid cell, calculating the median maximum inundation depth and the 5 and 95 % quantiles, as performed by Vorogushyn et al. (2010). These probabilistic hazard maps are the nal product of the proposed uvial hazard analysis. The median hazard maps serve as the expectation (in a statistical sense) for the ood hazard of a given probability of occurrence, while the 5 and 95 % percentile maps quantify the uncertainty, stemming from the natural variability of the inundation processes (natural or aleatory uncertainty; Merz and Thieken, 2005). Figure 6 summarizes the different steps of the uvial ood hazard analysis in a ow chart.

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

948 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

p = 0.5

Hau river Channel

p = 0.8

Range max. WL: 1.711.9 m a.s.l.

2

2

H [m a.s.l.]

H [m a.s.l.]

1

1

Range max. WL: 1.711.82 m a.s.l.

0 0 20 40 60 80 100 120 140

0 0 20 40 60 80 100 120 140

p = 0.9

Range max. WL: 1.761.95 m a.s.l.

0 0 20 40 60 80 100 120 140

p = 0.95

Range max. WL: 1.862.01 m a.s.l.

2

2

H [m a.s.l.]

H [m a.s.l.]

1

1

0 0 20 40 60 80 100 120 140

p = 0.98

Range max. WL: 1.862.05 m a.s.l.

0 0 20 40 60 80 100 120 140

p = 0.99

Range max. WL: 1.92.11 m a.s.l.

2

2

H [m a.s.l.]

H [m a.s.l.]

1

1

0 0 20 40 60 80 100 120 140

Time [h]

Time [h]

Figure 7. Boundary conditions for the uvial hazard analysis: Hau and Can Tho Rivers (blue) and the channels entering the modelling domain from theWest (green). p denotes the annual probabilities of non-exceedance. Simulation time is six days around the maximum water level in the Hau River. For each p level 140 simulations were performed. The spread of the boundary time series denotes the variability stemming from the bivariate extreme value statistic, which increases with higher annual probabilities of non-exceedance.

3.2 Pluvial hazard analysis

The pluvial hazard analysis comprises three components: a statistical analysis of extreme rainfall events in Can Tho, the derivation of spatially distributed synthetic rainstorms, and a 2-D hydraulic model. The hydraulic model for the pluvial hazard analysis is the same as described in Sect. 3.1.1, but with different boundary conditions. In order to exclude u-vial inundation the river boundaries were set to low water levels never exceeding the river banks and allowing for outow from the simulation domain. The boundaries of the model domain were set by adding spatially distributed rainfall volumes derived from the statistical analysis and synthetic rainstorms to each grid cell with a time step of 60 s. The statistical analysis and derivation of synthetic rainstorms are summarized in Fig. 6 and described in detail in the following sections.

3.2.1 Rainfall extreme value statistics

The basis for this analysis is the 30-year time series of hourly rainfall recorded at Can Tho airport shown in Fig. 4. The presence of many strong rainfall events besides the annual maxima calls for a peak-over-threshold (POT) analysis. POT enables a full exploitation of the information content of the time series including all hazardous storm events of the monsoon season, and thus a realistic estimation of rainfall probabilities. An inspection of the rainfall time series illustrates the very nature of the rainfall events in the region, which are convective tropical rainstorms. Almost all heavy rainfall events have a duration of 2 h, with the majority of events

lasting 1 h. This suggests that the storm duration for this

analysis can be realistically xed to maximum 1 h. In further analysis we even consider outbursts of shorter duration by

non-uniform rainfall intensities within the maximum duration of 1 h (cf. Sect. 3.2.2). This is also in concordance with local experiences. The construction of IDF curves is thus not required; a POT analysis of hourly rainfall is sufcient to describe the rainfall regime.

In order to exploit the full information content of the data series, it is assumed that the events lasting two hours in the time series are in fact one hour events extending over the change of the hour. Thus the sums of two hour events are assumed as an event lasting one hour and are used for the POT analysis. It is assumed that two rainfall events are independent if the temporal distance between them is at least 24 h. In order to determine the optimal threshold for the POT analysis several tests were conducted: a stability plot of the generalized Pareto distribution (GPD), which is used for the POT analysis, a Pareto quantile plot, a mean residual life plot and a KolmogorovSmirnov test. In addition, the stability of selected quantiles of the GP was tested depending on different thresholds. The results of the tests are shown in Fig. 8. All the tests indicate that a threshold of 18 or 19 mm h1 is an appropriate choice. Thus the threshold was xed at 18 mm h1.

Figure 9 shows the resulting maximum likelihood t of the GP to the POT series with a threshold of 18 mm h1 along with the empirical distribution, Gringorten plotting positions of the POT series, and a t of a GEV distribution to the annual maxima series for comparison. The POT quantiles are much higher compared to the annual maximum analysis depicting the likelihood of several extreme rainfall events per year (in the mean 12 rainfall events per year exceeding the threshold of 18 mm h1). The tted GP is the basis for the derivation of synthetic storm events and their associated probabilities.

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 949

0.2 Stability plot GP parameters

14

10 Pareto quantile plot

Tail index (shape) Scale

log (R [mm h ])

0.1

12

1819 mm h

10

0

10

-0.1

10 15 20 25 308

10 0 2 4 6 8 10

Threshold value [mm h ]

Theoretical quantiles

14 Mean residual life plot

KolmogorovSmirnow test

0.06

1

13

12

0.05

0.8

11

Mean excess

KS test statistic

p-value (significance)

10

19 mm h

9

0.04

0.6

8

7

0.03

0.4

6

5

18 mm h

18 mm h

4 0 10 20 30 40 50 60 70 80

0.02

10 15 20 25 300.2

Threshold value [mm h ]

Threshold [mm h ]

130

120

p=0.5

p=0.8 p=0.9

p=0.95

p=0.98

p=0.99

110

Quantile values [mm h ]

100

90

80

70

60

50 10 12 14 16 18 20 22 24 26 28 30

POT rainfall theshold [mm h ]

Figure 8. Threshold determination for the POT rainfall frequency analysis. Upper panel: analytical tests. For the KolmogorovSmirnov test the lowest KS statistic and the highest signicance indicates the threshold. For the other tests the threshold is approximated by the point where trajectories deviate from an approximate linear behaviour. The red lines indicate the threshold determined by the individual tests. Lower panel: empirical test. Quantiles (p = annual probability of non-exceedance) estimated with a maximum likelihood tted generalized

Pareto distribution and different POT thresholds. The red line indicates the threshold above which the estimated quantiles are approximately stable.

3.2.2 Synthetic storm events

In order to simulate inundation caused by heavy rainfalls the statistically derived rainfall intensities recorded at the rain gauge are translated into a spatial rainfall elds, based on the following assumptions:

1. the rainfall events do not cover the whole study area with uniform intensity

2. the extent of the convective rainfall cell is assumed to be circular

3. the intensity of the rainfall of the convective cell is highest at its centre and decreasing to the border

4. the intensity within the circular extent is distributed according to a Gaussian bell

5. the intensity along the border of the convective cell is 1/10 of the maximum intensity

6. the diameter of the storm cell increases with intensity

7. the location of the storm cell is stationary during the event duration of 1 h.

