Water Resources Research

RESEARCH ARTICLE Incorporating Antecedent Moisture Conditions and Intraevent

10.1029/2018WR023194
Variability of Rainfall on Flood Frequency Analysis in Poorly
Key Points:
• The methodology presented
Gauged Basins
incorporates the observed spatial L. Cea1 and I. Fraga2
and temporal variability of rainfall in
flood frequency analysis 1
Environmental and Water Engineering Group, Department of Civil Engineering, Universidade da Coruña, Coruña, Spain,
• The proposed method accounts for 2
the combined effect of antecedent CITIC, Universidade da Coruña, Coruña, Spain
soil moisture and rainfall intensity in
flood frequency estimates
• With the proposed approach, there is Abstract Rainfall-runoff models are frequently used for the prediction of flood discharges from extreme
no need to define synthetic design rainfall data in poorly gauged basins. However, the conventional methods used to estimate the frequency
hyetographs for a given return
period
of flood discharges from precipitation data do not incorporate a realistic representation of the spatial
and temporal structure of rainfall at the study site nor of the antecedent soil moisture content. Temporally
and spatially idealized rainfall patterns that do not represent properly the intraevent variability of
rainfall are in general assumed. In this paper, we propose a methodology for flood frequency analysis
Correspondence to:
L. Cea, based on the hydrological simulation of potentially hazardous real storm events. The aim of the proposed
luis.cea@udc.es method is to incorporate the impact of the spatial and temporal intraevent variability of rainfall on flood
discharge estimates without making strong assumptions about the spatial and temporal patterns of
Citation: storms. It also includes a calibrated relation between the antecedent rainfall depth and the soil infiltration
Cea, L., & Fraga, I. (2018). Incorporating capacity, in order to consider the combined effect of antecedent soil moisture content and rainfall
antecedent moisture conditions and intensity. The flood frequency estimates obtained with the proposed methodology are compared to those
intraevent variability of rainfall on flood
frequency analysis in poorly gauged estimated with a conventional method based on the definition of synthetic design hyetographs from local
basins. Water Resources Research, 54, intensity-duration-frequency curves. The analysis is presented in two watersheds of 84 and 353 km2
8774–8791. https://doi.org/10.1029/ located in the NW of Spain. Results highlight the role of the antecedent soil moisture in flood frequency
2018WR023194
analysis and show that the conventional methodology tends to overestimate the flood discharges when
Received 24 APR 2018 compared to the proposed approach.
Accepted 17 OCT 2018
Accepted article online 29 OCT 2018
Published online 8 NOV 2018 1. Introduction
Flood frequency analysis in poorly gauged basins (those that lack comprehensive streamflow data) relies
either on empirical regional regression equations or on rainfall-runoff transformation modeling. Regional
flood frequency analysis (RFFA) is commonly used when comprehensive streamflow data are available in
catchments with similar hydrological and climatological characteristics as the study basin, even though some
attempts have been done to develop these methods at the global scale (Smith et al., 2015). Although RFFA
methods have proven to give reliable flood frequency estimates and are generally accepted in engineering
practice, they remain a black box as they do not provide any knowledge about the physical processes that
generate the floods at a specific basin (Dawdy et al., 2012). Conversely, methods based on physically based
rainfall-runoff modeling can provide some understanding of the processes involved in the generation of
floods, which might be relevant for engineering applications and for climate change studies (Dawdy et al.,
2012). This paper deals with the application of rainfall-runoff transformation methods to flood frequency
analysis, focusing on the incorporation of rainfall variability and antecedent soil moisture content at the
event scale.
Conventional methods used in flood frequency estimation from precipitation data do not account properly
for the spatial and temporal variability of rainfall during storm events. However, it has been shown in several
studies that storm motion, which is one of the main causes of rainfall variability, has a relevant impact on
flood peaks (De Lima & Singh, 2002; Liang & Melching, 2015; Mizumura & Ito, 2011; Sigaroodi & Chen,
2016). Standard procedures assume synthetic patterns to distribute the rainfall depth in time and space.
The most common approach in engineering practice is to use the intensity-duration-frequency (IDF) curves
evaluated at the site of interest from point rainfall historical data. A family of IDF curves contains some infor-
©2018. American Geophysical Union. mation about the interevent temporal structure of rainfall, but this information is incomplete and needs to be
All Rights Reserved. complemented with a synthetic hyetograph in order to define the temporal pattern of the design

CEA AND FRAGA 8774

 Water Resources Research 10.1029/2018WR023194

hyetographs. One of the most common methods to do this is the Alternating Block Method (ABM), although
there are others such as a triangular hyetograph, the Soil Conservation Service (SCS) storms (NRCS, 2007), or
the HMR 52 storm (Hansen et al., 1982), which defines the spatial and temporal distribution of the probable
maximum precipitation for a watershed. None of these hyetographs represents properly the temporal varia-
bility of real events, especially when several rainfall intensity peaks occur during a storm. However, even if
temporal rainfall variability of real events shows substantial departure from idealized design storm hyeto-
graphs, the impacts of design storm assumptions, such as spatially uniform or idealized rainfall patterns,
on flood risk estimates are still poorly understood (Wright et al., 2013, 2014a, 2014b). This omission contrasts
with the well-known sensitivity of catchment response to rainfall distribution (Lobligeois et al., 2014; Niemi
et al., 2016; Shen et al., 2012).
Regarding the spatial distribution of rainfall, the most common approach is to use either a single basin-
averaged design hyetograph corresponding to a given return period in the whole catchment (assuming
therefore a spatially homogeneous rainfall) or to use different design hyetographs (all of them corresponding
to the same return period) at each subbasin of the hydrological model. In both cases, the peak rainfall inten-
sity associated to a given return period occurs simultaneously over the whole catchment. This is an unrealistic
assumption in medium and large catchments. In addition, it is unlikely that the rainfall depths registered at
different locations during the same event have the same return period.

Areal reduction factors (ARFs), defined as the ratio between extreme area-averaged rainfall and extreme
point rainfall, are commonly used to cope with this inconvenience (Asquith & Famiglietti, 2000; Bacchi &
Ranzi, 1996; NERC, 1975; NSSP, 1961; Rodriguez-Iturbe & Mejía, 1974; Sivapalan & Blöschl, 1998; Vaes et al.,
2005). If ARFs are known for the study catchment, effective rainfall depths can be computed by multiplying
point depths by the ARF. However, the estimation of ARFs is difficult to generalize and extrapolate to other
basins and storm conditions because they depend on the size, shape, and topography of the basin, on its
geographic location and on the rainfall duration. In addition, several works have suggested that ARFs depend
on return period (Sivapalan & Bloeschl 1998; Asquith & Famiglietti, 2000; Allen & DeGaetano 2005; Veneziano
& Langousis 2005), which constrains even more their applicability. For these reasons, commonly used ARFs
do not represent properly the real properties of extreme rainfall, and this might result in large errors in the
estimation of flood discharges (Wright et al., 2014a).

