See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223292116

Influence of Rainfall Spatial Variability on Flood Prediction

Article in Journal of Hydrology · March 2002

DOI: 10.1016/S0022-1694(01)00611-4

CITATIONS READS
238 751

4 authors, including:

Christophe Bouvier Leonardo Cisneros

Institute of Research for Development Qassim University
113 PUBLICATIONS 2,582 CITATIONS 4 PUBLICATIONS 275 CITATIONS

SEE PROFILE SEE PROFILE

Ramón Mora Domínguez

Universidad Nacional Autónoma de México
98 PUBLICATIONS 804 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Hydrological hazards in the Mediterranean region under climate change View project

Water reuse, regulation and policy View project

All content following this page was uploaded by Ramón Mora Domínguez on 06 February 2020.

The user has requested enhancement of the downloaded file.

 Journal of Hydrology 260 (2002) 216±230
www.elsevier.com/locate/jhydrol

In¯uence of rainfall spatial variability on ¯ood prediction

Patrick Arnaud a,*, Christophe Bouvier b, Leonardo Cisneros c, Ramon Dominguez c
a
ENGEES, 1 Quai Koch, BP 1039F, 67070 Strasbourg, France
b
IRD, BP 5045, 34032 Montpellier Cedex, France
c
UNAM, Instituto de Ingenieria, CD. Universitaria, Apdo. Postal 70-472, 04510 Mexico D.F., Mexico
Received 31 October 2000; revised 14 September 2001; accepted 18 December 2001

Abstract
This paper deals with the sensitivity of distributed hydrological models to different patterns that account for the spatial
distribution of rainfall: spatially averaged rainfall or rainfall ®eld. The rainfall data come from a dense network of recording rain
gauges that cover approximately 2000 km 2 around Mexico City. The reference rain sample accounts for the 50 most signi®cant
events, whose mean duration is about 10 h and maximal point depth 170 mm. Three models were tested using different runoff
production models: storm-runoff coef®cient, complete or partial interception. These models were then applied to four ®ctitious
homogeneous basins, whose sizes range from 20 to 1500 km 2. For each test, the sensitivity of the model is expressed as the
relative differences between the empirical distribution of the peak ¯ows (and runoff volumes), calculated according to the two
patterns of rainfall input: uniform or non-uniform. Differences in ¯ows range from 10 to 80%, depending on the type of runoff
production model used, the size of the basin and the return period of the event. The differences are generally moderate for
extreme events. In the local context, this means that uniform design rainfall combining point rainfall distribution and the
probabilistic concept of the areal reduction factor could be suf®cient to estimate major ¯ood probability. Differences are more
signi®cant for more frequent events. This can generate problems in calibrating the hydrological model when spatial rainfall
localization is not taken into account: a bias in the estimation of parameters makes their physical interpretation dif®cult and
leads to overestimation of extreme ¯ows. q 2002 Elsevier Science B.V. All rights reserved.
Keywords: Rainfall variability; Distributed hydrological model; Rainfall ®elds; Sensitivity; Flood estimation; Mexico

1. Introduction Many studies have dealt with the sensitivity of

hydrological models together with other uncertainties
The objective of this work was to study the sensi- related to spatial rainfall variability, i.e. with the
tivity of distributed hydrological models to the spatial spatial density of the rain gauge network and with
distribution of rainfall, and its in¯uence on the estima- the interpolation methods (Creutin et al., 1980;
tion of ¯ood probability. We compare both runoff Kuczera and Williams, 1992; FaureÁs et al., 1995;
volumes and peak ¯ows, calculated using either rain- Goovaerts, 2000; Wotling et al., 2000); with rainstorm
fall ®elds (i.e. hyetographs associated with different displacement (Surkan, 1974; Gupta and Waymire,
spatial observations) or spatially averaged rainfalls 1979; Marshall, 1980; Niemczynowicz, 1987, 1991).
(i.e. a synthetic hyetograph obtained by weighted These studies show that the spatial distribution of
average of the different observations). rainfall should be taken into account in any catchment
area. It in¯uences not only runoff volumes and peak
* Corresponding author. Fax: 133-3-88-24-82-84. ¯ows, but also the time shift of hydrographs (Dawdy
E-mail address: [email protected] (P. Arnaud). and Bergman, 1969; Wilson et al., 1979; Troutman,
0022-1694/02/$ - see front matter q 2002 Elsevier Science B.V. All rights reserved.
PII: S 0022-169 4(01)00611-4
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 217

