A joint modelling framework for daily extremes of river discharge and precipitation in urban areas

Abstract

Human settlements are often at risk from multiple hydro-meteorological hazards, which include fluvial floods, short-time extreme precipitation (leading to ′pluvial′ floods), or coastal floods. In the past, considerable scientific effort has been devoted to assessing fluvial floods. Only recently have methods been developed to assess the hazard and risk originating from pluvial phenomena, while little effort has been dedicated to joint approaches. The aim of this study was to develop a joint modelling framework for simulating daily extremes of river discharge and precipitation in urban areas. The basic framework is based on daily observations coupled with a novel precipitation disaggregation algorithm using nearest neighbour resampling combined with the method of fragments, to overcome data limitations and facilitate its transferability. The framework generates dependent time series of river discharge and urban precipitation that allow for the identification of fluvial flood days (daily peak discharge), days of extreme precipitation potentially leading to pluvial phenomena (maximum hourly precipitation) and combined fluvial-pluvial flood days (combined time series). Critical thresholds for hourly extreme precipitation were derived from insurance and fire service data.

Full text

A joint modelling framework for daily extremes of river discharge
and precipitation in urban areas
K. Breinl1, U. Strasser2, P. Bates3and S. Kienberger1
1 Department of Geoinformatics, Paris-Lodron University of Salzburg, Salzburg, Austria
2 Institute of Geography, University of Innsbruck, Innsbruck, Austria
3 School of Geographical Sciences, University of Bristol, Bristol, UK
Correspondence
Korbinian Breinl, Department of
Geoinformatics, Paris-Lodron University of
Salzburg, Schillerstrasse 30, 5020 Salzburg,
Austria
Tel: +43 (0)662 8044 7578
Email: korbinian.breinl@sbg.ac.at
DOI: 10.1111/jfr3.12150
Key words
Fluvial floods; hydrological modelling;
joint hazard modelling; pluvial floods;
stochastic modelling; weather generator.
Abstract
Human settlements are often at risk from multiple hydro-meteorological
hazards, which include fluvial floods, short-time extreme precipitation (leading
to ‘pluvial’ floods) or coastal floods. In the past, considerable scientific effort has
been devoted to assessing fluvial floods. Only recently have methods been devel-
oped to assess the hazard and risk originating from pluvial phenomena, whereas
little effort has been dedicated to joint approaches. The aim of this study was to
develop a joint modelling framework for simulating daily extremes of river
discharge and precipitation in urban areas. The basic framework is based on
daily observations coupled with a novel precipitation disaggregation algorithm
using nearest neighbour resampling combined with the method of fragments to
overcome data limitations and facilitate its transferability. The framework gen-
erates dependent time series of river discharge and urban precipitation that
allow for the identification of fluvial flood days (daily peak discharge), days of
extreme precipitation potentially leading to pluvial phenomena (maximum
hourly precipitation) and combined fluvial–pluvial flood days (combined time
series). Critical thresholds for hourly extreme precipitation were derived from
insurance and fire service data.
Introduction
Urban flooding can have multiple sources including fluvial
(river) floods, floods from high-intensity precipitation
(‘pluvial floods’) or coastal floods. Pluvial phenomena are
particularly likely in urban areas due to impervious surfaces
and are highly variable in time and space. Fluvial floods are
less variable and thus easier to predict and forecast. The
flooding intensity is related to the catchment hydrology.
Previous research mainly focused on fluvial flooding (e.g.
Bronstert, 1995; Apel et al., 2004; Bradbrook et al., 2005;
Hall et al., 2005; Merz et al., 2005; Apel et al., 2006; Büchele
et al., 2006; Grünthal et al., 2006; Apel et al., 2009; de Kok
and Grossmann, 2010; Veijalainen et al., 2010; Feyen et al.,
2012). More recent studies shifted the attention to pluvial
flooding (e.g. Falconer et al., 2009; Kazmierczak and Cavan,
2011; Priest et al., 2011; Blanc et al., 2012; Bradford et al.,
2012; Hurford et al., 2012; Zhou et al., 2012; Spekkers et al.,
2013): it has been recognised that pluvial floods may become
more frequent due to climate change (Larsen et al., 2009;
Madsen et al., 2009; Mailhot and Duchesne, 2010; IPCC,
2012; Zhou et al., 2012), a trend confirmed by the loss
experience of governments and insurers (Munich Re, 2013).
Moreover, only recent formulations of the shallow water
equations made pluvial hydraulic modelling feasible at
acceptable computational costs (e.g. ‘reduced complexity’
hydraulic codes by Hunter et al., 2007; Bates et al., 2010;
Sampson et al., 2013). Last but not least, the smaller scale of
pluvial phenomena might explain a lack of attention in the
past (Dawson et al., 2008).
Fluvial floods are typically assessed by hydrological mod-
elling (e.g. Driessen et al., 2010; Veijalainen et al., 2010) or
extreme value statistics (e.g. Keef et al., 2009; Lamb et al.,
2010), often combined with hydraulic modelling. Pluvial
floods have been mainly assessed by hydraulic modelling
(e.g. Morris et al., 2009; Chen et al., 2010; Zhou et al., 2012),
also combined with field measurements (Neal et al., 2009),
or with Geographic Information System (GIS) analyses
(Falconer et al., 2009). Little work was undertaken on joint
flood hazard assessment. Dawson et al. (2008) examined
fluvial and pluvial floods in a synthetic case study by means
of hydraulic modelling. Chen et al. (2010) hydraulically
simulated joint fluvial–pluvial scenarios in a village without
calculating joint probabilities, whereas Lian et al. (2013)
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd

examined the joint probability of tidal floods and extreme
precipitation in a coastal city.
The motivation behind this research was the question
whether fluvial and pluvial flood days may occur simulta-
neously and if this could be modelled with only a moderate
amount of observation data at hand. The resulting frame-
work works as follows (Figure 1): the fluvial component is
simulated with a hydrological model (HM), producing daily
mean discharge. The HM is driven with a stochastic daily
multisite weather generator (WG), which provides precipi-
tation and temperature for the entire river catchment as well
as precipitation for the urban domain. As the basic frame-
work works at a daily resolution and pluvial phenomena
occur at subdaily timescale, daily urban precipitation is
disaggregated to hourly maxima with a novel disaggregation
method using nearest neighbour resampling combined with
the method of fragments (MOF). Likewise, daily mean dis-
charge is converted into daily peaks, which are relevant for
fluvial floods. The two generated time series allow the iden-
tification of fluvial flood days (daily peak discharge), pluvial
flood days (maximum hourly urban precipitation) and com-
bined fluvial–pluvial flood days (combined time series).
Critical thresholds for precipitation potentially leading to
pluvial floods were derived from insurance data and opera-
tion records from the local fire service. The study was carried
out for the city of Salzburg (Austria), which is located at the
Salzach River and prone to both types of floods.
There are various reasons why the basic framework is
based on daily data, coupled with single-site precipitation
disaggregation for the urban area: first, many study areas
lack the availability of subdaily observation data. Hourly
modelling with a small number of hourly sites is often not
feasible, especially as the calibration of an HM ‘worsens radi-
cally with an excessive reduction of rain gauges’ (Bardossy
and Das, 2008). In the present study, only 5 hourly sites were
available with only 19 available simultaneous observation
years, which neither allows for a stable calibration of an HM
nor reliable probabilistic analyses. To overcome such data
limitations, daily precipitation at multiple sites can be
disaggregated to hourly values by using a single hourly ref-
erence site (e.g. Segond et al., 2006) or using weather radar
data (e.g. Verworn and Haberlandt, 2011). However, espe-
cially in mountainous areas, subdaily temporal patterns of
precipitation are more independent due to the strong influ-
ence of the relief (Mezghani and Hingray, 2009). Imposing
patterns from a single hourly site on all daily sites would
overestimate intersite correlations and introduce bias in the
HM output (Mezghani and Hingray, 2009). The use of radar
data would reduce the transferability of the concept. Second,
the stochastic generation of hourly precipitation is consid-
Figure 1 Simplified schematics of the modelling framework (adapted from Dawson et al., 2008).
2Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
98 Breinl et al.