Assumption 1 is based on local observations and has been conrmed by regional meteorologists in personal communication. Assumption 2 is also conrmed by meteorologists dealing with radar rainfall observations in the region. A similar assumption was also taken by Nuswantoro et al. (2014) for a storm generator for Jakarta in Indonesia, which has similar rainfall characteristics as the South of Vietnam. Assumptions 3 to 6 are also based on observations in the area and have also been used by Nuswantoro et al. (2014) for Jakarta. Assumption 4 is reasonable for tropical rainstorm cells. This approach of describing the rainfall intensity has been adopted from the weather generator of Willems (2001).For assumption 6 the extent of the large storm events was estimated at 8 km, based on detailed meteorological sim-

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

950 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

80 GP distribution fit to POT series precipitation Can Tho. threshold = 18 mm h

Probability density

(scaled to bin counts)

60

40

20

0 0 10 20 30 40 50 60 70 80

1

0.8

0.6

0.4

Fitted generalized pareto CDF

Empirical CDF 95 % conf. int.

0.2

0 0 10 20 30 40 50 60 70 80

Exceedances (R [mm h ] thresh)

0.99 (100)

Annual probability of nonexceedance

(return period T [a])

0.98 ( 50)

0.95 ( 20)

0.9 ( 10)

0.8 ( 5)

0.5 ( 2)

0 ( 1)

Empirical POTGringorten plotting positions GP (thresh=18)

GEV (AMS)

20 40 60 80 100 120 140

R [mm h ]

Figure 9. Fit of the GP to the hourly rainfall time series of Can Tho with POT threshold = 18 mm h

1. The lower gure shows the tted GP distribution with 95 % uncertainty bounds, the empirical POT distribution and the plotting positions for the POT series, along with a GEV distribution tted to the annual maxima series for comparison.

ulations of two large storm events in Can Tho (both around 80 mm h1) by Huong and Pathirana (2013). Table 1 lists the assumed relation of annual probability of non-exceedance p, rainfall intensity R(p), and extent of the storm cells at the full width at tenth of maximum (FWTM). The functional relation between FWTM and R(p) was empirically derived as FWTM = R(p) 90, based on the simulated storm events in

Huong and Pathirana (2013). Figure 10 shows two synthetic storm events resulting from this procedure for annual probabilities of non-exceedance of 0.5 and 0.99.

In order to compensate for the negligence of the movement of convective storm cells (assumption 7), the pluvial hazard analysis was embedded in a Monte Carlo analysis randomizing the location of the storm centres. Through this procedure, the random nature of the location of the maximum rainfall is captured, but the effect of moving rainfall

Table 1. Annual probability of non-exceedance p, associated rainfall intensity R(p) as estimated with the tted generalized Pareto distribution, and the assumed extent (FWTM, Full Width at Tenth of Maximum) of the synthetic rainstorm associated to the intensity.

p 0.5 0.8 0.9 0.95 0.98 0.99

R(p) [mm h

1 ] 59.79 72.54 82.43 92.55 106.20 116.90

FWTM [m] 5380 6528 7419 8329 9563 10 521

cells can also be mimicked. Analogously to the uvial hazard analysis 140 Monte Carlo runs with random selection of storm centres over the simulation domain for each probability level were conducted. In order to test the stability of the Monte Carlo procedure in terms of spatial distribution of maximum rainfall intensity, the maximum rainfall of the syn-

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 951

thetic rainstorms with random storm centres was evaluated at 25 evenly spaced grid cells over the simulation domain. For each grid cell the maximum rainfall intensity was extracted from the 140 synthetic storms for all probability levels. The reasoning behind this analysis is that a stable MC simulation would yield the rainfall intensity quantied by the extreme value statistics (cf. Sect. 3.2.1) at least once for the given probability level. As the box plots in Fig. 11 show, not all of the 25 grid cells received 100 % of the potential maximum rainfall, i.e. some underestimation occurs at some grid cells. However, for all but the annual probability of nonexceedance p of 0.5, this negative bias is very small (< 1 %) and thus negligible, particularly considering the uncertainties of the analysis. However, even for p = 0.5, the median of the

maximum rainfall of the grid cells is 98.4 % of the rainfall given by the GP distribution, i.e. the bias is 1.6 % only,

which can also be considered acceptable.

For simulating the inundation caused by the synthetic storm events, the events have to be disaggregated, i.e. the temporal resolution of one hour needs to be reduced to time steps appropriate for the hydraulic model. Instead of a simple uniform disaggregation we opted for a disaggregation with a distinct precipitation peak, which is more realistic for heavy convective rains. Thus the hourly intensity of the synthetic storm events was disaggregated into 60 time steps of 1 min by a normal distribution with [notdef] = 30 min and = 5 min. This

resulted in maximum rainfall intensities at 30 min after precipitation start and a concentration of the bulk of the precipitation in the 30 min surrounding the peak. This temporal dis-aggregation was applied to every pixel of the synthetic storm events. For the inundation simulation the rainfall amount of the disaggregated storm events was directly added to every pixel covered by the synthetic storms as surface water. The surface water was then routed by the 2-D hydraulic model with an overall simulation time of 3 h to allow for redistribution of the water after the end of the storm event. The resulting 140 maps of maximum inundation depths per probability level were then evaluated to create probabilistic ood hazard maps. This procedure is identical to the uvial hazard analysis (cf. Sect. 3.1.2).

3.3 Combined hazard analysis

The essential question for a combined uvial and pluvial hazard analysis is the question of dependency. At this point a clear distinction between dependence and coincidence is made: dependence infers a causal or functional relationship,i.e. in the given context one ood type would cause or inuence the other, either in its probability of occurrence or its magnitude. Coincidence does not include any relationship as dependent. It is rather dened as the chances that the two ood types occur at the same time. Following these denitions, for this particular study it can be assumed that uvial and pluvial ood events are completely independent from each other. Although they appear during the same season,

Figure 10. Examples of synthetic hourly rainfall events for annual probability of non-exceedance p = 0.5 (left panel) and annual

probability of non-exceedance p = 0.99 (right panel). Note that the

storms have different randomly drawn storm centre locations.

which is actually the prerequisite for a joint hazard analysis, the triggers for uvial and pluvial oods in the city of Can Tho are independent. Fluvial oods are caused by the interplay of run-off generation in the Mekong basin and tidal backwater effects. On the contrary, pluvial ood events are caused by local convective rainfalls, which are strong enough to create urban inundation, but the limited spatial extent and thus rainfall volume does not cause uvial inundation by the Hau River and channels.

The independencies of the ood triggering events have a direct consequence on the calculation of the probabilities of coinciding ood events: in a rst step the joint probability of occurrence within a ood season can be quantied by the product of the individual probabilities of occurrences. For example, a joint occurrence of a uvial ood event with an annual exceedance probability 0.5 and a pluvial ood event of probability 0.5 within the same ood season is 0.25. However, this joint probability of occurrence within the same ood season has to be corrected by the probability that the ood events actually occur at the very same time and cause a combined ood event. This probability is termed probability of coincidence in the given context. Thus the probability of occurrence of joint uvialpluvial ood events is generally calculated as the following:

P ( pl) = P () P (pl) P (co), (4) with P (co) as the probability of coincidence of uvial and pluvial ood events.