1983; Krajewski et al., 1991). Moreover, it signi®- may occur in calibrating the rainfall-runoff models if
cantly increases uncertainty about the estimation of the spatial organization of rainfall is not taken into
hydrological model parameters, and consequently account.
estimation of the extreme quantiles (Kuczera and
Williams, 1992; Obled et al., 1994).
Studies on the advantages of using rainfall ®elds as 2. Methodology
opposed to spatially averaged rainfall are relatively
rare. If the sensitivity of the hydrological model to 2.1. The study area and the data
these two patterns of rainfall representation is low,
the use of uniform areal rainfall derived from point The study was carried out in the region around
rainfall distribution combined with the probabilistic Mexico City, which is characterized by a vast, origin-
areal reduction concept (Brunet-Moret and Roche, ally endoreic basin surrounded by several mountain
1966; Rodriguez-Iturbe and Mejia, 1974; Lebel and chains. The mean altitude of the basin is 2240 m, but
Laborde 1988; Sivapalan and BloÈschl, 1998) is justi®- the surrounding mountains often exceed 3000 m and
able in order to generate design rainfall events. Conver- the highest volcanoes reach more than 5000 m (Popo-
sely, it is necessary to generate rainfall ®elds (Waymire catpetl and Ixatccihuatl, in the southeast). Both the
et al., 1984; Smith and Karr, 1985; Bell, 1987; relief and human activities have induced high variabil-
Rodriguez-Iturbe and Eagleson, 1987; Sivapalan and ity in soils and in land use. On the slopes, vegetation is
Wood, 1987; Yoo et al., 1996), which are used as abundant (pine or eucalyptus woods, extensive pastures
probabilized inputs in a distributed hydrological model. or agricultural areas) and andosoils are predominant,
Morin et al. (1995) showed on small Israeli basins while the ¯at central part of the catchment is character-
how spatial variability of rainfall, obtained by radar, ized by extensive urban areas and plastic clays.
can generate signi®cant variations in ¯ow modeling. The climate is of the tropical altitude type, charac-
Obled et al. (1994), using TOPMODEL for the Real terized by one short rainy season (here from June to
Collobrier 71 km 2 experimental catchment located in September). The mean annual rainfall is approxi-
southeast France, showed that ¯ows computed using mately 600 mm at an altitude of 2240 m, and up to
spatially averaged rainfall or rainfall ®eld are not very 1000 mm in the mountains. In spite of the tropical
different. However, as it is generally the case in this context, the rainfall intensity for short time steps is
type of study, conclusions may be very speci®c, rather low: for a 10-year return period, rainfall inten-
depending on the scale of the basin, rainfall variability sity for steps of 5, 60 min and 24 h are, respectively,
in the area, and the mechanisms involved in the 164, 46, 2.5 mm/h, with corresponding amounts of
generation of ¯ows. 6.8, 46, 60 mm. This is typical of the tropical or
In this study, rainfall data are from a dense rainfall equatorial altitude climates in South America, and is
network located in the Mexico City region. The sensi- comparable to other sites like Quito (Ecuador), La Paz
tivity of the distributed model to the rainfall input (Bolivia) (Bouvier et al., 1999).
(spatially averaged or rain ®eld) was studied by Rainfall records came from a rain gauge network
comparing the statistics for both peak ¯ows and covering an area of nearly 2500 km 2, and during the
volumes. To ensure wider application of the conclu- period 1988±1998, included 49 tipping bucket rain
sions of this study, these comparisons were performed gauges. Five recording rain gauges were judged to
for: be substandard and were discarded from this study.
Rain events were delimited in time by the absence
different types of rainfall-runoff conversion models of rainfall for a period of at least 1 h at all the stations.
different sizes of basin The duration of the events range between 9 and 126 h
different return periods (median 40 h), while maximal rainfall depths vary up
to 172 mm. The spatial correlations between point
The discussion focuses on the suitability to use cumulative rainfalls can be expressed through the
spatially averaged rainfall combined with areal reduc- climatologic variogram, which can be ®tted to a
tion factor for design rainfall, and on problems that spherical model (no nugget effect, sill ˆ 1.25,
218 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Table 1 where ai is the Thiessen coef®cient associated to i