erably more complex than of daily precipitation. Third, the
daily concept can be fairly easily transferred to other (data
scarce) study areas, as only subdaily precipitation for the
urban area (i.e. one site) is required.
Study area and data
The city of Salzburg is located at the Salzach River,one of the
major Alpine rivers. The Salzach catchment down to the city
of Salzburg has an area of 4637 km2(see Figure 2).
The catchment is characterised by mountains exceeding
3500 m.a.s.l. In our model experiment, daily totals of pre-
cipitation and daily mean temperature were available from
1987 to 2010. For precipitation, the stations 1–19 were used;
for temperature, the station 15, 16 and 19 (see Figure 2 and
Table 1).
Hourly urban precipitation data for gauge 19 could be
obtained for 1994–2012. Daily and hourly time series were
checked for consistency. Quality-checked daily mean and
peak discharge records in m3/s at the Salzach River gauge in
Salzburg were available from 1951 to 2010 and from 1976 to
2010, respectively. Figure 3 summarises the available data.
Hydrology and meteorology
The discharge regime of the Salzach River is dominated by
snowmelt. The maximum mean monthly discharge occurs in
June, the minimum in January. River floods mainly occur in
summer and are typically caused by a combination of
snowmelt and long-lasting spring and summer precipita-
tion. Between 1976 and 2010, the majority of annual peak
discharges occurred in July. The highest daily peak discharge
recorded was the August flood in 2002 with 2289 m3/s.
The 2013 June flood, for which no verified discharge data
yet exist, was estimated comparable with the 2002 event
with respect to the discharge amount; both events were
Figure 2 The Salzach catchment with precipitation gauges (code 1–19) and temperature recordings (at stations 15, 16 and 19) used. The
river gauge for calibration of the hydrological model is located in the city of Salzburg (adapted from Breinl et al. 2014).
Modelling framework for urban river discharge and precipitation 3
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 99

triggered by a ‘Vb’ weather situation (Mudelsee et al., 2004;
Kundzewicz et al., 2005) combined with snowmelt.
Pluvial floods are triggered by torrential precipitation
from thunderstorms or mesoscale convective systems, which
are frequent in summer (Sene, 2013). In the literature, a
typical critical threshold of precipitation is 30 mm/h (e.g.
Falconer et al., 2008; Parker et al., 2011; Priest et al., 2011;
Hurford et al., 2012).According to the German Meteorologi-
cal Service (DWD), ‘severe heavy precipitation’ is related
to >25 mm/h. The Canadian Atmospheric Environment
Service defines events >25 mm/h as ‘severe thunderstorms’
(WMO, 2003). In Austria, no such threshold exists.
Table 2 lists all precipitation days recorded in Salzburg-
Freisaal (gauge 19) exceeding a maximum of 25 mm/h since
1994.
The most recent event in June 2012 was the highest hourly
amount recorded since 1994, causing severe damage in the
city. As Table 2 indicates, extreme precipitation typically
occurs in summer.
Modelling framework
Assumptions and limitations
For the modelling framework, the following key assump-
tions were made:
• The focus is on urban flooding, i.e. the main urbanised
area of Salzburg (area approximately 30 km2).
• Rainfall that is relevant for pluvial floods is represented
by a single, centrally located urban precipitation
gauge [Salzburg-Freisaal (gauge 19), see Figure 2 and
Table 1].
• Rainstorms are simulated at hourly timescales.
• The Salzach River is assumed to be the only relevant fluvial
component.
The framework is applicable to small- to medium-sized
cities where a single precipitation gauge adequately repre-
sents local rainstorms and thus the pluvial component.
For larger cities, additional precipitation gauges would
likely be required to capture the spatial variability of
precipitation.
Figure 3 Data availability in the study area (gauge IDs refer to Figure 2 and Table 1).
Table 1 Overview of all gauges in the study area
ID Name Lat Long Altitude (m.a.s.l.)
1 Gerlos 47.2269 12.0444 1250
2 Paß Thurn 47.2997 12.4222 1200
3 Hochfilzen 47.4703 12.6217 960
4 Felbertauerntunnel 47.1181 12.5056 1650
5 Böckstein 47.0872 13.1158 1140
6 Flachau 47.3472 13.3958 910
7 Golling-Torren 47.5917 13.1575 473
8 Gosau 47.5892 13.5431 765
9 Hintersee 47.7611 13.2225 750
10 Bischofswiesen-Loipl 47.6525 12.9315 845
11 Hüttschlag 47.1772 13.2314 1030
12 Glanegg 47.7494 13.0175 450
13 Enzingerboden 47.1667 12.6333 1768
14 Schmittenhöhe 47.1786 12.7381 2102
15 Zell am See 47.3267 12.7953 767
16 St. Veit 47.3292 13.1550 747
17 Dienten 47.3908 13.0333 1265
18 Hüttau 47.4131 13.3156 720
19 Salzburg-Freisaal 47.7908 13.0525 420
Table 2 Days with precipitation exceeding 25 mm/h [time series
from Salzburg-Freisaal (gauge 19), 1994–2012]
Date Hourly total (≥25 mm)
18/05/1997 29.9
12/06/1997 39.3
26/06/1998 31.4
22/07/1998 25.4
09/08/1999 25.3
05/06/2000 44.6
03/08/2000 34.2
06/08/2000 26
10/05/2003 27.8
02/05/2004 29
12/05/2004 25.5
30/05/2005 42.1
29/06/2006 31.7
06/07/2006 33.9
19/08/2007 26.6
20/06/2012 45.5
4Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
100 Breinl et al.

Overview
The framework has three major components (see Figure 4,
right scheme, grey boxes), which are (i) a stochastic WG for
daily precipitation and mean temperature, (ii) an HM to
simulate daily mean discharge from synthetic meteorological
time series and (iii) a conversion algorithm (CA) to convert
daily urban precipitation into maximum hourly urban pre-
cipitation for the pluvial hazard and to convert daily mean
discharge into daily peak discharge for the fluvial hazard.
The city of Salzburg is schematised with a dashed rectangle
in Figure 4 (left scheme). Pluvial flood days are represented
by a single urban precipitation gauge (dot with dashed
outline), and fluvial flood days are represented by the river
gauge (upside-down triangle).
First, daily precipitation and mean temperature observa-
tions of all gauge stations (1) are taken to calibrate the daily
WG (2).Next, theWG generates synthetic daily time series of
precipitation and mean temperature of any length (3) at all
gauges. The daily catchment precipitation [i.e. the weighted
synthetic precipitation of all precipitation gauges except the
urban precipitation gauge at Salzburg-Freisaal (gauge 19), or
in other words, only the discharge generating precipitation
upstream] as well as synthetic temperature time series go
into the HM (4). The HM produces synthetic daily mean
discharge (5) at the catchment outlet/urban domain. A CA
(6) is used to simultaneously convert only the daily urban
precipitation into maximum hourly urban precipitation,
and synthetic daily mean discharge into daily peak discharge,
taking into account weather situations. As a result, the
framework generates daily peak discharge and maximum
hourly urban precipitation (7) at the urban domain
(hatched rectangles) for all synthetic days simulated by the
WG (1).
Daily peak discharges (7) are used to identify fluvial flood
days. The maximum hourly urban precipitation (7) makes
the identification of torrential rainstorms and thus pluvial
flood days possible. The two continuous time series also
allow identifying not only the probability and severity of
single hazards, but also potential combined fluvial–pluvial
flood days. For both phenomena, information on critical,
i.e. hazardous, thresholds is required. The following
sections provide a detailed description of all three major
components.
Figure 4 Detailed schematics of the modelling framework with schematics of the catchment and urban domain (left) and the modelling
sequence (right). Locations of gauges do not represent the locations in reality.
Modelling framework for urban river discharge and precipitation 5
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 101