In addition to the previous issues, when using a single design hyetograph for a given return period, it is diffi-
cult to establish the antecedent soil moisture content (AMC) that should be considered in the hydrological
model to compute rainfall-runoff. This becomes especially important since the impact of antecedent rainfall
on the soil infiltration capacity and thus on the generation of surface runoff is well known. The same flood
discharge can originate from a high rainfall intensity falling over a moderately wet terrain or from a medium
rainfall intensity falling over a saturated terrain. In most locations, soil moisture and extreme rainfall are
cyclostationary and partially correlated random processes. Therefore, the statistical distribution of flooding
depends on their joint distribution, which is in general unknown. However, conventional methods for the
evaluation of design hyetographs do not account properly for the dependence between extreme rainfall
and initial soil moisture conditions. Some authors have proposed methods to account for the joint variability
between these two variables using long-term continuous hydrological simulations, including in some cases
the generation of long-term synthetic series of weather variables (Blazkova & Beven, 2004; Brocca et al.,
2011; Camici et al., 2011; Falter et al., 2015; Moretti & Montanari, 2008). All these works highlight the impor-
tance of considering the joint probability of occurrence of AMC and extreme rainfall in flood
frequency analysis.

For the previous reasons, Wright et al. (2014a) suggest that, rather than updating methods that assume
strong simplifications on the structure of rainfall, as ARF-based methods, research and practice should focus
on flood estimation techniques that incorporate a realistic site-specific representation of the spatial and
temporal structure of storm rainfall. Following this line, in this paper we analyze the impact of the spatial
and temporal intra-event variability of rainfall on flood frequency analysis in poorly gauged basins using
rainfall-runoff models. We propose a methodology based on the hydrological simulation of potentially
hazardous real storm events. By poorly gauged basin, we refer to catchments with just a few measured
discharge data over a limited period of time, sufficient to calibrate a hydrological model but insufficient to
perform a statistical flood frequency analysis on streamflow data. The proposed method includes a

CEA AND FRAGA 8775

 Water Resources Research 10.1029/2018WR023194

Figure 1. Study region and location of the 11 rain gauges. The Mandeo and Mendo catchments, as well as the subbasins included in the hydrological model, are
shown in the left figure.

calibrated relation between the antecedent rainfall depth and the soil infiltration capacity during a storm
event. The advantage of this technique over standard procedures is that it does not make strong
assumptions about the spatial and temporal structure of rainfall and, at the same time, it naturally
considers the combined effect of AMC and rainfall intensity.
The flood frequency estimates obtained with the proposed methodology are compared to those computed
with a conventional method based on the evaluation of basin-averaged IDF curves and on the definition of
synthetic design hyetographs. The analysis is presented in two watersheds located in northwestern Spain
with drainage areas of 84 and 353 km2, respectively. The results show that the conventional methodology
is very sensitive to the AMC conditions assumed for a given return period, and it tends to overestimate the
flood discharges when compared to the proposed approach. Since it is not straightforward to establish which
AMC should be used when applying a conventional method, this results in a great source of uncertainty in
the results.
The following of this paper is organized as follows. Section 2 describes the physical properties of the study
catchments and the methodology proposed here to incorporate the combined effect of AMC and rainfall
variability in flood frequency analysis. The results obtained in the two study catchments with the proposed
and conventional methods are compared and discussed in section 3. Finally, the main conclusions and con-
tributions of the paper are summarized in section 4.

2. Study Area and Methodology

2.1. Study Catchments and Available Data
The two catchments used as study cases are those of the rivers Mandeo and Mendo (Figure 1). Both of them
are located in northwestern Spain and are affected by the energetic atmospheric interchanges between the
tropical and the polar regions. This produces frequent adverse storm events characterized by a large spatial
and temporal variability of rainfall (Cabalar-Fuentes, 2005). The areas of each catchment are 353 (Mandeo)
and 84 km2 (Mendo), while the lengths of the main rivers are roughly 50 (Mandeo) and 30 km (Mendo).
Average annual rainfall is 1,260 (Mandeo) and 1,240 mm (Mendo).
Both are predominantly rural catchments in which forest and brushland are the most extended land uses. In
the Mandeo catchment, the elevation varies from zero to 810 m above sea level, with an average elevation of
480 m and an average slope of 0.056. The elevation of the Mendo basin varies within 25 and 564 m, with an
average elevation of 344 m and an average slope of 0.052.

CEA AND FRAGA 8776

 Water Resources Research 10.1029/2018WR023194

Table 1 The orography and the orientation of the hillslopes have relevant effects
Pluviometers Used in the Present Study on the characteristics of the storm events in the study catchments. On
Pluviometer one hand, both catchments have a NNW orientation, which is the most fre-
Measuring Z Pannual P24
Id Name since (m) (mm) (mm) quent direction of arrival of the low pressure fronts in this region of Spain.
Therefore, the path of most of the arriving storms follows the upstream
R1 Arzúa 2009 362 1,484 84
direction of the basins, which has an attenuation effect on the peak dis-
R2 CIS Ferrol 2000 37 1,243 52
R3 Corno do Boi 2005 731 1,326 66 charges (Liang & Melching, 2015). On the other hand, due to the steep oro-
R4 Coruña Dique 2008 5 649 37 graphy of the hillslopes, the arriving low pressure fronts are quickly
R5 Guísamo 2012 176 1,145 56 uplifted on their path through the catchment. This contributes to increase
R6 Sambreixo 2000 496 1,040 55 the spatial and temporal variability of rainfall during storm events
R7 Vilalba 2001 684 1,186 56 (Cabalar-Fuentes, 2005). These effects have also been observed in other
R8 Mabegondo 2000 94 1,105 58
R9 Marco da Curra 2000 651 1,476 70
regions with similar climate and orography, such as the U.S. Pacific Coast
R10 Melide 2003 477 1,154 59 (Eiras-Barca et al., 2016).
R11 Olas 2005 401 1,492 70
Historical rainfall data with a time resolution of 10 min is available from the
Note. Z = elevation, Pannual = average annual precipitation, P24 = average rain gauge network of the regional meteorological agency MeteoGalicia.
annual maximum rainfall depth in 24 hr.
From this network, 11 tipping bucket pluviometers located around the
catchments were selected in order to interpolate the rainfall fields
(Figure 1 and Table 1). The average annual maximum rainfall depth in 24 hr (P24) registered at each station
varies between 37 and 84 mm. The average annual precipitation (Pannual) is strongly correlated with P24
(r = 0.90) and varies, with the exception of rain gauge R4, within 1,000 and 1,500 mm. Notice in Table 1 that
the rain gauge density varies over the time period with available data. This implies that the most recent
storms will have more realistic rainfall estimates than the earlier ones. Nonetheless, this does not pose a pro-
blem for the proposed methodology.

Radar data with spatial and temporal resolutions of 1 km and 5 min, respectively, are also available but only
for a limited number of rainfall events. Due to its limited length, the radar data could not be used in the sta-
tistical analysis of rainfall. However, it was used to verify that the 11 available pluviometers are sufficient to
represent the main features of the spatial structure of storm rainfall for aggregation levels larger or equal
to 1 hr. This will be shown later on in section 2.2.2.
Daily peak discharges measured since 2010 at two gauge stations (one on each catchment) managed by the
Regional Water Administration (Augas de Galicia) are also available. There is no detailed information about
the accuracy and reliability of these data. A preliminary analysis revealed several incoherent records,
characterized by very high peak discharges during low rainfall periods. Even if the most obvious erroneous
data were filtered, the uncertainty on the remaining discharge data is probably large. Despite the limited
accuracy ascribable to discharge measurements, the available data were used to do a rough calibration of
the hydrological model, as it will be detailed in the following sections.