Characteristics of the sample of 50 rain events station, Pi the cumulative rainfall at the i station,
Median Minimum Maximum and N the number of recording stations.
The spatial coef®cient of variation (CVS) was
Duration (h) 40 9 126 de®ned by:
Local maximum cumulative 85.8 70.1 172.2
v
rainfall (mm): PMAX uX
uN
Cumulative basin rainfall 21.9 2.0 80.7 t a P 2 PBAS†2
i i
(mm): PBAS
iˆ1
R ˆ PMAX=PBAS 3.4 1.54 20 CVS ˆ 1†
CVS median 0.73 0.27 3.03 PBAS
Rainfall gauges number 4 1 33
where P . PMAX=2
The spatial extent of rainfall may differ considerably
from one event to another: some are very localized
and others cover the whole catchment area, as shown
range ˆ 30 km). To give some idea of the spatial in Fig. 1. The ®rst event is a local rain event, which
extent of the rain ®elds, the mean area of the 10 mm mainly affects only one station and which represents
isohyet is about 500 km 2 (Cisneros et al., 1998). the highest maximum local cumulative rainfall
We ®nally retained the 50 most signi®cant events to (PMAX ˆ 172.2 mm; PBAS ˆ 27.5 mm) observed.
form our reference sample, which will be used as The second event is a widespread rain event, which
input data for the distributed hydrological model. affects the whole of the study area and represents the
These events are those with a rainfall depth of at highest cumulative basin rainfall (PMAX ˆ 140.2
least 70 mm at one of the stations. Some characteris- mm; PBAS ˆ 80.7 mm) observed.
tics of the sample are given in Table 1.
The cumulative basin precipitation (PBAS) was
2.2. The hydrological models
calculated by using the Thiessen polygon method:
MERCEDES (Bouvier et al., 1994; Bouvier and
X
N
PBAS ˆ ai Pi Delclaux, 1996), is a modeling toolkit, that contains
iˆ1 a wide variety of conceptual or physically based

Fig. 1. Maps of the isohyets of two extreme events.

 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 219

Fig. 2. In¯uence of the runoff production model on runoff.