Weather Generator (WG)
For the proposed framework, a suitable multisite
semiparametric WG has been developed (Breinl et al., 2014),
including an appropriate precipitation algorithm that has
been successfully tested in different climatic environments
(see Breinl et al., 2013). In this so-called ‘RSS weather
generator’, daily snapshots of precipitation fields (i.e.
catchment-wide precipitation patterns) called ‘amount
vectors’ are first clustered by k-means clustering into similar
classes, which represent similar type of patterns. These
classes are then simulated with a univariate Markov process,
removing the need for individual Markov models at each
site, as for example suggested by Wilks (1998) or Brissette
et al. (2007). Once time series of classes (i.e. time series of
precipitation patterns) are simulated, each single class is
replaced by a randomly drawn observed ‘amount vector’
matching the simulated class and month. To generate unob-
served extremes, precipitation amounts are randomly drawn
from parametric distributions (at each site separately) and
reshuffled according to ranks of the resampled amounts to
maintain temporal and spatial correlations. For more details
on the algorithm, the reader is referred to the publications
mentioned above. In Breinl et al. (2014), the WG was set up
for the Salzach catchment to simulate daily precipitation at
19 sites and daily mean temperature at 3 sites. Other than
developed in Breinl et al. (2014), we replaced the precipita-
tion gauge ‘Salzburg Airport’ by the precipitation gauge
‘Salzburg-Freisaal’ (gauge 19), which is closer to the city
centre, therefore better representing urban precipitation.
Hydrological Model (HM)
Model choice
Probabilistic analyses require multiple modelling runs to
reliably estimate uncertainty. Fully-distributed models can
simulate physical processes but are often computationally
expensive and data demanding (Arnold et al., 1998;
Romanowicz et al., 2005; Shrestha et al., 2006). Other chal-
lenges include uniqueness or equifinality (Beven, 2001).
Furthermore, a higher model resolution does not necessarily
mean a better model performance (Das et al., 2008). For this
reason, the fast conceptual model Hydrologiska Byråns
Vattenbalansavdelning (HBV), originally developed for
Scandinavian catchments by Bergström (Bergström, 1976,
1992, 1995), was chosen.HBV has been applied in numerous
studies (e.g. Uhlenbrook et al., 1999; Bergström et al., 2001;
Booij, 2005; Hagg et al., 2006; Bardossy and Das, 2008; Das
et al., 2008; Steele-Dunne et al., 2008; Gao et al., 2012; Beck
et al., 2013) and exists in a variety of versions. We used
HBV-light (Seibert, 1997, 1999,2000; Seibert and Vis, 2012),
which can simulate catchments utilising a lumped or
semidistributed setup. HBV-light simulates daily mean dis-
charge based on daily precipitation, daily mean temperature
and potential monthly evaporation. The model essentially
consists of four main components: a snow routine, a soil
moisture routine, a response routine and a routing routine.
Detailed descriptions of HBV are provided by Bergström
(1992,1995), Lindström et al. (1997) or Seibert (1997).
Model setup
We used HBV-light with three instead of two groundwater
boxes. A lumped model version with the maximum possible
number of 10 different elevation zones turned out to
perform as well as a semidistributed model version with
three subcatchments, but was faster. To reduce the risk of
overparameterisation, fixed values were used for the refreez-
ing coefficient CFR (0.05) and Water Holding Capacity
CWH (0.1) (J. Seibert, 2013, pers. comm.). The potential
monthly evaporation was calculated according to
Thornthwaite (1948). Precipitation values and heights of
all precipitation gauges (gauges 1–18) were weighted by
Thiessen polygons. For the temperature as a continuous phe-
nomenon, the arithmetic mean of the three gauges (15, 16
and 19) was used. The so-called‘GAP’ optimisation (Seibert,
2000) was applied in the calibration, which consists of a
genetic algorithm (GA) combined with Powell’s (P) quad-
ratically convergent method (Press et al., 2002) for subse-
quent fine tuning.
Model calibration
In conceptual models, different parameter sets can produce
similarly good results (‘model equifinality’, e.g. Beven and
Binley, 1992; Seibert, 1997; Beven, 2001).For this reason, the
GAP optimisation routine was used to generate 100 suitable
parameter sets. The calibration period (2000–2010) was
chosen as it comprises the extreme flood of August 2002.The
validation period was 1988–1997. The warm-up period was
1 year. Physically reasonable ranges of the model variables
were derived from different publications (Seibert, 1997,
2000; Seibert and Vis, 2012) and are listed in Table 3. Table 3
also gives the meaning of the variables, their units as well as
the corresponding model routines.
The final ranges of the 100 calibrated parameter sets are
listed in Table 4.
The widely applied Nash-Sutcliffe (NSE) goodness-of-fit
criterion (Nash and Sutcliffe, 1970) for model calibration
turned out to slightly underestimate extreme flows and over-
estimate winter flows, leading to bias in the water balance.
Two other objective functions were considered: whereas the
mean absolute residual error (MARE) [Eqn (1)] led to a
more ‘flashy’ calibration with a tendency to overestimate
smaller peaks and the flow variability, the lnNSE [Eqn (2)]
6Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
102 Breinl et al.

performed better than NSE with respect to spring and winter
flows but turned out to underestimate major flood peaks.
MARE =− −
∑
11
n
QQ
Q
obs sim
obs
(1)
ln lnQ lnQ
lnQ lnQ
n
obs sim
obs obs
NSE
where is the num
=− −
()
−
()
∑
∑
1
2
2
bber of observation days.
(2)
An equally weighted objective function f [Eqn (3)] of
the lnNSE and MARE turned out to be a reasonable
compromise.
fNSEMARE=∗ +∗05 05..ln (3)
Calibration and validation results are shown in Table 5.
The model performance appears to be satisfactory, which
can be explained by the dense precipitation gauge network
(Bardossy and Das, 2008). Figure 5 shows the calibration
(validation) results for the years with major peaks 2002
(1995) using the objective function consisting of lnNSE and
MARE [Eqn (3)], and the often applied NSE (see above).
Although the objective function shows similar performance
in the calibration period compared with NSE, differences are
more pronounced in the validation period, where NSE over-
estimates winter flows and tends to slightly underestimate
peaks. Results were comparable for other years of the cali-
bration and validation period. The bias in the simulation of
lower flows can be explained by the fact that NSE places
more emphasis on higher flows (by computing the differ-
ences between observations and predictions with squared
values) and less on lower flows (Krause et al., 2005). In
general, HBV-light tends to underestimate smaller winter
peaks, which has also been found by others (e.g.
Steele-Dunne et al., 2008).
The generally good simulation results indicate a reliably
calibrated HM. In the entire framework, the mean discharge
produced by the 100 parameter sets was used for analyses.
Conversion Algorithm (CA)
Daily mean discharge and daily urban precipitation are of
limited suitability for the characterisation of fluvial and
pluvial floods. Thus, an algorithm was used to convert daily
mean discharge into more suitable daily peak discharge and
daily urban precipitation into maximum hourly urban pre-
cipitation. A k-nearest neighbour algorithm (kNN) was
combined with the MOF. Nearest neighbour algorithms have
been widely applied in stochastic weather generation (e.g.
Brandsma and Buishand, 1997; Brandsma and Buishand,
Table 3 Ranges of HBV variables used in the GAP optimisation
Parameter Parameter (unit) Unit Model routine Min Max
TT Threshold temperature °C Snow routine −1.5 2.5
CFMAX Degree-day factor mm/°C/day Snow routine 1 10
SFCF Snowfall correction factor – Snow routine 0.4 1
CWH* Water holding capacity – Snow routine 0.1 0.1
CFR* Refreezing coefficient – Snow routine 0.05 0.05
FC Maximum soil moisture storage mm Soil moisture routine 50 600
LP Soil moisture value above which AET reaches PET – Soil moisture routine 0.3 1
BETA Parameter that determines the relative
contribution to runoff from rain or snowmelt
– Soil moisture routine 1 6
Cet Potential evaporation correction factor °C Soil moisture routine 0 0.5
K0 Storage (or recession) coefficient 0 mm/day Response routine 0.1 0.5
K1 Storage (or recession) coefficient 1 mm/day Response routine 0.01 0.2
K2 Storage (or recession) coefficient 2 mm/day Response routine 0.0001 0.1
UZL Threshold parameter mm Response routine 0 50
PERC Threshold parameter mm/day Response routine 1 6
MAXBAS Length of triangular weighting function Day Routing routine 1 3
*Fixed values to reduce the risk of over-parameterisation.
Table 4 Ranges of variables of the final 100 parameter sets
Parameter Mean Max Min
TT −0.53 −0.01 −1.17
CFMAX 5.15 6.59 2.80
SFCF 0.46 0.50 0.44
CWH* 0.10 0.10 0.10
CWR* 0.05 0.05 0.05
FC 454.30 599.99 142.55
LP 0.33 0.51 0.30
BETA 2.68 6.00 1.11
CET 0.25 0.38 0.17
K0 0.41 0.50 0.36
K1 0.06 0.08 0.05
K2 0.00 0.01 0.00
UZL 16.72 21.91 14.68
PERC 2.86 3.53 2.69
MAXBAS 2.55 2.61 2.42
*Fixed values to reduce the risk of over-parameterisation.
Modelling framework for urban river discharge and precipitation 7
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 103