2.2. Proposed Methodology

The proposed methodology is based on the hydrological simulation of real storm events in the study catch-
ments in order to obtain a series of flood peak discharges, which are then used to fit an extreme value dis-
tribution. The main steps of the methodology are listed below and described in the following of this section:

1. Select historical time series of rainfall data at an adequate time resolution for the study catchment.
2. Compute historical time series of precipitation averaged over the whole basin and over each of its main
subbasins (those that will be included in the hydrological model).
3. Find and select potentially hazardous storm events from the basin-averaged time series of precipitation
computed in step 2. This will typically be done using a peak-over-threshold criterion based on the rainfall
depth. Since the river discharge depends not only on the total rainfall depth but also on the spatial and
temporal variability of rainfall and on the AMC, it is convenient to define a relatively low peak-over-
threshold (POT) and let the hydrological model determine the storm events that generate highest river
discharges.
4. Model the selected storm events with a semidistributed or distributed hydrological model. The model
must be able to account for the spatial variability of rainfall in the simulations (i.e., do not use the
basin-averaged rainfall time series in the hydrological model but instead define spatially varying rainfall

CEA AND FRAGA 8777

 Water Resources Research 10.1029/2018WR023194

fields from the available data). At this step, if some streamflow data are available, a relation between the
antecedent rainfall depth and the infiltration parameters should be calibrated.
5. Extract the peak discharge for each rainfall event modeled in step 4, and fit an extreme value distribution
to the discharge data. The flood discharge for different return periods can be evaluated from the fitted
extreme value distribution.
2.2.1. Selection of Historical Time Series of Rainfall
The first step of the methodology is to obtain time series of rainfall that can be used to define its spatial and
temporal variability at the event scale during a historical time period. Most commonly, these will be point
rainfall data at several pluviometers located around and within the study catchment. In this case we have
used the 11 pluviometers shown in Figure 1 and in Table 1. The historical data obtained in this features spatial
and temporal variability in basin-scale rainfall, as represented by the observations from the rain gauges. In
some cases there might be data from meteorological radars that can be combined with rain gauge data to
obtain a more detailed characterization of the rainfall fields.
2.2.2. Basin-Averaged Time Series of Precipitation
The second step is to obtain basin-averaged and subbasin-averaged time series of precipitation from the
available rainfall data. The data registered by the 11 pluviometers of MeteoGalicia were spatially interpolated
using a natural neighbor interpolation and then averaged over the catchment surface. Other interpolation
techniques, as nearest neighbor interpolation and ordinary kriging, were tested, with similar results. More
accurate techniques such as kriging with external drift (KED) could not be used because the availability of
radar data in the study region is limited and could not be applied systematically to interpolate the rainfall
fields in the catchments. The spatial interpolation was done with a time resolution of 10 min, in order to
obtain historical time series of basin-averaged precipitation at 10-min resolution. The basin-averaged time
series of rainfall used in this work extended from 2001 to 2017 (17 years).

To verify the accuracy of the basin-averaged rainfall computed with natural neighbor interpolation, a number
of storm events in which radar data were available were selected, and the KED interpolation was used to esti-
mate the rainfall fields for those events. Time series of basin-averaged rainfall for the selected events com-
puted with KED and natural neighbor interpolations are compared in Figure 2. There is a strong similarity
between time series, especially for aggregation levels larger than 1 hr. For the 3-hr rainfall aggregation level,
the correlation coefficient is equal to 0.99 and the mean absolute error of the order of 0.1 mm/hr. For a 0.5-hr
rainfall duration, the discrepancies in the rainfall intensity is somewhat larger, but the dispersion around the
1:1 line is low (with correlation coefficients larger than 0.90 in all cases), and there is no significant bias (per-
cent bias lower than 8%). Nonetheless, this timescale is not relevant given the size of the study catchments. In
any case, as mentioned before, the radar data could not be included in our analysis since it is just available for
a limited number of events. If available in other case studies, it can be easily included in the methodology in
order to compute the basin-averaged rainfall time series.
2.2.3. Potentially Hazardous Rainfall Events
The historical basin-averaged 10-min series of rainfall were aggregated at different durations ranging from 1
to 24 hr (1, 3, 6, 12, 18, and 24 hr). From each of these, a number of potentially hazardous rainfall events were
selected using the POT method. The threshold for each aggregation level was defined as 70% of the average
annual maximum rainfall for that specific duration. A minimum 1-week lag was also imposed when applying
the POT method. One week is not enough to assure statistical independence of two consecutive events.
However, the storm events identified at this step will not be directly used to fit an extreme rainfall distribution
but just as input data in a hydrological model. Therefore, the statistical independence of events is not rele-
vant at this step. Conversely, it is difficult to assess a priori which rainfall events are more likely to generate
high discharges, due to the intraevent variability of rainfall and to the role of the AMC. Thus, it is more con-
venient to remain nonrestrictive at this stage and to select all the events that can be considered potentially
hazardous. Statistical independence of events will be imposed in later steps.

This criterion produced around 100 events in 17 years of available data (i.e., around six events per year). Other
slightly different thresholds (between 50% and 90% of the average annual maximum rainfall) and minimum
lags between events (from 5 to 15 days) were tested, with similar results at the end of the methodology.
2.2.4. Hydrological Modeling of the Potentially Hazardous Rainfall Events
From the previous step, a number of potentially hazardous storm events were obtained. These events were
introduced in a hydrological model to obtain the corresponding flood hydrographs and peak discharges. We

CEA AND FRAGA 8778

 Water Resources Research 10.1029/2018WR023194

Figure 2. Comparison between basin-averaged rainfall obtained with natural neighbor interpolation from the pluviometers data and from kriging with external drift
(KED) interpolation combining the pluviometers and radar data. r = Pearson correlation coefficient, MAE = mean absolute error, PBIAS = percent bias (Gupta et al.,
1999).

have used the semidistributed model HEC-HMS v4.2.1 (Scharffenberg & Fleming, 2006), although any other
semidistributed or distributed model could be used for the same purpose as long as the spatial variability
of rainfall is considered in the model.
The geospatial hydrology toolkit HEC-GeoHMS (Fleming & Doan, 2009) was used to divide the study catch-
ments in several subbasins, to generate the geometrical representation of the catchments in HEC-HMS, and
to obtain the main features of each hydrological element (subbasin areas, reach lengths, and average slopes).
The Mandeo catchment (353 km2) was subdivided in 34 subbasins with areas ranging from 0.2 to 30 km2,
while the Mendo catchment (84 km2) was discretized in 24 subbasins with areas ranging from 0.5 to 8 km2.
The rainfall time series were defined individually for each subbasin, from the subbasin-averaged values
obtained in step 2. In such a way, a realistic spatial and temporal structure of rainfall is included in
the simulations.