rainfall-runoff models. All of them operate with a discharge that produces additional runoff. Discharge
discretization of the catchment into regular grid is a function of both the level of the reservoir at the
squares whose size is de®ned by the user (in our beginning of the time interval and the intercepted
case, grid mesh length varied from 30 to 250 m, see rainfall. It can be considered as a variable storm-
Section 2.3). The precipitation Pm,t received by each runoff coef®cient whose value increases with the
grid mesh (m) at time step (t) is interpolated in space reservoir level, and varies from 0 to 100%:
from the observed hyetographs (see Section 2.4). For
each grid mesh (m) and each time step (t), the effective b2stocm;t 1 Pm;t c
Pem;t ˆ Pm;t
precipitation Pem,t that contributes to runoff is calcu- 2STOm
lated using different runoff production models. In this
if Pm;t # STOm 2 stocm;t
study, we tested successively three runoff production
models coupled with the same routing model. To
render further analysis more readable, each of the Pem;t ˆ Pm;t 2 STO 2 stocm;t ††
production models accounts only for one parameter,
but their different features may describe a rather broad 1 STO 2 stoct;m †
range of hydrological behaviors.
b2stocm;t 1 STO 2 stoct;m †c
Model 1 (complete interception): runoff begins £
2STOm
after the cumulative rainfall exceeds a threshold
value STO (in mm). Afterwards, the whole excess if Pm;t . STOm 2 stocm;t
cumulative rainfall is converted into runoff. This
model enables runoff due to soil saturation in water. where stocm;t being the ®lling level of the reservoir at
X time step t. This level was ®xed arbitrarily at zero at
Pem;t ˆ 0 if Pm;j , STOm
the beginning of the rain event.
jˆ0;t
This model enables some decreasing in®ltration,
X which can be compared to the transitory phase of a
Pem;t ˆ Pm;t when Pm;j . STOm
Hortonian process. In this case, the ®nal in®ltration
jˆ0;t
tends towards zero. It must also be noted that the soil
Model 2 (constant storm-runoff coef®cient): effec- reservoir level cannot decrease during a given event,
tive precipitation is directly proportional to precipita- which means that neither percolation into the deep aqui-
tion intensity. Wm (proportionality factor) can be fer nor evapotranspiration are taken into account at the
interpreted as a constant storm-runoff coef®cient event scale (this concerns both the models 3 and 1). This
over time. It can be considered for example as the would have been of course possible, but to ensure the
imperviousness coef®cient of an urban catchment. clarity of the analysis, we preferred the models we use to
be as simple as possible. The in¯uence of this simpli®ed
Pem;t ˆ Wm Pm;t
assumption is further discussed in Section 4.
Model 3 (partial interception): this is an inter- The graphs in Fig. 2 show the responses of the three
mediate case between models 1 and 2. Rainfall ®rst runoff production models subjected to a total rainfall
®lls a reservoir with a storage capacity STO (in mm), of 80 mm and parameterized in order to generate an
as in Model 1, but this reservoir has a continuous effective rainfall close to 60 mm. The corresponding
220 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Fig. 3. Routing model.

parameters are STO ˆ 20 mm for Model 1, W ˆ 0:75 celerity does not change with time, therefore the travel
for Model 2 and STO ˆ 30 mm for Model 3. time to the outlet is invariant for a given mesh. This
For a given grid mesh at a given time, each elemen- implies the same linearity properties as those of the
tary runoff contribution is then routed to the outlet, unit hydrograph, of which our routing model can be
using a lag and route model. The model needs two considered as a distributed form.
estimates (Fig. 3):
2.3. Reference basins
travel time Tm from the grid P mesh to the outlet,
which is expressed by Tm ˆ iˆ1;k Li =V0i ; where Four ®ctitious basins of different surface areas were
k is the number of grid mesh traveled between the de®ned. In order to mainly study the in¯uence of
grid mesh (m) and the outlet, Li and V0i are the spatial rainfall information, several simpli®ed
length and the ¯ow celerity associated with each assumptions were made.
grid mesh (i). First, we chose to conserve the same morphology
storage capacity Km during the travel time, which for all four basins. This was performed by applying
has units in time and corresponds to the parameter various reduction ratios to the grid mesh area: 1/1 for
of a linear storage model (Chocat, 1997, p. 867; the largest basin, then roughly 1/2, 1/4 and 1/8. The
Maidment, 1993, pp. 26±29). It is assumed that number of grid squares remains the same for all four
Km may be related to the travel time Tm by a linear basins (176 £ 232), but their size varies from 250 to
function Km ˆ K0m Tm ; where K0m is a non-dimen- 125, 65 and 30 m basins 1±4, respectively, and their
sional characteristic of the grid mesh (m). surface area is thus, respectively, 1531, 383, 103 and
22 km 2. Next, these basins were located in the study
So the routing model is fully determined by the V0 zone in such a way as to include both the highest
and K0 values of each grid mesh. At the outlet, the number of recording rain gauges and the strongest
elementary hydrograph due to an effective rainfall rainfalls (Fig. 4), in order to grab the highest possible
Pem;t of a mesh (m) over time ‰t0 ; t01Dt Š; is expressed spatial variability of the most signi®cant rainfalls.
by: Although very arti®cial, this method allowed to mini-
Q t† ˆ 0 if t , t0 1 Tm mize the sample effects due to irregular network
density or localized extreme events.
Q t† ˆ Pem;t =Km e2 t0 1Tm 2t†=Km if t . t0 1 Tm Second, the basins are considered to be spatially
homogeneous, i.e. both the runoff production condi-
The complete hydrograph results from the addition of tions (parameters STO or W) and the routing condi-
the elementary hydrographs provided by each grid tions (parameters V0 and K0) are constants on all grid
mesh at each time step. It must be noted that the V0 meshes. The initial humidity conditions are also
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 221