1998; Rajagopalan and Lall, 1999; Sharma and Lall, 1999;
Buishand and Brandsma, 2001; Sharif and Burn, 2007;
Leander and Buishand, 2009; King et al., 2012) or in simu-
lating hydrological time series (e.g. Lall and Sharma, 1996;
Shamseldin and OConnor, 1996). Nearest neighbour algo-
rithms typically involve selecting a specified number of data
vectors similar in characteristics to the vector of interest
(Sharif and Burn, 2007). The similarity is typically computed
by means of a distance measure (e.g. Euclidean distance).
One of the selected vectors is randomly resampled and used
to build a new time step in the simulation period. The
MOF has been used in precipitation disaggregation (e.g.
Maheepala and Perera, 1996; Srikanthan and McMahon,
2001; Wojcik and Buishand, 2003; Pui et al., 2009). The idea
behind the MOF is as follows (Srikanthan and McMahon,
2001): hourly precipitation observations (or any other
observations of high resolution) are standardised day by day
(or by another low-resolution unit) so that the sum of the
hourly precipitations on any day equals unity. This pro-
cedure gives nsets of fragments of hourly precipitations
fromarecordofndays. The generated synthetic daily pre-
cipitations are disaggregated by selecting a set of observed
fragments (in this case the selection is conducted using
nearest neighbour resampling) and multiplying the syn-
thetic daily precipitations by each of the hourly fragments to
produce synthetic hourly rainfalls.
Table 5 Goodness of fit results for Nash-Sutcliffe (NSE), lnNSE, mean absolute residual error (MARE) and the objective function (0.5 *
lnNSE +0.5 * MARE). The results are shown for the calibration and validation period
Period Time (10a)
NSE
(mean)
NSE
(max)
NSE
(min)
lnNSE
(mean)
lnNSE
(max)
lnNSE
(min)
MARE
(mean)
MARE
(max)
MARE
(min)
Obj. fct.
(mean)*
Obj. fct.
(max)*
Obj. fct.
(min)*
Calibration 01/01/2001–31/12/2010 0.86 0.87 0.82 0.88 0.89 0.87 0.86 0.86 0.85 0.87 0.87 0.86
Validation 01/01/1988–31/12/1997 0.74 0.76 0.68 0.81 0.82 0.80 0.82 0.82 0.81 0.82 0.82 0.80
*The objective function was used to calibrate HBV-light in the GAP optimisation routine.
Figure 5 HBV-light simulation results for 1 year of the 10-year calibration period (2002) and 1 year of the 10-year validation period
(1995), using the objective function [Eqn (3)] and Nash-Sutcliffe. Observations are shown in black, means of simulations are shown in red
and grey areas denote the 5th and 95th percentiles of simulations.
8Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
104 Breinl et al.

More specifically, the algorithm of this research works as
follows. First, in a kNN algorithm, each simulated synthetic
day is compared with an observed vector of daily mean
discharge, daily urban precipitation and daily catchment
precipitation (weight of all precipitation gauges except
gauge 19, see section Overview). A historical day where all
three variables are similar to those of the synthetic day and
for which hourly urban records exist is then selected. This
historical day is used in a second step to convert the syn-
thetic daily urban precipitation into a daily maximum
hourly value and daily mean discharge into daily peak dis-
charge (further details of algorithm given later in this
section). It is valid to assume that every combination of the
three variables represents specific atmospheric conditions.
This can be illustrated with two exemplary days: on the 10th
of May 2003, a maximum hourly urban precipitation sum of
27.8 mm was measured in Salzburg-Freisaal (gauge 19),
whereas the daily catchment precipitation (5.0 mm) as well
as daily mean discharge in the city of Salzburg (285.0 m3/s)
were comparatively low. The dominating large-scale circu-
lation pattern during this time was a high-pressure bridge
over Central Europe [see ‘Catalogue of Großwetterlagen
(circulation patterns) in Europe (1881–2009)’, Gerstengarbe
and Werner, 2010]. The daily urban precipitation sum in
Salzburg-Freisaal (gauge 19) was 33.6 mm, meaning that
82.8% of the precipitation (27.8/33.6 mm) fell within 1 h in
the early evening [19:50 Central European Time (CET) to
20:50 CET]. It is valid to assume that this precipitation event
in Salzburg was local and convective in nature. On the 7th of
August 2002, the maximum measured hourly precipitation
sum at Salzburg-Freisaal (gauge 19) was 2.9 mm. The daily
catchment precipitation (20.5 mm) as well as daily mean
discharge in the city of Salzburg (1046 m3/s) were compara-
tively high. The daily urban precipitation sum in Salzburg-
Freisaal (gauge 19) was also 20.5 mm, meaning that only
14.1% (2.9/20.5 mm) of the daily precipitation fell within
1 h in the late afternoon (17:10 CET to 18:10 CET). During
this time, the dominating large-scale circulation pattern
was a low-pressure system over Central Europe (see
Gerstengarbe and Werner, 2010), followed by a‘Vb’ weather
situation (see section Hydrology and meteorology and
Gerstengarbe and Werner, 2010) on the 9th of June that
caused wide-spread flooding across Europe. That is, the pre-
cipitation was likely frontal in nature. The ‘Vb’ weather
situation led to a peak discharge of 2289 m3/s on the 12th of
August in the city of Salzburg.
These two exemplary days show that weather situations
are inherently considered in the CA; the conversion of dis-
charge and precipitation is conditioned on days where the
synthetic as well as observed daily vectors of the three vari-
ables are similar.Technically,the kNN algorithm (first step of
CA) works as follows: take the vector Xof the three variables
of daily mean discharge, daily urban precipitation and daily
catchment precipitation on any synthetic day, e.g. the 1st of
January.
Xxxx t T
tttt
=
()
∀=
{}
,, ,,
,,,123 12…(4)
1. Choose all potential neighbours L from the daily obser-
vations from the previous and following zdays and from
all Nyears including the synthetic observation day,so that
L=N*(z*2 +1). That is, in case of z=10, all observed
days from the 22nd of December until the 11th of January
are taken into account.
Calculate the Manhattan distance between the synthetic
vector of the current day Xtand the observed vector Xlfor
day i,wherei=1,...,L
dXX
iti
=− (5)
Extract the k neighbours and compute a Kernel density
estimator that assigns higher probabilities to smaller devia-
tions according to Lall and Sharma (1996). For the jneigh-
bours, weights are calculated as follows:
wj
i
j
i
k
=
=
∑
1
1
1
(6)
with the cumulative probabilities pjdefined by
pw
ji
i
j
=
=
∑
1
(7)
Compute a uniform random number U(0,1) and compare
with pj. Choose wjwhere U is closest to pj.
The final disaggregation (MOF, i.e. second step of CA)
works as follows:
Compute the synthetic peak discharge ′
xt1, on day tfrom
the synthetic mean discharge x1,t as well as relationship
between the observed mean x1,land peak discharge ′
xl1, by
′=′
xx
xx
t
l
l
t1
1
1
1,
,
,
,(8)
Compute the synthetic maximum hourly urban precipi-
tation value ′
xt2, on day tfrom the synthetic daily urban
precipitation value x2,t as well as relationship between the
observed daily urban precipitation x2,1 and observed
maximum hourly precipitation x2,1 by
′=′
xx
xx
t
l
l
t2
2
2
2,
,
,
,(9)
The CA algorithm is illustrated in Figure 6 for one exem-
plary synthetic day. The synthetic day (top left) does not
provide maximum hourly urban precipitation as well as
peak discharge. In this exemplary case, two nearest neigh-
bours (right block) are available. If the first nearest neigh-
bour is drawn, the maximum hourly precipitation is
2.1/11.3 mm*12.7 mm =2.4 mm. The peak discharge is
Modelling framework for urban river discharge and precipitation 9
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 105

calculated by 336.0/290.0 m3/s*277.0 m3/s =320.9 m3/s. If
the second nearest neighbour is drawn, the maximum
hourly precipitation is 1.6/12.6 mm*12.7 mm =1.6 mm.
The peak discharge is then calculated by 323.0/276.0 m3/
s*277.0 m3/s =324.2 m3/s.
The choice of the number of nearest neighbours is often
heuristic. Sensitivity tests revealed that a high number of k
leads to a bias of hourly extremes, which, in the average,
significantly exceeded the observations. Thus, a low value of
kis recommended. For example, Pui et al. (2009) used k=1.
We conducted different modelling runs with 1 ≥k≤2 and a
varying window length z={10, 12, . . . 50}. However, even in
case of k=1, variability is introduced by the varying (para-
metric) synthetic daily precipitation amounts from the WG.
Definition of critical flood thresholds
To examine combined fluvial–pluvial flood days, critical
thresholds for river discharge (fluvial flooding) and extreme
precipitation (pluvial flooding) had to be defined. Flood
defence in the city of Salzburg provides protection up to
Q100 (Loizl, 2012a). The threshold for fluvial days was thus
set at the simulated Q100 value. Loss information from a
local insurer and operation records from the Salzburg fire
service helped to derive a suitable pluvial (=precipitation)
threshold. The insurance data set contains reported flood-
related losses (‘claims’) in the Salzburg city area reported by
clients from 2000 to 2013. It contains dates, flood sources,
monetary losses, types of buildings as well as addresses. Only
the years 2000–2012 were analysed due to unavailable pre-
cipitation data for 2013. Figure 7 shows the results for four
precipitation intensity classes. Most events (149) are related
to low intensities, only four had precipitation intensities
higher than 30 mm/h. Although a high percentage of mon-
etary losses is caused by high-intensity precipitation days
(26.54%), the highest percentage is related to low intensities
(47.16%). During low-intensity days in class 1 with
intensities between 0.1 and 10 mm/h, on average 1.21 claims
were reported.
Numbers of reported claims per event do not vary signifi-
cantly with increasing intensities (2.24 and 3.43 per event in
class 2 and 3), but a significant increase of claims occurs with
high intensities >30 mm/h (23.5 per event in class 4). This is
almost six times more (+585%) than in the third class (20–
30 mm). Days with more than 30 mm/h thus seem to be
most critical. This threshold is in line with results from other
studies (see section Hydrology and meteorology).
Data from the Salzburg fire service contain the number of
calls on heavy precipitation days (e.g. flooded basements).
For technical reasons, only information on days with pre-
cipitation intensities >15 mm/h could be provided. The
number of calls increases with increasing precipitation
(Figure 8). The most significant change can be detected
between the third (31.43) and second (9.04) classes
(+248%). Thus, 20 mm/h appears to be critical, but the
threshold is not as defined as in the insurance data set.
Figure 6 The conversion algorithm explained by one exemplary synthetic day and two nearest neighbours from the observations.
Figure 7 Insurance data: number of claims per date, number of
dates and total loss, shown for four different precipitation inten-
sity classes.
10 Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
106 Breinl et al.