CEA AND FRAGA 8779

 Water Resources Research 10.1029/2018WR023194

Table 2 Rainfall-runoff transformation was computed with the SCS unit hydro-
Catchment-Averaged Values of the Nine Curve Number (CN) Classes
graph. The lag time (Tlag) used in the SCS unit hydrograph was estimated
Considered in the Hydrological Model
as Tlag = 0.6Tc. The concentration time (Tc) of each subbasin was deter-
CN class CN1 CN2 CN3 CN4 CN5 CN6 CN7 CN8 CN9
mined using the TR55 method (USDA, 1986). TR55 computes Tc as the
Mandeo 51 56 61 66 71 76 80 85 90 sum of three travel times that correspond to sheet flow, shallow concen-
Mendo 53 56 60 63 66 69 73 76 79 trated flow, and channel flow. Each of them is computed from the topogra-
phical features (slope, length, and area) and surface roughness (Manning
coefficient) of the corresponding hydrological elements (hillslopes, shal-
low watercourses, and main channels). Using these estimations, the lag times computed for the different sub-
basins vary from 0.25 to 2.2 hr in the Mandeo catchment and from 0.3 to 1.5 hr in the Mendo catchment.
Flow routing along the main river network was computed using the kinematic wave equation. The average
slope of the river reaches was obtained from the Digital Terrain Model (DTM). The average values of the
cross-section width and Manning coefficient were determined from field inspection. Cross-section width var-
ies from 5 to 10 m, while the Manning coefficient varies from 0.03 to 0.06.
Infiltration losses were evaluated using the curve number (CN) method. A reference CN was assigned to each
subbasin according to its Hydrological Soil-Cover Complex (soil cover, practice, hydrological condition, and
hydrological soil group) for average soil moisture conditions. Since the infiltration capacity of a soil is largely
dependent on its AMC, the SCS defines three possible CN classes (dry, average, and moist) based on the
cumulative 5-day antecedent rainfall depth before the storm event (R5). Each CN class is defined by a factor
that multiplies the reference CN corresponding to average AMC conditions. These factors depend on the CN
itself and are lower than 1 (increase infiltration) for dry antecedent conditions and greater than 1 (reduce infil-
tration) for moist antecedent conditions.
The CN classes proposed by the SCS can be used if no calibration data at all is available at the study site. In our
case, there are two gauging stations, one on each catchment, with daily peak discharge data since 2010.
These data were used to calibrate a relation between different CN classes and the antecedent rainfall depth.
This was the only calibration performed on the hydrological model. The procedure followed is slightly differ-
ent to the standard SCS procedure and is therefore detailed in the following.
The reference CN, as well as the moist and dry AMC factors recommended by the SCS, was estimated indivi-
dually for each subbasin defined in the hydrological model. This gives three CN classes corresponding to dry,
average, and moist conditions, each one representing a spatial distribution of CN in the catchment. For each
subbasin, six additional CN classes were defined (three CN classes between the dry correction and the refer-
ence value, and three values between the reference value and the moist correction). The CN values for each
of these classes were linearly interpolated from the SCS recommended values. This gives nine possible CN
values for each subbasin. The catchment-averaged values of CN corresponding to the nine classes are shown
in Table 2. Notice that only the catchment-averaged values are shown in the table for the sake of conciseness,
but each subbasin has its own value of CN according to the Hydrological Soil-Cover distribution. In Table 2,
the classes CN1, CN5, and CN9 correspond respectively to the SCS values for dry, normal, and moist AMC.
Once the nine CN classes are defined, they must be related to the AMC of the soil. In order to do so, a number
of calibration events were selected from the available observed data. The calibration events were chosen as
those with available measured discharge data and at the same time classified as potentially hazardous
according to the criterion defined in step 3. This produced around 30 calibration events for each catchment.
The HMS model was run nine times for each calibration event (one run for each possible CN class). The peak
discharges obtained with the model were compared to the measured ones in order to establish, for each rain-
fall event, the CN class that gives the best agreement with the observed data. The optimum CN classes
obtained in this way were plotted against the antecedent rainfall depth in the n previous days (with varying
values of n between 1 and 120 days), and the following power regression was fitted to the data:


CNclass ¼ round a·Rbn (1)

where Rn is the rainfall depth on the n previous days, a and b are the constant parameters of the power
regression, and CNclass is an integer value between 1 and 9. From the n possible power regressions, the
one with the best agreement between numerical and observed peak discharges in terms of the

CEA AND FRAGA 8780

 Water Resources Research 10.1029/2018WR023194

Figure 3. Agreement between numerical and observed peak discharges in terms of Nash-Sutcliffe efficiency (NSE) as a
function of the number of antecedent days considered in equation (1) (left), and relation between CNclass and n day
antecedent rainfall depth that gives the best agreement (right).

Nash-Sutcliffe Efficiency (Nash & Sutcliffe, 1970) was chosen in order to relate the CN classes with its n-day
antecedent rainfall. This procedure was done independently for the Mandeo and Mendo catchments
(Figure 3).
Following this methodology, in the Mandeo catchment, the best predictor of CNclass was found to be R34 (i.e.,
the 34-day antecedent rainfall depth). In the Mendo catchment, the best calibration was obtained with R31. In
both cases, using R5 produced a significantly worse calibration. The fitted regressions for each catchment are
shown in Figure 3 and are given by


CNclass;Mandeo ¼ round 0:078·R0:868
34 (2)


CNclass;Mendo ¼ round 0:037·R0:928
31 (3)

The relations given by equations (2) and (3) are the only ones obtained from model calibration. Using R34 and
R34 as the CNclass predictors is coherent with the results obtained by Schulze (1982), who found that in humid
regions (as it is the study region), using the 30-day antecedent rainfall as a predictor of the AMC when apply-
ing the SCS method produced better results than using a 5-day period. Similar findings were reported by
Hope and Schulze (1982), who used a 15-day antecedent period when applying the SCS method to humid
regions of South Africa.
The CNs derived via calibration using the previous methodolgy differ substantially from the CNs that would
be assigned from standard guidance, as shown in Figure 4. The correlation between both values of CN is not
significant, and the differences are quite relevant. This means that the methodology proposed here needs
some streamflow data in order to calibrate a relation between CN and some predictor of the soil moisture

Figure 4. Scatterplot comparing the curve numbers (CNs) derived via calibration with the CNs that would be used from
standard guidance in the Mandeo (left) and Mendo (right) catchments. Each scatterpoint represents the CN assigned to
a specific subbasin of the hydrological model during a specific calibration event.

CEA AND FRAGA 8781

 Water Resources Research 10.1029/2018WR023194

Figure 5. Agreement between the observed and simulated peak discharges in the Mandeo (left) and Mendo (right) catch-
ments. MAE = mean absolute error, NSE = Nash-Sutcliffe efficiency, PBIAS = percent bias (Gupta et al., 1999).

conditions. Otherwise, rainfall-runoff transformation will not be properly represented by the

hydrological model.
The measured peak discharges versus the ones predicted with HEC-HMS for the calibration events using
equations (2) and (3) are shown in Figure 5, together with the following performance measures:
 2    
exp  exp  exp
∑ni¼1 QHMS
p;i  Qp;i ∑ni¼1 QHMS
p;i  Qp;i  ∑ni¼1 QHMS
p;i  Qp;i
NSE ¼ 1   2 MAE ¼ PBIAS ¼ 100 (4)
n ∑ni¼1 Qexp
∑ni¼1 Qexp
p;i  Qp
exp p;i

where n is the number of calibration events, QHMS

p;i is the peak discharge computed with HEC-HMS for each
calibration event, and Qexp
p;i is the measured peak discharge for each calibration event.