Fig. 4. Location of recording rain gauges and basin limits.

considered to be spatially homogeneous. Events are distance, spline function, kriging). However, these
assumed to be independent, and initial humidity methods did not signi®cantly modify the conclusions
conditions here correspond to dry conditions. of the study; this can be explained by a suf®cient
network density (Cisneros et al., 2000). Thus the
2.4. Tests simplest method, mentioned earlier, was adopted.
Tests were carried out for different models, differ-
The different tests consisted in comparing the ¯ows ent basin sizes, different time steps, and different para-
calculated according to the two patterns of rainfall meter values. In each case, the tests consisted of
input used by the distributed hydrological model: comparing the empirical distribution of runoff
uniform basin rainfall (UR: `uniform rainfall') and volumes and peak ¯ows, calculated for the 50 rain
rainfall ®eld (NUR: `non-uniform rainfall'). The ®rst, events as a function of either uniform or non-uniform
UR, is represented by a synthetic hyetograph obtained patterns of the rainfall variability. The empirical
by the weighted average of different hyetographs cumulative frequencies were calculated by Hazen
observed in space. This synthetic hyetograph is calcu- formula: F Xi † ˆ i 2 0:5†=n ˆ Probability X # Xi †;
lated for each time step using the Thiessen coef®cients, where i indicates the rank of the observation xi ranked
and each mesh of the basin receives the same quantity in ascending order, and n the total number of observa-
of rainfall. The second, NUR, is represented by the tions n ˆ 50†:
different hyetographs observed in space; the rainfall The comparison of the empirical distributions was
received by each mesh is that of the nearest station then expressed using the relative errors between the
(hypothesis of Thiessen polygon method). Other meth- quantiles of the same rank, calculated in the following
ods were tested for the construction of uniform rainfall way:
(centered hyetograph method shifted in time in order to
have the maximum intensities coincide), as well as XURi 2 XNURi
Ei %† ˆ £ 100 2†
other modes of spatial interpolation (weighted inverse XNURi
222 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Fig. 5. Relative errors on the quantiles of peak ¯ows and runoff volumes calculated from a UR and a NUR: in¯uence of the runoff production
model.