Facing two different thresholds of critical hourly precipi-
tation (20 and 30 mm), probabilistic analyses were con-
ducted assuming hourly thresholds of 20, 25 and 30 mm.
Validation of framework
The modelling framework was validated for (i) the fluvial
component, (ii) the pluvial component and (iii) combined
fluvial–pluvial flood days, using various metrics (Table 6).
For the analyses, 1000 synthetic time series of 5 times the
historical record length (120 years) were generated, which
corresponds to 120 000 synthetic years. As 1 year of each of
the 1000 time series was used as the warm-up period in HBV,
the final synthetic output was 119 000 years. The CA was run
42 times (two values for k, 21 values for z, section CA).
Results and discussion
Fluvial component (peak discharge)
The fluvial component is examined by analysing the mean
peak discharge, the timing of the maximum annual (AMAX)
hourly precipitation and by a frequency analysis of
maximum hourly precipitation (AMAX). The observed and
simulated mean monthly peak discharge is shown in
Figure 9 (upper left panel). There is a tendency to slightly
overestimate summer (+1.9% on average in June, July and
August) and slightly underestimate winter discharge (−6.3%
on average in December, January and February).
The timing of the observed and simulated annual peak of
discharge (AMAX) is shown in Figure 9 (upper right panel).
Although simulations generate more extremes in May and
fewer in July, one has to recall that only 35 AMAX observa-
tions were available for the validation. The frequency of
AMAX in late spring and over the summer is more evenly
distributed in the simulations. This matches well with the
observed peak discharge in Figure 9 (upper left panel).
Figure 9 (bottom panel) compares the model performance
with respect to the frequency of extremes (AMAX). Extreme
Value type 1 (EV1) plotting positions show that all observa-
tions are within the simulations.
The calculated fluvial flood quantiles (Table 7) have been
benchmarked with various publications. Klasinc et al. (2004)
and Loizl (2012b) give information on the Q100 year flood
in Salzburg, which ranges between 2200 and 2300 m3/s. Loizl
(2012b) also provides information on the 1/5 year flood
(‘about 1350 m3/s’). Thus,the literature indicates reasonable
simulations, especially as the methods used by these authors
are not explained in detail.
Pluvial component (maximum hourly
precipitation)
The pluvial component is examined by analysing the timing
of the maximum annual hourly precipitation (AMAX) and a
frequency analysis of maximum hourly precipitation
Table 6 Metrics to validate the modelling framework. Validation is conducted for the fluvial and pluvial component as well as their
relationship (i.e. combined fluvial–pluvial days)
Component Metrics Reason
Fluvial Mean peak discharge Valid water balance
Occurrence of AMAX (peak discharge) Valid timing of fluvial floods
Plotting positions of AMAX (peak discharge) Valid frequency of fluvial floods
Pluvial Occurrence of AMAX (hourly maxima) Valid timing of pluvial events
Plotting positions of AMAX (hourly maxima) Valid frequency of pluvial events
Fluvial–pluvial Inter-variable and temporal correlations
(variables are listed in Table 9)
Valid simulation of relationship between fluvial and pluvial
component (i.e. combined fluvial–pluvial events)
Figure 8 Operation records from the Salzburg fire service:
number of calls per date and number of dates are shown for three
different precipitation intensity classes.
Table 7 Simulated fluvial return periods (RP) (m3/s) of all simula-
tion runs: mean, 5% (prct5) and 95% (prct95) percentiles
RP prct5 Mean (sim) prct95
2 957.91 996.37 1036.06
5 1250.68 1317.03 1390.10
10 1441.22 1529.34 1629.71
20 1624.16 1732.98 1857.59
50 1856.34 1996.59 2156.57
100 2029.99 2194.12 2381.76
Modelling framework for urban river discharge and precipitation 11
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 107

(AMAX). As Figure 10 (upper panel) shows, the CA
framework reproduces the timing of extreme precipitation
well.
Some deviations are obvious; however, the observed plot
is based on only 19 observations of AMAX and there is, for
example, no reasonable explanation for a drop of the fre-
quency in July but short observed time series. The frame-
work simulates comparable frequencies of extreme
precipitation over summer, which is reasonable. The model
also performs satisfactorily with respect to the frequency of
extremes (Figure 10, lower panel).
Benchmark figures for hourly precipitation could be
obtained from the online portal ‘ehyd’, run by the Federal
Ministry of Agriculture in Austria (Lebensministerium,
2013). The portal provides information on design storm
depths (in millimetre) for grid points (6 ×6 km) for all
Austria, calculated by a weighting method of the output of a
convective precipitation model and statistical analysis of
observation data. Information is available for return periods
of up to 100 years.
Simulation results and values from the‘ehyd’ portal match
well (Table 8), which indicates reasonable simulation results.
Combined fluvial–pluvial flood days
To validate the ability of the framework to simulate com-
bined flood days, the observed and simulated intervariable
and temporal correlations of multiple variables (see Table 9)
have been examined.
The intervariable correlations (Pearson) between all com-
binations of variables/time series (Table 9) are shown in
Figure 11 (upper panel) for the observations and the mean
of all simulations.
Figure 9 Simulated and observed mean peak discharge (m3/s) in all months (upper left panel), timing of observed and simulated annual
peak discharge (AMAX) (upper right panel) and frequency analysis of observed and simulated annual peaks (AMAX) (bottom panel). Dots
show plotting positions of observations, crosses the mean of the simulations. The grey area denotes the 5th and 95th percentiles of
simulations.
Table 8 Simulated return periods (RP) of hourly precipitation
(mm) of all simulation runs: mean, 5% (prct5) and 95% (prct95)
percentiles of simulations and recommended hourly design storm
depths by the Federal Ministry of Agriculture (ehyd online portal)
RP prct5 Mean (sim) prct95 ehyd
2 22.22 24.36 27.10 24.80
5 29.58 33.47 39.26 33.10
10 34.29 39.51 47.57 39.10
20 38.74 45.30 55.61 45.20
50 44.64 52.79 66.06 53.00
100 49.04 58.40 73.85 59.00
12 Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
108 Breinl et al.

The correlation structure between all variables is fairly
well preserved. The average deviation is +1.5% with the
highest deviation of +12.1% between variable 2 and 5 and
the lowest between variable 3 and 4 (−0.4%). For the lag1-
autocorrelation, the average deviation is +3.6%. The highest
deviation is related to the simulated daily urban precipita-
tion (variable 2) with −20.2%, which is an inherent issue of
the vast majority of stochastic WGs (Breinl et al., 2014). The
lowest deviation is related to maximum hourly urban pre-
cipitation (+4.5%).
Daily peak discharges and maximum hourly urban pre-
cipitation were compared with respect to combined flood
days. The critical threshold for fluvial flood days was set at
the simulated Q100 value (2194.12 m3/s, Table 7) and three
different thresholds of 20, 25 and 30 mm/h were used for
pluvial flood days (section Definition of critical flood
thresholds). In the analysis, the window of combined flood
days was varied from 1 (critical fluvial and pluvial thresh-
olds on the same day) up to 121 (60 days prior and after
day of interest). As a wide window can comprise more than
one fluvial or pluvial day, a combined flood within any
width of the window is defined as an event where at least
one fluvial and one pluvial day were detected. Although
such a wide window does not represent a ‘combined’ flood
in its actual sense, analyses are useful for the fire service or
the civil protection in regard to preparedness and logistics.
Table 10 summarises the simulation results of combined
floods.
Combined floods on the same day are unlikely and have a
high average return period, ranging from 2542.22 (assuming
20 mm) to 6239.70 years (assuming 30 mm). Critical thresh-
olds within 1 week are more likely. All calculated return
periods have large confidence intervals. Moreover, return
periods strongly vary with the assumption on the critical
threshold of pluvial floods. For example, combined floods
on the same day have a 2.5 times higher return period when
assuming 30 mm as critical, compared with 20 mm (6239.70
versus 2542.22 years). The uncertainties with respect to criti-
cal precipitation thresholds have to be considered by risk
managers.
Figure 12 (left panel) analyses the catchment precipitation
up to 10 days prior to the actual flood day for fluvial, pluvial
and combined flood days, averaged for all simulations. The
analysis was conducted for a pluvial threshold of 25 mm.
Fluvial flood days are caused by increased precedent
Figure 10 Simulated and observed timing of observed and simu-
lated annual maximum hourly precipitation (AMAX) (upper
panel) and frequency analysis of observed and simulated
annual maxima of hourly precipitation (lower panel). Dots show
plotting positions of observations, crosses the mean of the simu-
lations. The grey area denotes the 5th and 95th percentiles of
simulations.
Table 9 Variables and corresponding IDs used in the correlation
analysis
Variable ID
Daily catchment precipitation 1
Daily urban precipitation 2
Maximum hourly urban precipitation 3
Daily mean discharge 4
Daily peak discharge 5
Figure 11 Upper panel: Observed and simulated intervariable cor-
relation (Pearson) between all possible combinations of variables
(Table 9). Lower panel: Lag-1 autocorrelation of variables in
Table 9. White bars denote the observation; black bars the mean
of all simulations.
Modelling framework for urban river discharge and precipitation 13
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 109