Once the relations between the antecedent rainfall and CN are established from the calibration events, equa-
tions (2) and (3) are used to assign a CN class to each one of the potentially hazardous rainfall events identi-
fied in step 3. These are modeled in HEC-HMS, using the measured temporal and spatial variability of rainfall
in the catchments, to obtain the estimated peak discharge for each potentially hazardous rainfall event.
It should be noticed that the antecedent rainfall is chosen as the predictor of the CN because the infiltration
capacity of the soil depends largely on the initial soil moisture conditions, which in general are strongly
related to the rainfall depth during the days previous to the flood event. However, the present availability
of satellite data opens up the possibility of using directly the soil moisture content, instead of the antece-
dent rainfall, as the predictor of the CN for each storm event. To explore this possibility, we have used
the satellite data from the Soil Moisture Active Passive (SMAP) mission, which gives estimations of the soil
moisture content at a 10-km resolution. Figure 6 shows the CN calibrated with the proposed methodology

Figure 6. Soil Moisture Active Passive (SMAP) root soil moisture versus calibrated curve number (CN) in the storm events
identified since 2015 (available SMAP data). Mandeo (left) and Mendo (right) catchments.

CEA AND FRAGA 8782

 Water Resources Research 10.1029/2018WR023194

in the storm events registered since 2015, against the root soil moisture obtained from SMAP at the
beginning of those events. The correlation coefficient between both variables is very high (0.92 in the
Mandeo and 0.85 in the Mendo), which suggests that soil moisture data obtained from SMAP has a great
potential to be used as an efficient predictor of the CN in future applications of the methodology.
However, SMAP data are only available since 2015, and therefore, it could not be fully incorporated in the
study cases presented in this paper. Therefore, in the present study, we have maintained the antecedent
rainfall to derive the CN.
2.2.5. Extreme Value Distribution of the Peak Discharge
The peak discharges obtained with HEC-HMS for each event identified in step 3 were used to estimate the
annual exceedance probability distribution of discharge at both catchments. To do so, the POT method
was applied to the modeled peak discharges in order to obtain a statistically independent sample of flood
discharges. This is necessary because (1) not all the potentially hazardous rainfall events produce high dis-
charges and (2) the lag period used for the selection of rainfall events in step 2 is too small to ensure statistical
independence. The POT method was applied with a threshold equal to 50% of the average annual maximum
discharge and a minimum lag period between flood events of 1 month. This criterion produced roughly one
flood event per year.
Once the peak discharges sample was obtained, the exceedance probability of each discharge was estimated
using the plotting position formula of Gringorten (1963), and an extreme value distribution was fitted to the
data. The formula of Gringorten is optimized to plot the largest observations of a Gumbel distribution.
However, it is common in practice to try several extreme value distributions and select the one that gives
the best fit to the data. In this case we have tested the two-parameter Gumbel distribution and the three-
parameter Generalized Extreme Value distribution. A three-parameter distribution is more flexible to fit the
data, but it is also more sensitive to outlying events, since it can easily follow singularities of the data set
(Cunnane, 1985). Some water administrations recommend using a three-parameter distribution only when
the length of the measured data exceeds a minimum number of 50 years (Midttømme et al., 2011).
Otherwise, a two-parameter distribution is preferred. Taking into account the previous considerations, the
Gumbel distribution was chosen to characterize the extreme flood discharge at both catchments.

2.3. Conventional Methodology

Conventional methods used in flood frequency analysis from rainfall data are based on the evaluation of IDF
curves and on the definition of synthetic hyetographs for the study region. The practical application of con-
ventional methods might differ in details such as the way in which the IDF curves are computed, the shape
and parameters used to define the synthetic hyetographs, the definition of ARF, or the way in which rainfall is
distributed over the study catchment. The main drawbacks of conventional methods have been discussed in
the section 1 and originate mainly from the fact that they do not account properly for the real temporal and
spatial structure of rainfall during storm events, nor for the AMC of the soil. These drawbacks have already
been pointed out by other researchers (Wright et al., 2014a).

In this study we have used a conventional method based on the estimation of basin-averaged IDF curves for
the study catchments and on the definition of synthetic hyetographs for a given return period using the ABM.
Those are quite standard procedures in flood frequency analysis. The specific implementation used in this
study is described in the following of this section. The same raw rainfall data were used in the proposed
and conventional methods.
2.3.1. Basin-Averaged IDF Curves
The IDF curves were computed from the historical time series of basin-averaged rainfall at 10-min resolution,
computed as detailed in section 2.2.2. This was done independently for the Mandeo and Mendo catchments.
In addition, and just for comparison purposes, we have computed the IDF curves at each of the 11 rain gauge
stations (i.e., using just the rainfall data registered by each station). Using basin-averaged and site-specific IDF
curves takes into account the spatial variability of rainfall in the study catchment and avoids the need of using
generic ARF derived for other catchments or regions. Notice that very often in practice, conventional meth-
ods are applied using generic ARF, at best derived for the climatic region where the study catchment is
located. Those kinds of implementations based on ARF account for point-to-area rainfall differences, but they
do not include the spatial variability of rainfall in the hydrological simulations.

CEA AND FRAGA 8783

 Water Resources Research 10.1029/2018WR023194

Figure 7. Estimated catchment-averaged intensity-duration-frequency curves corresponding to return periods of 2, 10, 25, and 50 years.

The POT method was applied to the basin-averaged time series considering different rainfall aggregation
levels (durations of 0.5, 1, 3, 6, 12, 18, and 24 hr). For each aggregation level, the rainfall threshold was defined
as 70% of the average annual maximum rainfall depth for that specific duration. The minimum lag period
between peaks was established as 30 days. This criterion produced around 70 rainfall events on each catch-
ment. For each duration, the exceedance probability of the sampled data was estimated using the formula of
Gringorten (1963). Following the considerations mentioned in section 2.2.5, three candidate extreme value
distributions were tested: Gumbel (two parameters), sqrtETmax (two parameters), and Generalized Extreme
Value (three parameters). The sqrtETmax (SQRT-Exponential Type Distribution of Maximum) was proposed in
Etoh et al. (1987) specifically for annual maximum series of total rainfall depth during a single rainstorm or
rainfalls of about 24 hr.
The best fit to the data was obtained with the Gumbel distribution, and it was therefore the one used to com-
pute the IDF curves shown in Figure 7. Due to the relatively short length of the available time series (17 years),
the IDF curves were only computed for return periods of up to 50 years. It is remarkable in Figure 7 the differ-
ences between the basin-averaged IDFs and those estimated from point rainfall measurements at each sta-
tion. This is in direct relation with the ARF. For short rainfall durations, the point rainfall intensities are in
general larger than the basin-averaged intensities because the maximum rainfall does not occur over the
whole catchment simultaneously. The differences in the point rainfall IDFs is due to the nonhomogeneity
of the statistical characteristics of rainfall over the study region (see Table 1). For those small aggregation
levels, the basin-averaged intensities in the Mendo catchment are larger than in the Mandeo catchment
because the later one has a larger surface, and thus, the ARF is smaller. For larger aggregation levels, the
basin-averaged IDF in the Mandeo and Mendo tend to the same value because the spatial correlation of rain-
fall increases with the rainfall duration. Those differences are reflected in the basin-averaged design hyeto-
graph computed with the ABM (Figure 8), which has a higher peak intensity in the smaller
catchment (Mendo).

CEA AND FRAGA 8784

 Water Resources Research 10.1029/2018WR023194

Figure 8. Catchment-averaged synthetic hyetographs for return periods of 2, 10, 25, and 50 years.