In this study P (co) is estimated by the typical length of the ood season in relation to the duration of the uvial ood events. The period of the ood season in which uvial oods in Can Tho typically appear is mid-September to the end of November, i.e. lasting about 76 days (cf. Sect. 2.1). Flood peaks of the Mekong with high water levels last typically about six days (cf. Sect. 3.1). Within these six days, 12 distinct ood peaks, i.e. periods of high water levels, occur due to the semi-diurnal tidal regime. The high water levels differ only slightly, as shown in Fig. 7. Furthermore, sensitivity runs with the hydraulic model have shown that the maximum inundation depths of a combined ood event do not

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

952 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

and pluvial analysis. Figure 6 illustrates the combined hazard analysis in the overall process ow chart.

4 Results and discussion

4.1 Fluvial hazard

Figure 12 shows the probabilistic uvial hazard maps comparing the 5, 50, and 95 % percentile maps for all selected probability levels. The inundation maps clearly show the inundation pathways. Inundation of the built-up area starts around the junction of the city channel and the Can Tho River, at the 90 bend of the city channel in the East of the simulation domain, along the open sewer channel connected to the Can Tho River, and low lying areas around the city channel after the 90 bend of the channel in the east of the simulation domain. From these locations the inundation is progressing into the urban area. A particular distinct feature is the inundation of the 30 Thng 4 road (Road of the30 April) crossing the sewer channel. This is in accordance with observations during inundation events.

The area most susceptible to inundation is the part of the Ci Kh ward between the Hau River and the wide part of the city channel. This area is always inundated except for the 5 % quantile map of p level 0.5. This is also the area of the deepest inundation.

Figure 12 shows that the maximum inundation depths and extents have an increasing annual probability of nonexceedance, as expected. In the median quantile maps, the mean maximum inundation depth, excluding the channel and boundary cells, evaluates to 0.31 m for the p level of 0.5 and increases to 0.36 m for p level 0.99 (cf. Table 2). This rather slight increase in mean inundation depth is accompanied by a substantial increase in the inundated area from 2.37 km2 (p = 0.5) to 5.29 km2 (p = 0.99). In the 50 % quantile maps

almost all the city area is inundated for p level 0.99.

A further noteworthy feature of Fig. 12 is the difference between the quantile maps. There are considerable differences in inundation extent, and for the higher p levels also in inundation depth. This uncertainty is caused by the different pairs of maximum discharge and volume of the synthetic ood events, stemming from the extreme value analysis of the uvial boundary described in Sect. 3.1.2. In the particular context of Can Tho city with prevailing short term ood events, the maximum discharge of the synthetic events and thus the maximum water levels play a dominant role. The maximum water level controls the overbank water level and inundation duration, because with higher maximum water levels the period of overbank water levels during a tidal cycle of six hours is also longer. Contrary to the maximum water level, the overall volume of the whole ood season does not have any impact on the local inundation in Can Tho city.Thus the range of the maximum ood discharges determined by the 140 random MC derived boundaries per p level causes

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

1

MC maximum rainfall of 25 grid cells normalized to R(T)

0.99

0.98

0.97

0.96

0.95

0.94

0.93

0.92 0.5 0.8 0.9 0.95 0.98 0.99

Probability of nonexceedance p

Figure 11. Estimation of the bias introduced by the assumptions taken for the synthetic rainstorm events and the random location of the storm centres in the Monte Carlo analysis. For each probability level the maximum of the 140 synthetic storm events was extracted for 25 grid cells evenly spaced over the simulation domain. The maximum rainfall was normalized to the value given by the generalized Pareto distribution tted to the POT rainfall series. A value of 1 thus indicates zero bias. The box plots show the distribution of the maximum rainfall among the 25 grid cells.

differ signicantly, if the rainstorm event occurs exact at the time of highest water level or if it occurs within 3 h around

high water levels. This means that the sensitive time window for coincidence of a uvial and pluvial event is 12 ood peaks 6 h duration of high water levels = 3 days within a

ood period of 76 days. Considering the average number of rainfall events in the critical uvial ood period, which amounts to 5, P (co) evaluates to ve 3/76 = 0.1974. In

order to account for the unavoidable uncertainty in the assumption taken for this calculation, a value of P (co) = 0.2

was used for calculating the joint uvialpluvial ood probabilities.

For the combined uvialpluvial hazard analysis a set of joint ood events was simulated by combining uvial and pluvial ood events with the same individual probability of occurrence, i.e. uvial ood events of a probability of occurrence of 0.5 were combined with pluvial ood events of a probability of occurrence of 0.5, uvial events of 0.8 occurrence probability with pluvial events of 0.8 occurrence probability, and so forth. A complete permutation of different probability levels for uvial and pluvial ood events was not performed in order to obtain the same number of probability levels as in the individual hazard analysis, but also to keep the required simulation time in manageable limits.

Technically, the combined uvial and pluvial ood hazard maps were derived by simulating the 140 uvial ood scenarios per probability level and adding a synthetic rainstorm event with a random storm centre location at the time of the maximum water level of the uvial scenario boundary. From the 140 combined scenarios per probability level probabilistic hazard maps were derived just as for the uvial

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 953

Figure 12. Probabilistic uvial ood hazard maps for Can Tho showing maximum inundation depths: for the selected p levels (annual probabilities of non-exceedance) the median (50 % quantile) maps and associated 5 and 95 % quantile maps are shown illustrating the uncertainty of the hazard estimation. Maximum inundation depths < 0.02 m are indicated as no inundation.

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

954 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

Table 2. Statistical evaluation of probabilistic hazard maps: mean maximum inundation depths and inundated area for every hazard map (annual probabilities of non-exceedance and percentile maps). For the statistics the channel and the river boundary cells were excluded, as well as cells with maximum inundation depths < 0.02 m.

5 % percentile maps 50 % percentile maps 95 % percentile maps

Annual Mean Inundated Mean Inundated Mean Inundated probability maximum area [km

2

2 ] maximum area [km

of non- inundation inundation inundation exceedance depth [m] depth [m] depth [m]

Fluvial hazard maps

0.5 0.44 1.16 0.31 2.37 0.31 2.740.8 0.31 2.27 0.31 3.33 0.32 3.730.90 0.31 2.83 0.32 3.71 0.33 4.410.95 0.31 3.45 0.33 4.30 0.35 5.100.98 0.32 4.05 0.35 4.98 0.37 5.540.99 0.33 4.34 0.36 5.29 0.40 5.97

Pluvial hazard maps

0.5 0.11 0.25 0.06 3.32 0.09 5.430.8 0.05 1.28 0.07 4.41 0.10 5.910.90 0.05 3.07 0.08 5.12 0.11 6.230.95 0.06 3.87 0.09 5.56 0.11 6.490.98 0.08 4.62 0.11 6.18 0.12 6.770.99 0.09 5.32 0.11 6.43 0.13 6.94

Fluvialpluvial hazard maps

0.95 0.27 2.04 0.20 4.28 0.18 5.980.992 0.23 3.43 0.22 5.53 0.22 6.520.998 0.21 4.65 0.22 6.11 0.24 6.860.9995 0.23 5.25 0.24 6.48 0.28 7.110.99992 0.24 6.04 0.27 6.97 0.31 7.320.99998 0.25 6.43 0.29 7.14 0.35 7.46

the uncertainty in the ood hazard in Can Tho city as shown in Fig. 12.