(frontal) precipitation over several days. Pluvial flood days
are related to increased (convective) precipitation up to 2
days prior to the event. Combined flood days are related to
weather situations similar to the ones causing fluvial floods
and may be caused by high-intensity rainfall cells within
frontal systems, as described by Houston et al. (2011).
Fluvial flood days have a certain delay related to the peak of
precipitation, which represents the catchment response time.
The same delay can be observed for combined flood days.
The catchment delay also explains the low probability of
combined floods on the same day, as the probability of com-
bined floods significantly increases with only a small increase
of the observation window (see Table 10). The right panel of
Figure 12 shows the same analysis, but conducted for the
simulated daily urban precipitation. Results are similar. In
case of combined flood days, urban precipitation amounts
however further increase until the actual flood day. On
pluvial flood days, catchment precipitation amounts are on
average only 4.4 times higher than the annual average of
catchment precipitation. Daily urban precipitation amounts
are however 12 times higher than the annual urban average.
This indicates the local convective character of precipitation
cells leading to urban pluvial floods.
Summary and conclusion
This paper presents a joint probabilistic modelling frame-
work to assess the hazard from fluvial and pluvial flood days
for the city of Salzburg. The basic modelling framework runs
at daily time steps to overcome data limitations and facilitate
its transferability. The framework consists of a WG, the HM
HBV-light and a CA consisting of a kNN and the MOF. The
CA converts daily mean discharge into peak discharge and
daily urban precipitation into maximum hourly precipita-
tion, which allows identifying fluvial, potentially pluvial and
combined flood days. The CA takes into account vectors of
daily mean discharge, daily urban precipitation and daily
catchment precipitation, which are treated as a proxy for
weather situations. Combined floods are rare, particularly
when occurring on the same day. Insurance data and opera-
tion records of the fire department indicate a critical precipi-
tation threshold in Salzburg between 20 and 30 mm/h,
which is in line with other studies. The critical hourly thresh-
old defining extreme precipitation as ‘pluvial’ has a large
impact on the simulated probabilities. The analysis of com-
bined flood days has been conducted assuming three differ-
ent thresholds (20, 25 and 30 mm/h). The framework is
Table 10 Return periods of combined fluvial–pluvial flood days, assuming three different thresholds for pluvial days (5th percentile,
mean and 95th percentile). A window of 1 means combined events on the same day; a window of 7 days means at least one fluvial and
pluvial event within 1 week etc.
Window
20-mm threshold 25-mm threshold 30-mm threshold
prct5 Mean prct95 prct5 Mean prct95 prct5 Mean prct95
1 1461.92 2542.22 6611.11 2468.88 4684.16 23 800.00 3067.01 6239.70 29 750.00
7 299.15 403.78 901.52 417.54 660.41 3541.67 512.05 911.88 7628.21
31 215.03 275.65 467.03 323.02 446.65 1101.85 418.13 639.62 1876.97
61 180.91 226.00 327.46 269.35 355.48 658.91 370.26 526.16 1131.18
121 147.39 178.83 228.14 206.74 262.38 393.26 296.02 384.99 610.26
Figure 12 Left panel: Analysis of catchment precipitation for up to 10 days prior to the actual flood day (left panel) for the three flood
types. Right panel: Analysis of daily urban precipitation for up to 10 days prior to the actual flood day for the three flood types.
14 Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
110 Breinl et al.

designed for urban rainstorms at hourly resolution, but
could be adapted to simulate rainstorms at finer scales. This
would however complicate the definition of critical pluvial
precipitation thresholds, which are usually defined at hourly
resolution. In future research, the framework could be
coupled with a hydraulic model to translate the simulated
discharge and rainfall into inundated areas and monetary
risk.
Acknowledgements
This research has been funded by the EU 7th Framework
Marie Curie ITN project ‘CHANGES’ under Grant Agree-
ment No. 263953. Observation data were provided by the
Central Institution for Meteorology and Geodynamics
(ZAMG), the Federal Ministry of Agriculture, Forestry, Envi-
ronment and Water Management and the German Meteoro-
logical Service (DWD). The authors would like to thank
Allianz-Elementar in person of Rupert Pichler and Claudia
Hutterer for providing the insurance data as well as the Salz-
burg fire service in person of fire director Eduard Schnöll
and his colleagues for providing the operation records. Com-
ments by Markus Stowasser and Thea Turkington are appre-
ciated. The authors would like to thank the anonymous
reviewers for their helpful and constructive comments. The
paper expresses the view of the authors and not the Austrian
and German organs of government.
References
Apel H., Thieken A.H., Merz B. & Blöschl G. Flood risk assess-
ment and associated uncertainty. Nat Hazards Earth Syst Sci
2004, 4, (2), 295–308.
Apel H., Thieken A., Merz B. & Blöschl G. A probabilistic mod-
elling system for assessing flood risks. Nat Hazards 2006, 38,
(1), 79–100.
Apel H., Aronica G., Kreibich H. & Thieken A. Flood risk analy-
ses – how detailed do we need to be? Nat Hazards 2009, 49,
(1), 79–98.
Arnold J.G., Srinivasan R., Muttiah R.S. & Williams J.R. Large
area hydrologic modeling and assessment – part 1: model
development. J Am Water Resour Assoc 1998, 34, (1), 73–89.
Bardossy A. & Das T. Influence of rainfall observation network
on model calibration and application. HydrolEarthSystSci
2008, 12, (1), 77–89.
Bates P.D., Horritt M.S. & Fewtrell T.J. A simple inertial formu-
lation of the shallow water equations for efficient two-
dimensional flood inundation modelling. JHydrol2010, 387,
(1–2), 33–45.
Beck H.E., Bruijnzeel L.A., van Dijk A.I.J.M., McVicar T.R.,
Scatena F.N. & Schellekens J. The impact of forest regenera-
tion on streamflow in 12 mesoscale humid tropical catch-
ments. HydrolEarthSystSci2013, 17, (7), 2613–2635.
Bergström S. Development and application of a conceptual runoff
model for Scandinavian catchments. Lund: Dept. of Water
Resour. Engineering, Lund Inst of Technol./Univ. of Lund,
1976. Bull. Ser. A.
Bergström S. The HBV model: its structure and applications.
Norrköping: Swedish Meteorological and Hydrological Insti-
tute (SMHI), Hydrology, 1992.
Bergström S. The HBV model (Chapter 13). In: V.P. Singh, ed.
Computer models of watershed hydrology. Highlands Ranch,
CO: Water Resources Publications, Highlands Ranch, 1995,
pp. 443–476.
Bergström S., Carlsson B., Gardelin M., Lindström G.,
Pettersson A. & Rummukainen M. Climate change impacts
on runoff in Sweden – assessments by global climate models,
dynamical downscaling and hydrological modelling. Clim Res
2001, 16, (2), 101–112.
Beven K. How far can we go in distributed hydrological model-
ling? HydrolEarthSystSci2001, 5, (1), 1–12.
Beven K. & Binley A. The future of distributed models – model
calibration and uncertainty prediction. Hydrol Process 1992,
6, (3), 279–298.
Blanc J., Hall J.W., Roche N., Dawson R.J., Cesses Y., Burton A.
& Kilsby C.G. Enhanced efficiency of pluvial flood risk esti-
mation in urban areas using spatial–temporal rainfall
simulations. J Flood Risk Manage 2012, 5, (2),
143–152.
Booij M.J. Impact of climate change on river flooding assessed
with different spatial model resolutions. JHydrol2005, 303,
(1–4), 176–198.
Bradbrook K., Waller S. & Morris D. National floodplain
mapping: datasets and methods – 160 000 km in 12 months.
Nat Hazards 2005, 36, (1), 103–123.
Bradford R.A., O’Sullivan J.J., van der Craats I.M., Krywkow J.,
Rotko P., Aaltonen J., Bonaiuto M., De Dominicis S., Waylen
K. & Schelfaut K. Risk perception – issues for flood manage-
ment in Europe. Nat Hazards and Earth System Sciences 2012,
12, (7), 2299–2309.
Brandsma T. & Buishand T.A. Rainfall generator for the Rhine
basin: single-site generation of weather variables by nearest-
neighbour resampling. De Bilt: KNMI, 1997. KNMI-
publication 186-I.
Brandsma T. & Buishand T.A. Simulation of extreme precipita-
tion in the Rhine basin by nearest-neighbour resampling.
HydrolEarthSystSci1998, 2, (2–3), 195–209.
Breinl K., Turkington T. & Stowasser M. Stochastic generation
of multi-site daily precipitation for applications in risk man-
agement. JHydrol2013, 498, (0), 23–35.
Breinl K., Turkington T. & Stowasser M. Simulating daily pre-
cipitation and temperature: a weather generation framework
for assessing hydrometeorological hazards. Meteorol Appl
2014.
Brissette F.P., Khalili M. & Leconte R. Efficient stochastic gen-
eration of multi-site synthetic precipitation data. JHydrol
2007, 345, (3–4), 121–133.
Modelling framework for urban river discharge and precipitation 15
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 111