2.3.2. Synthetic Hyetographs

From the basin-averaged IDFs, the synthetic hyetographs for return periods of 2, 10, 25, and 50 years were
defined using the ABM (Figure 8). The block duration was fixed to 0.5 hr in all cases, which is short enough
given the estimated concentration time of the catchments (10 and 5 hr).
The hyetographs shown in Figure 8 were used as the rainfall input in HEC-HMS, in order to compute the flood
discharge for different return periods using the conventional approach. The discretization and parameters
introduced in HEC-HMS models are the same as those used in the proposed methodology, in order to restrict
the differences between approaches to the way in which rainfall variability and AMC is considered.

3. Results and Discussion

3.1. Estimation of Flood Frequency
The flood discharges obtained with the proposed and conventional methods described in the previous sec-
tion were compared for several return periods. The aim of this comparison is to analyze and quantify the rele-
vance of considering site-specific rainfall patterns during storm events. Hence, the same hydrological model
(same structure and same parameters) was used in both approaches. The model structure and parameters
were already described in section 2.2.4. The only difference in the parameterization used with both meth-
odologies is the definition of the CN, since the proposed approach accounts for the variability of AMC, while
the conventional method does not. Therefore, we have applied the conventional approach using the three
AMC considered in the SCS method (dry, normal, and moist), in order to compare the results with those of
the proposed methodology. For comparison purposes, we have also computed the flood discharges using
the RFFA at the global scale presented in Smith et al. (2015), which only requires the location, area, and aver-
age annual precipitation in the study basins.

CEA AND FRAGA 8785

 Water Resources Research 10.1029/2018WR023194

Figure 9. Exceedance probability of flood discharge data and fitted Gumbel distribution in the Mandeo (left) and Mendo
(right) catchments, computed with the proposed methodology.

Figure 9 shows the Gumbel distribution fitted to the daily peak discharge, computed for the Mandeo and
Mendo catchments following the proposed methodology. The flood discharges given by the Gumbel distri-
bution for the return periods of 2, 10, 25, and 50 years are shown in Table 3 and vary between 232 and
527 m3/s in the Mandeo catchment and from 20 to 56 m3/s in the Mendo catchment. Notice that the 90%
confidence intervals in Figure 9 are rather large, especially for the highest return periods, due to the low num-
ber of sample data (the whole study is done with 17 years of available rainfall data). If more years of rainfall
data were available, the confidence intervals should reduce.
The standard approach is strongly dependent on the chosen AMC (dry, normal, or moist). In general, in the
study region the maximum river discharges occur in the wet season, when the soil is under normal or moist
conditions. As it could be expected, the dry AMC underestimates the flood discharges in all cases, and there-
fore, it should not be used in a flood frequency analysis. However, it is difficult to determine a priori which of
the other two AMC conditions (normal or moist) is preferable to use in a conventional flood frequency ana-
lysis. In the Mandeo catchment, for instance, the normal AMC underestimates the flood discharges roughly
by a factor 2 when compared to the proposed methodology. Conversely, the moist AMC overestimates the
discharge, especially for low return periods. On the Mendo catchment, the normal AMC gives very similar
results to the proposed method, while the moist conditions strongly overestimate the flood discharge
(always when compared to the proposed method).
In addition, the chosen AMC has a large impact on the results, making the flood discharge estimation very
sensitive to the decision of using one or another AMC. Discharges computed assuming moist conditions
are, on average, three times larger than those computed assuming normal conditions.
Figure 10 shows the relation between the observed peak discharges and the antecedent rainfall depth for
different number of antecedent days (5 and 34 days in the Mandeo and 5 and 31 days in the Mendo). The
5-day antecedent rainfall (R5) is the one used in the standard application of the SCS method, while the 34-

Table 3
Daily Peak Discharges for Different Return Periods, Estimated With the Proposed and Standard Methodologies, as Well as With the Regional Flood Frequency Analysis
(RFFA) at the Global Scale of Smith et al. (2015)
3 3
Mandeo—QT (m /s) Mendo—QT (m /s)
T
(years) Proposed Standard dry Standard normal Standard moist RFFA Proposed Standard dry Standard normal Standard moist RFFA

2 232 0 92 411 167 20 2 16 76 25

10 389 17 169 558 326 39 6 38 126 49
25 468 26 214 634 414 49 9 52 155 62
50 527 34 250 690 482 56 12 65 176 72
Note. Results obtained with the standard methodology are shown for the three different soil moisture content conditions considered in the Soil Conservation
Service curve number method (dry, normal, and moist).

CEA AND FRAGA 8786

 Water Resources Research 10.1029/2018WR023194

Figure 10. Observed peak discharge measured at the gauge stations of Mandeo (upper) and Mendo (lower) catchments versus antecedent rainfall. The correlation
coefficient between antecedent rainfall and peak discharge is indicated (r).

and 31-day antecedent rainfall (R34 and R31) are the ones used in equations (2) and (3) of the proposed
methodology to estimate the infiltration parameter CN for each rainfall event. Figure 10 shows that in both
catchments, there is a significant correlation between Qp and the antecedent rainfall depth (R34, R31, or R5).
The Spearman’s rank-order correlation coefficients vary between 0.54 and 0.77, with significance levels (p
value) lower than 0.005 in all cases. This is coherent with the results shown in Figure 3 and reinforces the
importance of considering the correct AMC. Despite this correlation, Figure 10 also emphasizes that the
highest peak discharges do not always occur under the same AMC conditions, and thus, it is important to
consider event-specific AMC, or at least the possible combinations of AMC and rainfall intensity.
Coming back to the conventional approach applied to the study basins, if we estimate the AMC according to
R5 and Figure 10, in the Mendo catchment we would say that the moist AMC is the most appropriate to
estimate flood discharges. However, as shown in Table 3, this would produce discharges that are three times
larger than those obtained with the proposed methodology. Moreover, given the increasing trend between
the antecedent rainfall and the peak discharge shown in Figure 10, the AMC that should be considered in a
conventional method should be dependent on the return period. Compared to the standard approaches, the
proposed methodology has the advantage that it incorporates in a natural way the AMC in order to establish
the infiltration parameters to be used during each potentially hazardous storm event.
Table 3 also presents the flood discharges computed with the RFFA at the global scale presented in Smith
et al. (2015). The results of Smith et al. (2015) were obtained using a global database of discharge data and
depend only on the catchment area, average annual rainfall, and climate region. This method can therefore
be applied to any region in the world provided that the drainage area and average annual precipitation in the
basin are known. In our case, both study basins are in the temperate climate zone. The flood discharges
obtained (Table 3) lie within the ones computed with the conventional methodology using normal and moist
AMC conditions. This is physically coherent since, as mentioned before, in humid regions it is expected that

CEA AND FRAGA 8787

 Water Resources Research 10.1029/2018WR023194

Figure 11. Comparison of real and synthetic alternating block hyetographs for several events.