This uncertainty is also shown in Table 2 comparing the inundated areas between the quantile maps. These underline the visual impression of Fig. 12. The numbers indicate that the differences in inundation extent is higher between the 5 and 50 % quantile maps compared to the 50 and 95 % quantile maps. This effect is less visible in the numbers of the mean maximum inundation depths. The 5 % quantile map of p level 0.5 shows even higher mean maximum inundation depths than the higher quantile maps. This is a statistical effect caused by the relatively few inundated grid cells, which are along the main inundation hot spots, all of them showing comparatively high maximum inundation depths.

4.2 Pluvial hazard

The pluvial hazard maps are shown in the same format as the uvial hazard maps in Fig. 13. On rst glance the distinctively different characteristics of the pluvial hazard maps become obvious. The hazard maps neither show featured ow

2 ] maximum area [km

]

paths nor particular hot spots of inundation. This is caused by the nature of the boundary conditions providing spatially distributed input, in combination with the Monte Carlo selection of random storm centre locations. As a result the plu-vial hazard maps look much more uniform compared to the uvial hazard maps. The inundation patterns show the topo-graphical depressions as given by the DEM rather than actual ow paths. It has to be mentioned that the proposed methodology of random storm locations and variable spatial extent and intensity of the synthetic rainstorms yields characteristically different hazard maps compared to the assumption of spatially uniform coverage of the whole simulation domain. Spatially uniform rainstorms would produce different quantile maps, which would look more uniform over the whole simulation domain. And even more importantly, the inundation depths would be deeper due to the higher rainfall volume given by a uniform coverage compared to the synthetic storm events derived in Sect. 3.2.2.

As shown in Table 2 and visible in Fig. 13, the mean maximum inundation depths of the pluvial maps are much lower for all p levels compared to the uvial hazard maps. For the

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 955

Figure 13. Probabilistic pluvial ood hazard maps for Can Tho showing maximum inundation depths: for the selected p levels (annual probabilities of non-exceedance) the median (50 % quantile) maps and associated 5 and 95 % quantile maps are shown illustrating the uncertainty of the hazard estimation. Maximum inundation depths < 0.02 m are indicated as no inundation.

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

956 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

50 % quantile maps they range between 6 and 11 cm. However, the inundation extent is higher compared to the uvial maps due to the spatially distributed rainstorms. For the 50 % quantile maps the inundation area is always roughly 1 km2 larger.

The uncertainty of the pluvial hazard maps as shown by the 5 and 95 % quantile maps is of particular interest. This uncertainty is considerably larger compared to the uvial hazard maps. As clearly visible in Fig. 13 and quantied in Table 2 the difference between the quantile maps in terms of inundation extent is much larger with the pluvial hazard maps, particularly for the lower p levels. The uncertainty range given by the quantile maps is actually in the same range as between the median of the lowest and highest p level.This holds true for both mean maximum inundation depths and inundation extent, as shown in Table 2. Solely the mean maximum inundation depth between the quantile maps for p level 0.5 does not follow this rule, because of the same statistical effect in calculating the mean maximum inundation depth for the 5 % quantile map. The wide uncertainty range becomes particularly obvious by comparing the 95 % quantile map for p level 0.5 and the 5 % quantile map for p level 0.99, for both mean maximum inundation depths and inundation extent. These maps are almost identical.

This phenomenon is a direct consequence of the random nature of the location of the storm centres. In the absence of dened inundation pathways, this random distribution of the storm centres causes the different inundation patterns and thus the characteristics of the quantile maps. The 95 % quantile map of the p level 0.5 basically shows inundations caused by the centre of a small rainstorm hitting by chance a dened are of the simulation domain. On the contrary, the 5 % quantile map of the p level 0.99 show the inundation caused by the borders of a large storm that actually hit another part of the city. Due to the different storm magnitudes this results in similar quantile maps for different p levels. Assuming that the assumptions for the derivation of the synthetic storm events (cf. Sect. 3.2.2) are reasonable, this is not only a consequence of the proposed method, but also of the very nature of the rainstorms causing inundation, and thus a characteristic feature of pluvial hazard maps under a pre-dominantly convective rainfall regime.

4.3 Combined uvialpluvial hazard

The joint uvialpluvial hazard maps (Fig. 14) combine the characteristics of the individual hazard maps. They show the same distinct inundation hot spots with deeper maximum inundation depths and ow pathways as the uvial maps, complemented by the spatially more uniform but lower inundation depths of the pluvial hazard maps. In general the uvial inundation is dominant, i.e. causes the deeper inundations, while the pluvial inundation plays a role only in places where no uvial inundation occurs. These areas are mainly the yet unnamed industrial park in the An Ha ward at the north-

west corner of the simulation domain, and the Thi Bnh ward west of the 90 bend of the domain boundary.

Interestingly, the maximum inundation depths in the uvial-dominated areas differ only slightly (max. a few mm) from the uvial hazard maps. This indicates that the uvial processes are not signicantly altered by the additional plu-vial input. This is caused by the smaller duration and thus smaller ood water volume of the rainstorms compared to the uvial inundation.

The interplay of uvial and pluvial inundation results in overall larger inundation extents compared to both uvial and pluvial hazard maps alone (cf. Table 2). The mean maximum inundation depths, however, are lower compared to the uvial maps, which is again a statistical effect accounting for the larger number of inundated cells with low inundation depths stemming from the pluvial input.

The percentile maps quantifying the natural uncertainty also show features of both uvial and pluvial uncertainty maps. In the uvial-dominated inundation areas, uncertainty within each probability level is smaller than the difference between the median of all probability levels. The uncertainty within the p levels is skewed, i.e. the difference in inundation extent between the 5 and 50 % percentile maps is larger than between the 50 and 95 % percentile maps. However, for the pluvial-dominated inundation areas the same large uncertainty as for the pluvial hazard maps can be observed.

Sensitivity of coincidence

The core assumption of the combined hazard analysis is the exact coincidence of the rainstorm with the maximum water level in the Hau River. In order to rule out artefacts caused by this assumption, a parallel combined hazard analysis, in which the rainstorm can occur randomly in a time window of 3 h around the maximum water level was performed. A

comparison of the nal probabilistic hazard maps with the exact match maps showed only minor differences in maximum inundation depth and extent in the range of a few millimetres. Given the nature of the processes and the uncertainties in the DEM and input data this is negligible. This means that in the combined hazard analysis the specic point in time of coincidence of rainstorm and uvial inundation is not of major importance, at least for the used criteria of maximum inundation depth and extent. It can be suspected that the timing of the rainstorm within the uvial event plays a more pronounced role if uvial and pluvial events of different probabilities of occurrence are combined, e.g. a uvial event with p = 0.5 combined with a pluvial event of p = 0.99. However,

as these combinations were not simulated this can only be speculated with the current results at hand.