Bronstert A. River flooding in Germany: influenced by climate
change? Phys Chem Earth 1995, 20, (5–6), 445–450.
Buishand T.A. & Brandsma T. Multisite simulation of daily pre-
cipitation and temperature in the Rhine Basin by nearest-
neighbor resampling. Water Resour Res 2001, 37, (11),
2761–2776.
Büchele B., Kreibich H., Kron A., Thieken A., Ihringer J., Oberle
P., Merz B. & Nestmann F. Flood-risk mapping: contribu-
tions towards an enhanced assessment of extreme events and
associated risks. Nat Hazards Earth Syst Sci 2006, 6, (4), 485–
503.
Chen A.S., Djordjevic´ S., Leandro J. & Savic´ D.A. An analysis of
the combined consequences of pluvial and fluvial flooding.
Water Sci Technol 2010, 62, (7), 1491–1498.
Das T., Bardossy A., Zehe E. & He Y. Comparison of conceptual
model performance using different representations of spatial
variability. JHydrol2008, 356, (1–2), 106–118.
Dawson R.J., Speight L., Hall J.W., Djordjevic S., Savic D. &
Leandro J. Attribution of flood risk in urban areas.
J Hydroinform 2008, 10, (4), 275–288.
Driessen T.L.A., Hurkmans R.T.W.L., Terink W., Hazenberg P.,
Torfs P.J.J.F. & Uijlenhoet R. The hydrological response of the
Ourthe catchment to climate change as modelled by the HBV
model. HydrolEarthSystSci2010, 14, (4), 651–665.
Falconer R., Smyth P. & Maani L. Pluvial extreme events risk
appraisal techniques with recent applications in Ireland and
the UK. In: Irish National Hydrology Conference 2008. Dublin:
Irish National Committees for IHP and ICID, 2008, pp.
43–52.
Falconer R.H., Cobby D., Smyth P., Astle G., Dent J. & Golding
B. Pluvial flooding: new approaches in flood warning,
mapping and risk management. J Flood Risk Manage 2009, 2,
(3), 198–208.
Feyen L., Dankers R., Bódis K., Salamon P. & Barredo J. Fluvial
flood risk in Europe in present and future climates. Clim
Change 2012, 112, (1), 47–62.
Gao H.K., He X.B., Ye B.S. & Pu J.C. Modeling the runoff and
glacier mass balance in a small watershed on the Central
Tibetan Plateau, China, from 1955 to 2008. Hydrol Process
2012, 26, (11), 1593–1603.
Gerstengarbe F.W. & Werner P.C. Katalog der Großwetterlagen
Europas (1881–2009) nach Paul Hess und Helmut Brezowsky,
7th revised edition. Potsdam: Potsdam institute climate
impact research (PIK), 2010. PIK-Report No. 100.
Grünthal G., Thieken A., Schwarz J., Radtke K., Smolka A. &
Merz B. Comparative risk assessments for the city of cologne
– storms, floods, earthquakes. Nat Hazards 2006, 38, (1),
21–44.
Hagg W., Braun L.N., Weber M. & Becht M. Runoff modelling
in glacierized Central Asian catchments for present-day and
future climate. Nord Hydrol 2006, 37, (2), 93–105.
Hall J.W., Sayers P.B. & Dawson R.J. National-scale assessment
of current and future flood risk in England and Wales. Nat
Hazards 2005, 36, (1), 147–164.
Houston D., Werritty A., Bassett D., Geddes A., Hoolachan A. &
McMillan M. Pluvial (rain-related) flooding in urban areas:
the invisible hazard. York: Joseph Rowntree Foundation,
2011.
Hunter N.M., Bates P.D., Horritt M.S. & Wilson M.D.
Simple spatially-distributed models for predicting flood
inundation: a review. Geomorphology 2007, 90, (3–4), 208–
225.
Hurford A.P., Priest S.J., Parker D.J. & Lumbroso D.M. The
effectiveness of extreme rainfall alerts in predicting surface
water flooding in England and Wales. Int J Climatol 2012, 32,
(11), 1768–1774.
IPCC. Managing the risks of extreme events and disasters to
advance climate change adaptation. A special report of working
groups I and II of the Intergovernmental Panel on Climate
Change. Cambridge, UK and New York, NY: Cambridge Uni-
versity Press, 2012.
Kazmierczak A. & Cavan G. Surface water flooding risk to
urban communities: analysis of vulnerability, hazard and
exposure. Landsc Urban Plan 2011, 103, (2), 185–197.
Keef C., Svensson C. & Tawn J.A. Spatial dependence in extreme
river flows and precipitation for Great Britain. JHydrol2009,
378, (3–4), 240–252.
King L.M., McLeod A.I. & Simonovic S.P. Simulation of histori-
cal temperatures using a multi-site, multivariate block
resampling algorithm with perturbation. Hydrol Process 2012,
28, (3), 905–912.
Klasinc R., Heigerth G., Breitenstein S., Vosnjak S. & Steinman
F. Die neue Eisenbahnbrücke über die Salzach – Umbau in
einem hochwassergefährdeten Fluss. Modellversuche und
vergleichende Numerik, translated by Wallgau, Germany:
Lehrstuhl für Wasserbau und Wasserwirtschaft. Technische
Universität München, 2004, 82–91.
de Kok J.-L. & Grossmann M. Large-scale assessment of flood
risk and the effects of mitigation measures along the Elbe
River. Nat Hazards 2010, 52, (1), 143–166.
Krause P., Boyle D.P. & Bäse F. Comparison of different effi-
ciency criteria for hydrological model assessment. Adv Geosci
2005, 5, 89–97.
Kundzewicz Z.W., Ulbrich U., Brucher T., Graczyk D., Kruger
A., Leckebusch G.C., Menzel L., Pinskwar I., Radziejewski M.
& Szwed M. Summer floods in central Europe – Climate
change track? Nat Hazards 2005, 36, (1–2), 165–189.
Lall U. & Sharma A. A nearest neighbor bootstrap for
resampling hydrologic time series. Water Resour Res 1996, 32,
(3), 679–693.
Lamb R., Keef C., Tawn J., Laeger S., Meadowcroft I., Surendran
S., Dunning P. & Batstone C. A new method to assess the risk
of local and widespread flooding on rivers and coasts. J Flood
Risk Manage 2010, 3, (4), 323–336.
Larsen A.N., Gregersen I.B., Christensen O.B., Linde J.J. &
Mikkelsen P.S. Potential future increase in extreme one-hour
precipitation events over Europe due to climate change.
Water Sci Technol 2009, 60, (9), 2205–2216.
16 Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
112 Breinl et al.