flood discharges occur when the AMC are somewhere between normal and moist. Moreover, RFFA
discharges are quite similar to the ones obtained with the proposed methodology, with absolute relative
differences within 10% and 30%, which is remarkable considering that the regional regressions presented
in Smith et al. (2015) have been developed at a global scale.
3.2. Relevance of the Intraevent Temporal Variability of Rainfall
In order to further analyze the relevance of the temporal variability of rainfall in the evaluation of peak dis-
charges, an equivalent synthetic hyetograph was computed for each of the potentially hazardous events.
This was done by first computing the intensity-duration curve of each event and then applying the ABM with
that intensity-duration curve. The alternating block hyetograph computed for each event in such a way has
the same maximum intensities that the real event for duration periods of 0.5, 1, 2 h, etc. The only difference
between the real and the alternating block hyetographs is the way in which rainfall intensities are organized
in time. Figure 11 shows the real and synthetic alternating block hyetographs computed for several represen-
tative events. In some events, both hyetographs are quite similar. This is for instance the case of the events
occurring on 01 December 2012, 14 November 2010, and 14 January 2016 in Figure 11. However, in other
events the difference between real and synthetic hyetographs is remarkable. Some events have two peak
intensities that cannot be captured by the alternating block hyetograph (e.g., those of 02 December 2015
and 25 October 2006), while in other events, the intensity remains with almost constant high values during
several hours (e.g., on the event of 07 January 2016). In order to quantify the effect that these differences in
the temporal pattern of rainfall have on the computed peak discharges, the synthetic hyetographs for all the
potentially hazardous events were modeled with HEC-HMS, using exactly the same hydrological parameters
(including the same CN) as in the real events. As shown in Figure 12, it is not evident whether the alternating
block hyetograph is more or less conservative than the real measured hyetographs. Actually, in some events
the alternating block hyetograph produces higher peak discharges, while in others it predicts lower dis-
charges, without a clear trend between these differences and the peak discharge. It is remarkable that the
peak discharge estimates obtained with the ABM are unbiased with respect to the ones obtained with the

CEA AND FRAGA 8788

 Water Resources Research 10.1029/2018WR023194

Figure 12. Comparison of peak discharges for each rainfall event computed using the measured hyetographs and the
alternating block hyetographs.

observed hydrographs, and the dispersion on the data shown in Figure 12 is relatively low. This suggests that
intraevent temporal variability might not be so relevant in this case.

4. Conclusions
The two most important processes when modeling rainfall-runoff transformation during storm events are
rainfall and infiltration. However, the methods commonly used to estimate the frequency of flood discharges
in poorly gauged basins make strong simplifications regarding these two processes. Compared to
conventional methods, the methodology proposed in this paper accounts for the observed intraevent spatial
and temporal variability of rainfall, as well as for site-specific antecedent soil moisture conditions. For these
reasons, it provides a more site-specific representation of rainfall-runoff transformation than
conventional methods.
Conventional design storm methods assume that a basin-average T-year rainfall falls uniformly over the
whole catchment. Even if ARF or similar techniques can be used to transform point rainfall to basin-averaged
precipitation, the hypothesis of a spatially uniform rainfall remains unrealistic. The proposed methodology
copes with this problem by working with the observed spatial and temporal variability of rainfall as
represented by the available rain gauge data.
Another problem when using conventional methods is to determine the soil infiltration parameters that
should be considered in the hydrological model, since infiltration is strongly dependent on the AMC, which
might vary from one storm to another. In the study basins, the maximum peak discharges do not always
occur when the soil is under moist conditions, and this makes it difficult to decide which AMC should be
considered when applying the conventional methodology. However, this decision has a very high impact
on the computed flood discharges. In the study basins, there is a factor 3 between the discharges computed
assuming either normal or moist AMC. By modeling each storm event independently, the proposed method
accounts in a natural way for the AMC without making additional hypothesis. In the simulations presented
here, antecedent rainfall is used to account for AMC. However, the recent availability of satellite soil moisture
data from the SMAP mission opens up the possibility of using these data instead of antecedent rainfall to
account directly for AMC in future studies.
The main shortcoming of the proposed methodology is that it requires some streamflow data in order to
calibrate the relation between antecedent rainfall and CN. This prevents its applicability to basins with no
discharge data at all. In order to extend the applicability of the method from poorly gauged basins to
ungauged basins, a geographically generalizable relationship between antecedent rainfall and CN would
be required. This possibility should be explored in the future through the application of the method to a
much larger number of catchments.
Given the last developments in hydrological models and the increasing capacity of computers, we consider
that flood frequency analysis in poorly gauged basins should focus on incorporating an accurate

CEA AND FRAGA 8789

 Water Resources Research 10.1029/2018WR023194

representation of rainfall and infiltration during individual storm events, rather than in the definition of spe-
cific design storms or AMC for a given return period.