Spatial coincidence of the ood triggers also proved to be of minor importance. The fact that the maximum inundation depths of the combined analysis do not differ much from the uvial analysis could indicate that with the given set of 140 Monte Carlo combinations, the random storm centres do

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 957

Figure 14. Probabilistic combined uvialpluvial ood hazard maps for Can Tho showing maximum inundation depths: for the derived p levels (annual probabilities of non-exceedance), the median (50 % quantile) maps and associated 5 and 95 % quantile maps are shown illustrating the uncertainty of the hazard estimation. Note that the p levels are different to the uvial and pluvial hazard maps due to the discrete combinations of uvial and pluvial p levels. Maximum inundation depths < 0.02 m are indicated as no inundation.

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

958 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

not fall in the areas where the uvial inundation is highest.

An increase of the Monte Carlo runs could clarify this question. However, although the simulation time with the models and hardware at hand are in acceptable limits, simulation time is still a prohibitive factor in this respect.

4.4 Implications for risk assessment

Combining uvial and pluvial ood events in a hazard analysis has implications for estimating ood risk. In ood risk assessments (FRAs), the occurrence of uvial, pluvial and combined ood events and their probabilities have to be taken into account. While a combination of uvial and pluvial hazard without any interaction can be considered straightforwardly as the sum of the individual hazards and associated risks, the inclusion of the combined hazard requires the consideration that parts of the uvial and pluvial scenarios coincide. In the following the formal consideration of this coincidence in a risk assessment is derived for the calculation of expected annual damages (EADs). EADs are dened as the product of the annual probability of exceedance of a given ood event and the damage it inicts. For a discrete set of scenarios, i.e. probability levels as used in this study, EADs are formulated as the following:

EAD =

n

EAD(,pl, pl) =

n

Xi=1[Delta1]

P ()i P ( pl)i

[bracketrightbig] D()i

+

n

Xi=1[Delta1]

P (pl)i P ( pl)i

[bracketrightbig] D(pl)i

Xi=1[Delta1]P ( pl)i D( pl)i. (8)

This formulation is valid for the presented combination of uvial and pluvial probability levels, which are identical in each combination. If a complete permutation of the probability levels is performed, the additional scenarios and their probabilities must also be considered analogously as in the special case presented here. Formally this is expressed as an extension of the formula above:

EAD(,pl, pl) =

n

Xi=1[Delta1]

+

n

"P ()i

n

Xj=1P ( pl)i,j

[bracketrightBigg]

D()i

+

n

Xi=1[Delta1]

"P (pl)i

n

Xj=1P ( pl)i

[bracketrightBigg]

D(pl)i

+

n

Xi=1n

Xi=1[Delta1]Pi Di, (5)

where [Delta1]P is the increment of annual probability of exceedance = [Delta1](1 p), with p as the annual probability of

non-exceedance; D is the damage inicted; i is the numerator of the probability levels considered, and n is the number of probability levels. As this formulation is an approximation of the integration of a (hypothetical) continuous risk curve, the annual probability of exceedance and the damage have to be interpolated between the p levels (Ward et al., 2011; Merz et al., 2009). Using a linear interpolation [Delta1]Pi and Di are dened as the following:

Di =

1

2 (Di+1 + Di) (6)

[Delta1]Pi = Pi+1 Pi. (7)

If the uvial () and pluvial (pl) hazards were not coinciding, the EADs for combined uvial and pluvial risk would be simply the sum of EAD() and EAD(pl). However, in the presented study the uvial and the pluvial hazard can coincide, and this has to be taken into account in the calculation of the overall EAD. Formally this is achieved by reducing the annual probability of exceedance of the individual hazards by the annual probability of exceedance of the combined hazard:

Xj=1[Delta1]P ( pl)i,j D( pl)i,j . (9)

4.5 Limitations of the approach

Although the presented results appear to be plausible and are accompanied by uncertainty estimations, it is clear that such comprehensive approaches have their limitations. This section identies and discusses the most important limitations.

The developed methodology considers the natural (aleatory) uncertainty only, i.e. the uncertainty stemming from the natural variability of the ood triggering events.In the uvial analysis the variability of the occurrences of ood peak and ood volume, and the hydrograph shape at the upper boundary of the Mekong Delta were taken into account in this respect. In the pluvial analysis natural variability was mapped by the magnitude and location of rainstorm events. However, the quantication of these uncertainties is also plagued with uncertainty, which is termed epistemic uncertainty. This generally stems from imperfect knowledge of the processes, insufcient data or models. For example, for the uvial analysis Dung et al. (2015) showed that the bivariate frequency analysis is associated with considerable epistemic uncertainty stemming from a limited length of the data series. This uncertainty is not considered in the presented study. Technically, a consideration of this uncertainty source would mean an extension of the Monte Carlo analysis, in which even more pairs of ood peak Qmax and ood volume V per probability level have to be considered. Whether this is feasible, is mainly a question of simulation time. The resulting percentile maps would certainly show a wider range compared to this study. The same holds true if other epistemic uncertainty sources were considered.

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 959

The pluvial analysis also contains epistemic uncertainty which was also not quantied. The epistemic uncertainty is associated to the assumptions listed in Sect. 3.2.2. The assumptions on storm extent and spatial distribution of intensity have particularly large impacts on the inundation simulation and thus the probabilistic hazard maps. Unfortunately, no data for the validation of the assumptions were available to the authors, and thus a quantication of the uncertainty caused by the assumptions cannot be performed. An evaluation of radar rain data of the area, as e.g. performed in Barnolas et al. (2010), would be of great benet for the char-acterisation and thus the validation of the assumptions taken.

The analysis of the spatial distribution of maximum rainfall intensities within the Monte Carlo analysis revealed a small negative bias (cf. Sect. 3.2.2). While this bias is already small, a increase of the random storm centre location,i.e. a higher number of Monte Carlo runs, would diminish this bias. In a test the bias was reduced to virtually zero in a MC run with 1000 random storm centres. This would of course increase the computational time considerably, particularly for the combined uvialpluvial events. The question of what is an acceptable bias is thus a compromise between desired accuracy and practical considerations. Although the test for the bias introduced by the assumption of the synthetic rainstorms is not encompassing, it can be stated for the presented test case that if a bias of less than 1 % is regarded acceptable, the quotient derived by the division of the area of the simulation domain [m2] by the product of Monte Carlo runs and FWTM [m] of the synthetic storms should be smaller than 16. This may be used as a rough guideline for similar studies.

Further sources of epistemic uncertainty are the hydraulic model and the quality of the DEM. While the horizontal resolution is acceptable for a study of this scale, the quality of the vertical information has to be critically assessed. Because the DEM was interpolated from elevation information of topographical maps, it has to be expected that the accuracy is actually not sufcient for urban inundation simulation. Additional uncertainty is caused by the lack of calibration and validation of the hydraulic model due to insufcient data. However, sensitivity runs with different distributions and values of roughness parameters indicate that the uncertainty in maximum inundation depths is dominated by the topography and not by the parameterization of roughness values.

These uncertainties have to be taken into account when using the derived hazard maps for actual ood management planning. It is recommended to use the maps mainly for the identication of the most affected areas, and to use the inundation depths with caution.