where XUR and XNUR are the values of the quantiles Modeling based on runoff production by a constant
resulting from the distribution of the peak ¯ows (or runoff coef®cient (Model 2), is not very sensitive to
runoff volumes) calculated from the UR and from the the spatial distribution of rainfall. Errors in runoff
NUR, respectively. volumes are very small, and may even be zero for
most of the cumulative frequencies. This can be
explained, in this particular case, by the linearity of
3. Results the transformation of precipitation into effective rain-
fall. Thus, the overestimation of peak ¯ows obtained
3.1. In¯uence of model type with NUR is only due to the effect of the position of
storm cells on the effective rainfall routing. This effect
In this part, the models were applied to the largest generates errors ranging from 220 to 210% on peak
basin S ˆ 1531 km2 †: Calculations were carried out ¯ow quantiles resulting from the two patterns of rain-
at an hourly time step. fall representation.
In contrast, runoff production by surpassing of a
3.1.1. In¯uence of the runoff production model threshold on cumulative rainfall (Model 1), may
The runoff production parameter was ®xed at strongly underestimate peak ¯ows and runoff volumes
STO ˆ 20 mm for Model 1, W ˆ 0:75 for Model 2 calculated using a uniform rainfall. The errors are
and STO ˆ 30 mm for Model 3 (see example in maximum (2100%) for 21 events, for which the
Section 2.2, Fig. 2). These values guarantee the ¯ows calculated with UR are zero. For these events,
same mean rainfall-runoff coef®cient for the 50 smoothing due to the spatial average rainfall is signif-
studied events. The routing parameters were ®xed at icant, and the cumulative uniform rainfall does not
V0 ˆ 1 m=s and K0 ˆ 0:7 for each grid mesh. These exceed the imposed threshold (here, 20 mm), whereas
values are considered to be suitable in the studied area this threshold may be exceeded locally in several
(Bouvier et al., 1994). places. Such events are generally not very widespread
For each of the three runoff production models, events. For more signi®cant ¯ows, the spatial average
coupled with the same routing model, the empirical rainfall exceeds the imposed threshold and the rainfall
frequency distributions of the peak ¯ows and the excess fully participates in the runoff. Furthermore,
runoff volumes were calculated from the sample of when rainfall considerably exceeds the imposed
50 rain events. The relative errors (Ei) de®ned in Eq. threshold, Model 2 acts like Model 1, with a runoff
(2) were calculated for the different models of runoff coef®cient close to 100%, and that is the reason why
production. They were drawn against the correspond- we ®nd similar errors of around 210% in both
ing cumulative frequency i 2 0:5†=50 (Fig. 5). To models. Finally, results obtained with Model 2 are
facilitate the interpretation of results, we have quite different, but it is worth noting that in the case
smoothed the curves by a moving average calculated of an extreme event, the differences resulting from
on three values (Fig. 5). using UR and NUR are not very great.
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 223

Fig. 6. Relative errors on the quantiles of peak ¯ows and runoff volumes calculated from a UR and a NUR: in¯uence of the runoff production
parameter for the models 1 and 3.

Model 3 presents an intermediate behavior, which three strongest peak ¯ows calculated are due to very
is in agreement with the way effective rainfall is widespread events: these concern, respectively, 31, 37
calculated. The threshold effect, caused by the priority and 16 out of 44 stations, where basin rainfall is over
given to reservoir ®lling, is moderated by its variabil- PMAX/2 (the median value is 4, cf. Table 1), and
ity over time, allowing runoff from the beginning of represent three out of four of the strongest basin rain-
the rainfall. But, in so doing, the reservoir storage falls observed (81.2, 55.9 and 59.7 mm).
effect is effective for longer, and affects higher rain-
falls. That is the reason why errors are bigger for rare 3.1.2. In¯uence of the runoff production parameters
events than in models 1 and 2, about 230 to 220%. The conclusions did not change when parameters of
In any runoff production model, relative error the three runoff production models were modi®ed. For
decreases with an increase in cumulative frequency, Model 2, we checked that we obtained exactly the
which can be explained as follows: ®rstly, as seen same relative errors regardless of the value repre-
earlier in models 1 and 3, the threshold effect of the sented by coef®cient W. For the other models,
runoff production model decreases towards the comparison was made using STO ˆ 10 mm (weak
extreme events. Absolute errors vary little, thus rela- value) instead of the initial values (high value),
tive errors decrease (see relative error de®nition: Eq. STO ˆ 20 mm for Model 1, STO ˆ 30 mm for
(2)); secondly, the biggest ¯ows are generally Model 3 (Fig. 6).
obtained for rain events that affect the whole basin By increasing the effect of storage in Model 1, rela-
and have lower spatial heterogeneity of cumulative tive errors increase for frequent events, but remain
rainfall, and are therefore less sensitive to smoothing identical for rare events. In the case of Model 3, the
due to the spatial average. In the sample used, the increase in error is constant (more or less 20%) over
the whole range of cumulative frequencies. It can be
thus concluded that the storage capacity of soil in a
pure cumulative saturation soil process (Model 1)
would have a strong effect overall on frequent events,
whereas it affects rare events by means of some kind
of Hortonian process (Model 3).