Acknowledgments References
This work was financially supported by
the Spanish Ministry of Economy and Allen, R. J., & DeGaetano, A. T. (2005). Areal reduction factors for two eastern United States regions with high rain-gauge density. Journal of
Competitiveness (Ministerio de Hydrologic Engineering, 10(4), 327–335.
Economía y Competitividad) within the Asquith, W. H., & Famiglietti, J. S. (2000). Precipitation areal-reduction factor estimation using an annual-maxima centered approach. Journal
project “CAPRI: Probabilistic flood of Hydrology, 230(1–2), 55–69. https://doi.org/10.1016/S0022-1694(00)00170-0
prediction with high resolution Bacchi, B., & Ranzi, R. (1996). On the derivation of the areal reduction factor of storms. Atmospheric Research, 42(1–4), 123–135. https://doi.
hydrologic models from radar rainfall org/10.1016/0169-8095(95)00058-5
estimates” (CGL2013-46245-R). Ignacio Blazkova, S., & Beven, K. (2004). Flood frequency estimation by continuous simulation of subcatchment rainfalls and discharges with the aim
Fraga received financial support from of improving dam safety assessment in a large basin in the Czech Republic. Journal of Hydrology, 292(1–4), 153–172. https://doi.org/
the Xunta de Galicia (Centro Singular de 10.1016/j.jhydrol.2003.12.025
Investigación de Galicia accreditation Brocca, L., Melone, F., & Moramarco, T. (2011). Distributed rainfall-runoff modelling for flood frequency estimation and flood forecasting.
2016–2019) and the European Union Hydrological Processes, 25(18), 2801–2813. https://doi.org/10.1002/hyp.8042
(European Regional Development Cabalar-Fuentes, M. (2005). Los temporales de lluvia y viento en Galicia. Propuesta de clasificación y análisis de tendencias (1961-2001).
Fund-ERDF). The historical rainfall data Investigaciones Geográficas (Esp), 36.
from the pluviometers used in this work, Camici, S., Tarpanelli, A., Brocca, L., Melone, F., & Moramarco, T. (2011). Design soil moisture estimation by comparing continuous and storm-
as well as the discharge data at the two based rainfall-runoff modeling. Water Resources Research, 47, W05527. https://doi.org/10.1029/2010WR009298
gauge stations, can be freely Cunnane, C. (1985). Factors affecting the choice of distribution for flood series. Hydrological Sciences Journal, 30(1), 25–36. https://doi.org/
downloaded from the webpage of the 10.1080/02626668509490969
Galician Meteorological Agency Dawdy, D. R., Griffis, V. W., & Gupta, V. K. (2012). Regional flood-frequency analysis: How we got here and where we are going. Journal of
MeteoGalicia (http://www.meteogalicia. Hydrologic Engineering, 17(9), 953–959. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000584
gal/observacion). The DTM of both De Lima, J. L. M. P., & Singh, V. P. (2002). The influence of the pattern of moving rainstorms on overland flow. Advances in Water Resources,
catchments can be freely downloaded 25(7), 817–828. https://doi.org/10.1016/S0309-1708(02)00067-2
from the webpage of the Centro Eiras-Barca, J., Brands, S., & Miguez-Macho, G. (2016). Seasonal variations in North Atlantic atmospheric river activity and associations with
Nacional de Información Geográfica anomalous precipitation over the Iberian Atlantic Margin. Journal of Geophysical Research: Atmospheres, 121, 931–948. https://doi.org/
(http://centrodedescargas.cnig.es). 10.1002/2015JD023379
Etoh, T., Murota, A., & Nakanishi, M. (1987). SQRT-exponential type distribution of maximum. In V. P. Singh (Ed.), Hydrologic frequency mod-
eling (pp. 253–264). Dordrecht: Springer. https://doi.org/10.1007/978-94-009-3953-0_17
Falter, D., Schröter, K., Dung, N. V., Vorogushyn, S., Kreibich, H., Hundecha, Y., et al. (2015). Spatially coherent flood risk assessment based on
long-term continuous simulation with a coupled model chain. Journal of Hydrology, 524, 182–193. https://doi.org/10.1016/j.
jhydrol.2015.02.021
Fleming, M. J., & Doan, J. H. (2009). HEC-GeoHMS geospatial hydrologic modelling extension: User’s manual version 4.2. Davis, CA: US Army
Corps of Engineers, Institute for Water Resources, Hydrologic Engineering Centre.
Gringorten, I. I. (1963). A plotting rule for extreme probability paper. Journal of Geophysical Research, 68(3), 813–814. https://doi.org/10.1029/
JZ068i003p00813
Gupta, H. V., Sorooshian, S., & Yapo, P. O. (1999). Status of automatic calibration for hydrologic models: Comparison with multilevel expert
calibration. Journal of Hydrologic Engineering, 4(2), 135–143. https://doi.org/10.1061/(ASCE)1084-0699(1999)4:2(135)
Hansen, E. M., L. C. Schreiner, and J. F. Miller (1982), Application of probable maximum precipitation estimates: United States east of the 105th
meridian (Vol. 55). US Department of Commerce, National Oceanic and Atmospheric Administration.
Hope, A. S., & Schulze, R. E. (1982). Improved estimates of stormflow volume using the SCS curve number method. In V. P. Singh (Ed.), Rainfall-
runoff relationships (pp. 419–431). Littleton: Water Resources Publications.
Liang, J., & Melching, C. S. (2015). Experimental evaluation of the effect of storm movement on peak discharge. International Journal of
Sediment Research, 30(2), 167–177. https://doi.org/10.1016/j.ijsrc.2015.03.004
Lobligeois, F., Andréassian, V., Perrin, C., Tabary, P., & Loumagne, C. (2014). When does higher spatial resolution rainfall information improve
streamflow simulation? An evaluation using 3620 flood events. Hydrology and Earth System Sciences, 18(2), 575–594. https://doi.org/
10.5194/hess-18-575-2014
Midttømme, G., Pettersson, L. E., Holmqvist, E., Nøtsund, Ø., Hisdal, H., & Sivertsgård, R. (2011). Retningslinjer for flomberegninger (Guidelines
for flood calculations). Retningslinjer no. 04/2011 (59 pp.). Retrieved from www.nve.no
Mizumura, K., & Ito, Y. (2011). Influence of moving rainstorms on overland flow of an open book type using kinematic wave. Journal of
Hydrologic Engineering, 16(11), 926–934. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000398
Moretti, G., & Montanari, A. (2008). Inferring the flood frequency distribution for an ungauged basin using a spatially distributed rainfall-
runoff model. Hydrology and Earth System Sciences, 12(4), 1141–1152. https://doi.org/10.5194/hess-12-1141-2008
Nash, J. E., & Sutcliffe, J. V. (1970). River flow forecasting through conceptual models. Part I—A discussion of principles. Journal of Hydrology,
10(3), 282–290. https://doi.org/10.1016/0022-1694(70)90255-6
Natural Resources Conservation Service (2007). National engineering handbook: Chapter 16 Hydrographs. Washington, DC: Natural Resources
Conservation Service.
NERC (1975). Flood studies report. UK: Natural Environmental Research Council.
Niemi, T. J., Guillaume, J. H., Kokkonen, T., Hoang, T. M., & Seed, A. W. (2016). Role of spatial anisotropy in design storm generation:
Experiment and interpretation. Water Resources Research, 52, 69–89. https://doi.org/10.1002/2015WR017521
NSSP (1961). National severe storms project objectives and basic design. Report no. 1, U.S. Weather Bureau.
Rodriguez-Iturbe, I., & Mejía, J. M. (1974). On the transformation of point rainfall to areal rainfall. Water Resources Research, 10(4), 729–735.
https://doi.org/10.1029/WR010i004p00729
Scharffenberg, W. A., & Fleming, M. J. (2006). Hydrologic modeling system HEC-HMS: User’s manual. US Army Corps of Engineers, Hydrologic
Engineering Center.
Schulze, R. E. (1982). The use of soil moisture budgeting to improve stormflow estimates by the SCS curve number method (p. 63).
Pietermaritzburg: University of Natal, Department of Agricultural Engineering.
Shen, Z., Chen, L., Liao, Q., Liu, R., & Hong, Q. (2012). Impact of spatial rainfall variability on hydrology and nonpoint source pollution
modeling. Journal of Hydrology, 472, 205–215.

CEA AND FRAGA 8790

 Water Resources Research 10.1029/2018WR023194

Sigaroodi, S. K., & Chen, Q. (2016). Effects and consideration of storm movement in rainfall–runoff modelling at the basin scale. Hydrology and
Earth System Sciences, 20(12), 5063–5071. https://doi.org/10.5194/hess-20-5063-2016
Sivapalan, M., & Blöschl, G. (1998). Transformation of point rainfall to areal rainfall: Intensity-duration-frequency curves. Journal of Hydrology,
204(1–4), 150–167. https://doi.org/10.1016/S0022-1694(97)00117-0
Smith, A., Sampson, C., & Bates, P. (2015). Regional flood frequency analysis at the global scale. Water Resources Research, 51, 539–553.
https://doi.org/10.1002/2014WR015814
USDA, Soil Conservation Service (1986). Technical Release 55: Urban Hydrology for Small Watersheds (2nd ed.).
Vaes, G., Willems, P., & Berlamont, J. (2005). Areal rainfall correction coefficients for small urban catchments. Atmospheric Research, 77(1–4),
48–59. https://doi.org/10.1016/j.atmosres.2004.10.015
Veneziano, D., & Langousis, A. (2005). The areal reduction factor: A multifractal analysis. Water Resources Research, 41, W07008. https://doi.
org/10.1029/2004WR003765
Wright, D. B., Smith, J. A., & Baeck, M. L. (2014a). Critical examination of area reduction factors. Journal of Hydrologic Engineering, 19(4),
769–776. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000855
Wright, D. B., Smith, J. A., & Baeck, M. L. (2014b). Flood frequency analysis using radar rainfall fields and stochastic storm transposition. Water
Resources Research, 50, 1592–1615. https://doi.org/10.1002/2013WR014224
Wright, D. B., Smith, J. A., Villarini, G., & Baeck, M. L. (2013). Estimating the frequency of extreme rainfall using weather radar and stochastic
storm transposition. Journal of Hydrology, 488, 150–165. https://doi.org/10.1016/j.jhydrol.2013.03.003

CEA AND FRAGA 8791