5 Conclusions

This study develops a methodology for a combined uvial and pluvial ood hazard analysis. The methodology is exemplarily developed for Can Tho city, which is the economic centre of the Vietnamese part of the Mekong Delta. Both u-vial and pluvial inundation processes cause regular ooding in Can Tho city. Both uvial and pluvial hazard analyses contain a dedicated statistical part to estimate the probabilities of occurrence of oods of different magnitudes. Synthetic ood events were derived based on this frequency analysis, which provide the boundary conditions for a 2-dimensional inundation modelling of the central part of the city. With the help of the hydraulic model, maximum inundation depths were determined for every ood scenario for Can Tho city.

The two ood hazards were not only considered independent from each other, but also in combination, i.e. occurring at the same time. Because uvial and pluvial ooding occurs in the same period of the year, these events can coincide. As the triggering events are essentially independent from each other, the probability of occurrence of coinciding events can be directly calculated from the probability of occurrences of the individual events and the probability of coincidence of the events. The resulting inundation events bear features of both uvial and pluvial inundation processes. The presented method is novel not only in terms of the specic application for Can Tho city, but also in general, as a similar study has not been published so far. The presented methods can be transferred to other cities in a similar tropical setting and adapted to different climate zones.

The hazard analyses also include uncertainty estimations.

For the uvial, pluvial, and combined hazard the natural uncertainty, i.e. variability of ood events with identical probabilities of occurrence, was taken into account. This facilitated the derivation of probabilistic hazard maps showing the maximum inundation depths with an expectation, i.e. the median hazard map, and percentile maps quantifying the natural uncertainty in a spatially explicit manner. These maps can serve as a basis for ood management and mitigation plans, for a ood risk assessment for Can Tho city, or for cost-benet analyses of ood protection measures. The derived hazard maps are available as GeoTiffs ready for use in a GIS in the Supplement related to this paper.

However, for usage of the hazard maps in practical applications the limitations, resp. the epistemic uncertainty which was not quantied have to be taken into account. In this respect the quality of the digital elevation model and the assumptions underlying the derivation of synthetic rainstorm events have the highest impact on the nal hazard estimates, particularly the maximum inundation depths. Thus for the time being, the maps should be mainly used for the determination of the most affected areas, and the maximum inundation depths should be taken as rough guidance only. Considering the prominent damage pathways inundation of houses by water levels in the streets exceeding road curbs a possi-

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016

960 H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City

ble way of direct usage of the maps could be to target ood mitigation plans to areas with maximum inundation depths exceeding a certain threshold, e.g. 10 cm. This would eliminate the random character of the pluvial maps and thus the inuence of the assumptions taken on rainstorm extent and storm centre locations. In order to increase the overall condence in the hazard maps, the assumptions on rains storms should be validated, preferably by an analysis of a series of rain radar maps. Additionally, an improved DEM should be obtained and the hydraulic simulations repeated.

In any case, the presence of two or more hazard sources and their combination have to be taken into account in any risk assessment in a statistically sound and consistent manner. In general, the probability of occurrence of combined ood events reduces the probabilities of occurrence of individual uvial and pluvial events of the same magnitude, which needs to be reected in the risk calculations. How this can be achieved, has been exemplarily shown for a hypothetical calculation of expected annual damages (EADs), a standard risk variable.

Given the fact that the hydrometeorological setting of Can Tho city is similar to many low-lying cities in the tropics, the presented methodology can be used as a blueprint for ood hazard analysis in these cities. Due to the convective nature of tropical rainfall, the developed method for a probabilistic pluvial ood hazard analysis should be easily transferable.The uvial part, however, should be modied to match the hydrological situation at hand. In most cases this means a simplication compared to the presented study, because such complicated hydraulic situations as in the Mekong Delta are not very frequent.

However, if the independence of the uvial and pluvial ood events is not given, the approach has to be modied in order to take the dependence structure into account. This may be achieved by analysis of sufcient long time series of river discharge/water level and precipitation or by an approach similar to Breinl et al. (2015), utilizing a large-scale weather generator and a hydrological model.

The Supplement related to this article is available online at http://dx.doi.org/10.5194/nhess-16-941-2016-supplement

Web End =doi:10.5194/nhess-16-941-2016-supplement .

Acknowledgements. This study was conducted within the framework of the WISDOM project which was funded by the German Ministry of Education and Research BMBF (FKZ 033L040DN) and the Vietnamese Ministry of Science and Technology MOST.Funding by the ministries and support from various Vietnamese administrative bodies is gratefully acknowledged.

The article processing charges for this open-access publication were covered by a ResearchCentre of the Helmholtz Association.

Edited by: M.-C. LlasatReviewed by: two anonymous referees

References

Apel, H., Thieken, A. H., Merz, B., and Blschl, G.: Flood risk assessment and associated uncertainty, Nat. Hazards Earth Syst.Sci., 4, 295308, doi:http://dx.doi.org/10.5194/nhess-4-295-2004

Web End =10.5194/nhess-4-295-2004 http://dx.doi.org/10.5194/nhess-4-295-2004

Web End = , 2004.

Apel, H., Merz, B., and Thieken, A. H.: Quantication of uncertainties in ood risk assessments, Int. J. River Basin Manage., 6, 149162, 2008.

Apel, H., Aronica, G., Kreibich, H., and Thieken, A.: Flood risk analyses how detailed do we need to be?, Nat. Hazards, 49, 7998, 2009.

Arrighi, C., Brugioni, M., Castelli, F., Franceschini, S., and Mazzanti, B.: Urban micro-scale ood risk estimation with parsimonious hydraulic modelling and census data, Nat. Hazards Earth Syst. Sci., 13, 13751391, doi:http://dx.doi.org/10.5194/nhess-13-1375-2013

Web End =10.5194/nhess-13-1375- http://dx.doi.org/10.5194/nhess-13-1375-2013

Web End =2013 , 2013.

Barnolas, M., Rigo, T., and Llasat, M. C.: Characteristics of 2-D convective structures in Catalonia (NE Spain): an analysis using radar data and GIS, Hydrol. Earth Syst. Sci., 14, 129139, doi:http://dx.doi.org/10.5194/hess-14-129-2010

Web End =10.5194/hess-14-129-2010 http://dx.doi.org/10.5194/hess-14-129-2010

Web End = , 2010.

Bates, P. D., Horritt, M. S., and Fewtrell, T. J.: A simple inertial formulation of the shallow water equations for efcient two-dimensional ood inundation modelling, J. Hydrol., 387, 3345, doi:http://dx.doi.org/10.1016/j.jhydrol.2010.03.027

Web End =10.1016/j.jhydrol.2010.03.027 http://dx.doi.org/10.1016/j.jhydrol.2010.03.027

Web End = , 2010.

Blanc, J., Hall, J. W., Roche, N., Dawson, R. J., Cesses, Y., Burton,A., and Kilsby, C. G.: Enhanced efciency of pluvial ood risk estimation in urban areas using spatial-temporal rainfall simulations, J. Flood Risk Manage., 5, 143152, doi:http://dx.doi.org/10.1111/j.1753-318X.2012.01135.x

Web End =10.1111/j.1753- http://dx.doi.org/10.1111/j.1753-318X.2012.01135.x

Web End =318X.2012.01135.x , 2012.