Leander R. & Buishand T.A. A daily weather generator based on
a two-stage resampling algorithm. JHydrol2009, 374, (3–4),
185–195.
Lebensministerium. ehyd, [online], 2013. Available: http://
ehyd.gv.at/ [accessed 10 March 2014].
Lian J.J., Xu K. & Ma C. Joint impact of rainfall and tidal level
on flood risk in a coastal city with a complex river network: a
case study of Fuzhou City, China. HydrolEarthSystSci2013,
17, (2), 679–689.
Lindström G., Johansson B., Persson M., Gardelin M. &
Bergström S. Development and test of the distributed
HBV-96 hydrological model. JHydrol1997, 201, (1–4), 272–
288.
Loizl R. Hochwasserschutz Stadt Salzburg. Salzburg: Land Salz-
burg, Fachabteilung Wasserwirtschaft, 2012a.
Loizl R. Naturgefahrenmanagement: Prävention aus Sicht der
Schutzwasserwirtschaft. In: A. Kanonier, ed. FORUM
Raumplanung: Raumplanung und Naturgefahrenmanagement.
Vienna: Österreichische Gesellschaft für Raumplanung,
2012b, pp. 9–28.
Madsen H., Arnbjerg-Nielsen K. & Mikkelsen P.S. Update of
regional intensity-duration-frequency curves in Denmark:
tendency towards increased storm intensities. Atmos Res
2009, 92, (3), 343–349.
Maheepala S. & Perera B.J.C. Monthly hydrologic data genera-
tion by disaggregation. JHydrol1996, 178, (1–4), 277–291.
Mailhot A. & Duchesne S. Design criteria of urban drainage
infrastructures under climate change. J Water Resour Plann
Manage 2010, 136, (2), 201–208.
Merz B., Thieken A. & Apel H. What can go wrong? Flood risk
assessment and scenario analysis. In: J. Van Alphen, E.
Van Beek & M. Taal, eds. Floods, from defence to management.
Nijmegen, The Netherlands: Taylor & Francis Group, 2005,
pp. 673–679.
Mezghani A. & Hingray B. A combined downscaling-
disaggregation weather generator for stochastic generation of
multisite hourly weather variables over complex terrain:
development and multi-scale validation for the Upper Rhone
River basin. JHydrol2009, 377, (3–4), 245–260.
Morris M., Bryant R., Waller S., Hunter N., Lamb R., Crossley
A. & Balmbra V. An innovative approach to pluvial flood risk
assessment, Irish National Hydrology Conference 2009, 2009,
68–78.
Mudelsee M., Borngen M., Tetzlaff G. & Grunewald U. Extreme
floods in central Europe over the past 500 years: role of
cyclone pathway ‘Zugstrasse Vb’. J Geophys Res 2004, 109,
(D23), 1–21.
Munich Re. Topics geo – natural catastrophes 2012. Analyses,
assessments, positions. Munich: Munich Re, 2013.
Nash J.E. & Sutcliffe J.V. River flow forecasting through concep-
tualmodelspartI–adiscussion of principles. JHydrol1970,
10, (3), 282–290.
Neal J.C., Bates P.D., Fewtrell T.J., Hunter N.M., Wilson M.D. &
Horritt M.S. Distributed whole city water level measure-
ments from the Carlisle 2005 urban flood event and com-
parison with hydraulic model simulations. JHydrol2009,
368, (1–4), 42–55.
Parker D.J., Priest S.J. & McCarthy S.S. Surface water flood
warnings requirements and potential in England and Wales.
Appl Geogr 2011, 31, (3), 891–900.
Press W.H., Teukolsky S.A., Vetterling W.T. & Flannery B.P.
Numerical recipes in C++: the art of scientific computing, 2nd
edition. Cambridge, UK; New York: Cambridge University
Press, 2002.
Priest S.J., Parker D.J., Hurford A.P., Walker J. & Evans K.
Assessing options for the development of surface water flood
warning in England and Wales. J Environ Manage 2011, 92,
(12), 3038–3048.
Pui A., Sharma A. & Mehrotra R. A comparison of alternatives
for daily to sub-daily rainfall disaggregation. In: R.S.
Anderssen, R.D. Braddock & L.T.H. Newham, eds. 18th
World IMACS/MODSIM Congress. Cairns: The Australian
National University Canberra, 2009, pp. 3535–3541. 13–17
July 2009.
Rajagopalan B. & Lall U. A k-nearest-neighhor simulator for
daily precipitation and other weather variables. Water Resour
Res 1999, 35, (10), 3089–3101.
Romanowicz A.A., Vanclooster M., Rounsevell M. & La Junesse
I. Sensitivity of the SWAT model to the soil and land use
data parametrisation: a case study in the Thyle catchment,
Belgium. Ecol Modell 2005, 187, (1), 27–39.
Sampson C.C., Bates P.D., Neal J.C. & Horritt M.S. An auto-
mated routing methodology to enable direct rainfall in high
resolution shallow water models. Hydrol Process 2013, 27, (3),
467–476.
Segond M.L., Onof C. & Wheater H.S. Spatiat-temporal disag-
gregation of daily rainfall from a generalized linear model.
JHydrol2006, 331, (3–4), 674–689.
Seibert J. Estimation of parameter uncertainty in the HBV
model. Nord Hydrol 1997, 28, (4–5), 247–262.
Seibert J. Regionalisation of parameters for a conceptual rainfall-
runoff model. Agric Forest Meteorol 1999, 98–9, 279293.
Seibert J. Multi-criteria calibration of a conceptual runoff
model using a genetic algorithm. HydrolEarthSystSci2000,
4, (2), 215–224.
Seibert J. & Vis M.J.P. Teaching hydrological modeling with a
user-friendly catchment-runoff-model software package.
HydrolEarthSystSci2012, 16, (9), 3315–3325.
Sene K. Urban flooding. In: K. Sene, ed. Flash floods. United
Kingdom: Springer, 2013, pp. 293–311.
Shamseldin A.Y. & OConnor K.M. A nearest neighbour linear
perturbation model for river flow forecasting. JHydrol1996,
179, (1–4), 353–375.
Sharif M. & Burn D.H. Improved K-nearest neighbor weather
generating model. JHydrolEng2007, 12, (1), 42–51.
Sharma A. & Lall U. A nonparametric approach for daily rain-
fall simulation. Math Comput Simul 1999, 48, (4–6), 361–
371.
Modelling framework for urban river discharge and precipitation 17
J Flood Risk Management •• (2015) ••–•• © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
J Flood Risk Management 10 (2017) 97–114 © 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd
Modelling framework for urban river discharge and precipitation 113

Shrestha R., Tachikawa Y. & Takara K. Input data resolution
analysis for distributed hydrological modeling. JHydrol2006,
319, (1–4), 36–50.
Spekkers M.H., Kok M., Clemens F.H.L.R. & ten Veldhuis J.A.E.
A statistical analysis of insurance damage claims related to
rainfall extremes. HydrolEarthSystSci2013, 17, (3), 913–
922.
Srikanthan R. & McMahon T.A. Stochastic generation of
annual, monthly and daily climate data: a review. Hydrol
EarthSystSci2001, 5, (4), 653–670.
Steele-Dunne S., Lynch P., McGrath R., Semmler T., Wang S.Y.,
Hanafin J. & Nolan P. The impacts of climate change on
hydrology in Ireland. JHydrol2008, 356, (1–2), 28–45.
Thornthwaite C.W. An approach toward a rational classification
of climate. Geogr Rev 1948, 38, 55–94.
Uhlenbrook S., Seibert J., Leibundgut C. & Rodhe A. Prediction
uncertainty of conceptual rainfall-runoff models caused by
problems in identifying model parameters and structure.
Hydrol Sci J 1999, 44, (5), 779–797.
Veijalainen N., Lotsari E., Alho P., Vehviläinen B. & Käyhkö J.
National scale assessment of climate change impacts on
flooding in Finland. JHydrol2010, 391, (3–4), 333–350.
Verworn A. & Haberlandt U. Spatial interpolation of hourly
rainfall – effect of additional information, variogram infer-
ence and storm properties. HydrolEarthSystSci2011, 15,
(2), 569–584.
Wilks D.S. Multisite generalization of a daily stochastic precipi-
tation generation model. JHydrol1998, 210, (1–4), 178–191.
WMO. Guidelines on cross-border exchange of warnings.Geneva,
Switzerland: World Meteorological Organization, 2003.
Public Weather Service, PWS-9: TD No. 1179.
Wojcik R. & Buishand T.A. Simulation of 6-hourly rainfall and
temperature by two resampling schemes. JHydrol2003, 273,
(1–4), 69–80.
Zhou Q., Mikkelsen P.S., Halsnaes K. & Arnbjerg-Nielsen K.
Framework for economic pluvial flood risk assessment con-
sidering climate change effects and adaptation benefits.
JHydrol2012, 414, 539–549.
18 Breinl et al.
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management •• (2015) ••–••
© 2015 The Chartered Institution of Water and Environmental Management (CIWEM) and John Wiley & Sons Ltd J Flood Risk Management 10 (2017) 97–114
114 Breinl et al.