Breinl, K., Strasser, U., Bates, P., and Kienberger, S.: A joint modelling framework for daily extremes of river discharge and precipitation in urban areas, J. Flood Risk Manage., doi:http://dx.doi.org/10.1111/jfr3.12150

Web End =10.1111/jfr3.12150 http://dx.doi.org/10.1111/jfr3.12150

Web End = , in press, 2015.

Burton, A., Kilsby, C. G., Fowler, H. J., Cowpertwait, P. S. P., and OConnell, P. E.: RainSim: A spatialtemporal stochastic rainfall modelling system, Environ. Model. Softw., 23, 1356-1369, doi:http://dx.doi.org/10.1016/j.envsoft.2008.04.003

Web End =10.1016/j.envsoft.2008.04.003 http://dx.doi.org/10.1016/j.envsoft.2008.04.003

Web End = , 2008.

Chan, F. K. S., Mitchell, G., Adekola, O., and McDonald, A.: Flood Risk in Asias Urban Mega-deltas: Drivers, Impacts and Response, Environ. Urbaniz. Asia, 3, 4161, doi:http://dx.doi.org/10.1177/097542531200300103

Web End =10.1177/097542531200300103 http://dx.doi.org/10.1177/097542531200300103

Web End = , 2012.

Chen, A. S., Djordjevic, S., Leandro, J., and Savic, D. A.: An analysis of the combined consequences of pluvial and uvial ooding, Water Sci. Technol., 62, 14911498, 2010.

Chinh, D. T., Bubeck, P., Dung, N. V., and Kreibich, H.: The

2011 ood event in the Mekong Delta: preparedness, response, damage and recovery of private households and small businesses, Disasters, doi:http://dx.doi.org/10.1111/disa.12171

Web End =10.1111/disa.12171 http://dx.doi.org/10.1111/disa.12171

Web End = , in press, 2016.de Almeida, G. A. M. and Bates, P.: Applicability of the local inertial approximation of the shallow water equations to ood modelling, Water Resour. Res., 49, 48334844, doi:http://dx.doi.org/10.1002/wrcr.20366

Web End =10.1002/wrcr.20366 http://dx.doi.org/10.1002/wrcr.20366

Web End = , 2013.de Bruijn, K. M., Diermanse, F. L. M., and Beckers, J. V. L.: An advanced method for ood risk analysis in river deltas, applied to societal ood fatality risk in the Netherlands, Nat. Hazards Earth Syst. Sci., 14, 27672781, doi:http://dx.doi.org/10.5194/nhess-14-2767-2014

Web End =10.5194/nhess-14-2767- http://dx.doi.org/10.5194/nhess-14-2767-2014

Web End =2014 , 2014.

Dimitriadis, P., Tegos, A., Oikonomou, A., Pagana, V., Koukouvinos, A., Mamassis, N., Koutsoyiannis, D., and Efstratiadis,A.: Comparative evaluation of 1D and quasi-2D hydraulic mod-

Nat. Hazards Earth Syst. Sci., 16, 941961, 2016 www.nat-hazards-earth-syst-sci.net/16/941/2016/

H. Apel et al.: Combined uvial and pluvial urban ood hazard analysis: Can Tho City 961

els based on benchmark and real-world applications for uncertainty assessment in ood mapping, J. Hydrol., 534, 478492, doi:http://dx.doi.org/10.1016/j.jhydrol.2016.01.020

Web End =10.1016/j.jhydrol.2016.01.020 http://dx.doi.org/10.1016/j.jhydrol.2016.01.020

Web End = , 2016.

Dung, N. V., Merz, B., Brdossy, A., Thang, T. D., and Apel, H.:

Multi-objective automatic calibration of hydrodynamic models utilizing inundation maps and gauge data, Hydrol. Earth Syst.Sci., 15, 13391354, doi:http://dx.doi.org/10.5194/hess-15-1339-2011

Web End =10.5194/hess-15-1339-2011 http://dx.doi.org/10.5194/hess-15-1339-2011

Web End = , 2011.Dung, N. V., Merz, B., Brdossy, A., and Apel, H.: Handling uncertainty in bivariate quantile estimation An application to ood hazard analysis in the Mekong Delta, J. Hydrol., 527, 704717, doi:http://dx.doi.org/10.1016/j.jhydrol.2015.05.033

Web End =10.1016/j.jhydrol.2015.05.033 http://dx.doi.org/10.1016/j.jhydrol.2015.05.033

Web End = , 2015.

Falter, D., Schrter, K., Dung, N. V., Vorogushyn, S., Kreibich,H., Hundecha, Y., Apel, H., and Merz, B.: Spatially coherent ood risk assessment based on long-term continuous simulation with a coupled model chain, J. Hydrol., 524, 182193, doi:http://dx.doi.org/10.1016/j.jhydrol.2015.02.021

Web End =10.1016/j.jhydrol.2015.02.021 http://dx.doi.org/10.1016/j.jhydrol.2015.02.021

Web End = , 2015.

Fewtrell, T. J., Duncan, A., Sampson, C. C., Neal, J. C., and Bates, P.D.: Benchmarking urban ood models of varying complexity and scale using high resolution terrestrial LiDAR data, Phys. Chem.Earth, 36, 281291, doi:http://dx.doi.org/10.1016/j.pce.2010.12.011

Web End =10.1016/j.pce.2010.12.011 http://dx.doi.org/10.1016/j.pce.2010.12.011

Web End = , 2011.

Hall, J. W., Sayers, P. B., and Dawson, R. J.: National-scale Assessment of Current and Future Flood Risk in England and Wales, Nat. Hazards, 36, 147164, 2005.

Hundecha, Y., Pahlow, M., and Schumann, A.: Modeling of daily precipitation at multiple locations using a mixture of distributions to characterize the extremes, Water Resour. Res., 45, W12412, doi:http://dx.doi.org/10.1029/2008wr007453

Web End =10.1029/2008wr007453 http://dx.doi.org/10.1029/2008wr007453

Web End = , 2009.

Huong, H. T. L. and Pathirana, A.: Urbanization and climate change impacts on future urban ooding in Can Tho city, Vietnam, Hydrol. Earth Syst. Sci., 17, 379-394, doi:http://dx.doi.org/10.5194/hess-17-379-2013

Web End =10.5194/hess-17-379- http://dx.doi.org/10.5194/hess-17-379-2013

Web End =2013 , 2013.

Huth, J., Kuenzer, C., Wehrmann, T., Gebhardt, S., Tuan, V. Q., and Dech, S.: Land Cover and Land Use Classication with TWOPAC: towards Automated Processing for Pixel- and Object-Based Image Classication, Remote Sensing, 4, 25302553, 2012.

Kuenzer, C., Campbell, I., Roch, M., Leinenkugel, P., Tuan, V., and Dech, S.: Understanding the impact of hydropower developments in the context of upstreamdownstream relations in the Mekong river basin, Sustain. Sci., 8, 565584, doi:http://dx.doi.org/10.1007/s11625-012-0195-z

Web End =10.1007/s11625-012- http://dx.doi.org/10.1007/s11625-012-0195-z

Web End =0195-z , 2013a.

Kuenzer, C., Guo, H., Schlegel, I., Tuan, V. Q., Li, X., and

Dech, S.: Varying Scale and Capability of Envisat ASAR-WSM, TerraSAR-X Scansar and TerraSAR-X Stripmap Data to Assess Urban Flood Situations: A Case Study of the Mekong Delta in Can Tho Province, Remote Sensing, 5, 51225142, 2013b.Manh, N. V., Dung, N. V., Hung, N. N., Merz, B., and Apel, H.:

Large-scale suspended sediment transport and sediment deposition in the Mekong Delta, Hydrol. Earth Syst. Sci., 18, 3033 3053, doi:http://dx.doi.org/10.5194/hess-18-3033-2014

Web End =10.5194/hess-18-3033-2014 http://dx.doi.org/10.5194/hess-18-3033-2014

Web End = , 2014.

McMillan, H. K. and Brasington, J.: End-to-end ood risk assessment: A coupled model cascade with uncertainty estimation, Water Resour. Res., 44, W03419, doi:http://dx.doi.org/10.1029/2007wr005995

Web End =10.1029/2007wr005995 http://dx.doi.org/10.1029/2007wr005995

Web End = , 2008.

Merz, B. and Thieken, A. H.: Separating natural and epistemic uncertainty in ood frequency analysis, J. Hydrol., 309, 114132, doi:http://dx.doi.org/10.1016/j.jhydrol.2004.11.015

Web End =10.1016/j.jhydrol.2004.11.015 http://dx.doi.org/10.1016/j.jhydrol.2004.11.015

Web End = , 2005.

Merz, B., Elmer, F., and Thieken, A. H.: Signicance of high probability/low damage versus low probability/high damage ood events, Nat. Hazards Earth Syst. Sci., 9, 10331046, doi:http://dx.doi.org/10.5194/nhess-9-1033-2009

Web End =10.5194/nhess-9-1033-2009 http://dx.doi.org/10.5194/nhess-9-1033-2009

Web End = , 2009.

Merz, B., Hall, J., Disse, M., and Schumann, A.: Fluvial ood risk management in a changing world, Nat. Hazards Earth Syst. Sci., 10, 509527, doi:http://dx.doi.org/10.5194/nhess-10-509-2010

Web End =10.5194/nhess-10-509-2010 http://dx.doi.org/10.5194/nhess-10-509-2010

Web End = , 2010.

MunichRe: Natural Disasters 2014: https://www.munichre.com/touch/naturalhazards/en/natcatservice/annual-statistics/index.html

Web End =https://www.munichre.com/ https://www.munichre.com/touch/naturalhazards/en/natcatservice/annual-statistics/index.html

Web End =touch/naturalhazards/en/natcatservice/annual-statistics/index. https://www.munichre.com/touch/naturalhazards/en/natcatservice/annual-statistics/index.html

Web End =html , last access: 25 February 2015.

Nuswantoro, R., Diermanse, F., and Molkenthin, F.: Probabilistic ood hazard maps for Jakarta derived from a stochastic rainstorm generator, J. Flood Risk Manage., doi:http://dx.doi.org/10.1111/jfr3.12114

Web End =10.1111/jfr3.12114 http://dx.doi.org/10.1111/jfr3.12114

Web End = , in press, 2014.

Renaud, F. G., Syvitski, J. P. M., Sebesvari, Z., Werners, S. E., Kremer, H., Kuenzer, C., Ramesh, R., Jeuken, A., and Friedrich, J.: Tipping from the Holocene to the Anthropocene: How threatened are major world deltas?, Curr. Opin. Environ. Sustain., 5, 644654, doi:http://dx.doi.org/10.1016/j.cosust.2013.11.007

Web End =10.1016/j.cosust.2013.11.007 http://dx.doi.org/10.1016/j.cosust.2013.11.007

Web End = , 2013.

Syvitski, J. P. M.: Deltas at risk, Sustain. Sci., 3, 2332, doi:http://dx.doi.org/10.1007/s11625-008-0043-3

Web End =10.1007/s11625-008-0043-3 http://dx.doi.org/10.1007/s11625-008-0043-3

Web End = , 2008.

Syvitski, J. P. M. and Higgins, S.: Going under: The worlds sinking deltas, New Scientist, 216, 4043, doi:http://dx.doi.org/10.1016/S0262-4079(12)63083-8

Web End =10.1016/S0262- http://dx.doi.org/10.1016/S0262-4079(12)63083-8

Web End =4079(12)63083-8 , 2012.

Takagi, H., Ty, T. V., Thao, N. D., and Esteban, M.: Ocean tides and the inuence of sea-level rise on oods in urban areas of the Mekong Delta, J. Flood Risk Manage., 8, 292300, doi:http://dx.doi.org/10.1111/jfr3.12094

Web End =10.1111/jfr3.12094 http://dx.doi.org/10.1111/jfr3.12094

Web End = , 2015.

Thompson, C. M. and Frazier, T. G.: Deterministic and probabilistic ood modeling for contemporary and future coastal and inland precipitation inundation, Appl. Geogr., 50, 114, doi:http://dx.doi.org/10.1016/j.apgeog.2014.01.013

Web End =10.1016/j.apgeog.2014.01.013 http://dx.doi.org/10.1016/j.apgeog.2014.01.013

Web End = , 2014.

Vorogushyn, S., Merz, B., Lindenschmidt, K. E., and Apel,H.: A new methodology for ood hazard assessment considering dike breaches, Water Resour. Res., 46, W08541, doi:http://dx.doi.org/10.1029/2009WR008475

Web End =10.1029/2009WR008475 http://dx.doi.org/10.1029/2009WR008475

Web End = , 2010.

Ward, P. J., de Moel, H., and Aerts, J. C. J. H.: How are ood risk estimates affected by the choice of return-periods?, Nat. Hazards Earth Syst. Sci., 11, 31813195, doi:http://dx.doi.org/10.5194/nhess-11-3181-2011

Web End =10.5194/nhess-11- http://dx.doi.org/10.5194/nhess-11-3181-2011

Web End =3181-2011 , 2011.

Wassmann, R., Hien, N. X., Hoanh, C. T., and Tuong,T. P.: Sea Level Rise Affecting the Vietnamese Mekong Delta: Water Elevation in the Flood Season and Implications for Rice Production, Climatic Change, 66, 89107, doi:http://dx.doi.org/10.1023/B:CLIM.0000043144.69736.b7

Web End =10.1023/B:CLIM.0000043144.69736.b7 http://dx.doi.org/10.1023/B:CLIM.0000043144.69736.b7

Web End = , 2004.

Wheater, H., Chandler, R., Onof, C., Isham, V., Bellone, E., Yang, C., Lekkas, D., Lourmas, G., and Segond, M.-L.: Spatial-temporal rainfall modelling for ood risk estimation, Stoch. Environ. Res. Risk A., 19, 403416, 2005.

Willems, P.: A spatial rainfall generator for small spatial scales,J. Hydrol., 252, 126144, doi:http://dx.doi.org/10.1016/S0022-1694(01)00446-2

Web End =10.1016/S0022-1694(01)00446-2 http://dx.doi.org/10.1016/S0022-1694(01)00446-2

Web End = , 2001.

www.nat-hazards-earth-syst-sci.net/16/941/2016/ Nat. Hazards Earth Syst. Sci., 16, 941961, 2016