3.1.3. In¯uence of the routing model parameters

The in¯uence of the routing model was analyzed by
changing the value of the V0 parameter from 0.25 to
4 m/s K0 ˆ 0:7†: This range of variation is large and
arbitrary but it covers a wide range of situations that
Fig. 7. Relative errors on the peak ¯ows quantiles calculated from a may be encountered. Production Model 3 was selected
UR and a NUR: in¯uence of the routing model parameters. for the test, with STO ˆ 30 mm:
224 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Table 2
Characteristics of the four basins studied

Basin 1 Basin 2 Basin 3 Basin 4

Surface area (km 2) 1531 383 103 22

Number of recording stations 44 20 6 4
R median (de®ned in Table 1) 3.4 2.7 1.6 1.4

The peak ¯ow values do, of course, increase with to now is a relatively large basin (1531 km 2), in which
transfer speed, but the relative errors presented in Fig. the spatial variability of rainfall is documented by 44
7 remain very similar for each value of V0. This can rain gauges. In order to study the in¯uence of basin
be explained by the linearity of the routing model, size, we examined the results obtained using other
which means that the peak ¯ows are proportional to smaller ®ctitious basins, whose characteristics are
the runoff volumes: since runoff volume errors were summarized in Table 2.
not changed in the different tests because the same Comparison was performed using the `intermedi-
production model was used, linearity of the routing ate' model: runoff production Model 3 with STO ˆ
model ensures that peak ¯ows error are identical. 30 mm; V0 ˆ 1 m=s and K0 ˆ 0:7 for the routing
Using runoff production models 1 and 2 leads to the model, and an hourly time step (Fig. 8).
same conclusions. It can be seen that the relative errors are generally
In this part of the study, we were also interested in much greater for the largest basins (1 and 2) than for
the in¯uence of the time step on results. Calculations the smallest (3 and 4), although there is a relative
were thus carried out with a shorter time step (5 min), decrease in these differences towards the rare frequen-
which gave the same results as with a 60 min time cies. This can be explained by the fact that the spatial
step. This is mainly due to the fact that the residence variability of rainfall increases with catchment size.
time in the basins studied is still much longer than the This result is in agreement with the assumption that
time step investigated. Any in¯uence of the time step the mean spatial extent of rainfall is around a few
would only be effective if it exceeded the residence hundred square kilometers in the Mexico basin. This
time. This may occur in a small basin with a high is also con®rmed by the value of the range of the
transfer speed, where the hourly time step is some- spherical variogram model, 30 km. In the same way,
times insuf®cient. the analysis of the spatial coef®cient of variation
(CVS) shows that mean values decrease from 0.73 to
3.2. In¯uence of basin size 0.29 with an increase in the size of the basin (Table 3).
However, a bias may be caused by different number of
The basin that has been used for the simulations up rain gauges, which is very low in the smallest basins.

Fig. 8. Relative errors on the quantiles of peak ¯ows and runoff volumes, calculated from a UR and a NUR: in¯uence of basin size.
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 225

Table 3
Median CVS values for station sampled

Basin size (km 2) 1531 383 103 22

CVS median (all stations) 0.73 (basin 1) 0.67 (basin 2) 0.35 (basin 3) 0.29 (basin 4)
CVS median (1 station out of 2) 0.72 (basin 1) 0.70 (basin 2) 0.32 (basin 3) 0.18 (basin 4)
CVS median (median of 100 ± 0.63 0.46 0.23
generated basins)
Range of CVS median ± 0.55±0.67 0.32±0.61 0±0.46

Obviously, if this number is reduced to one, both rela- 3.3. In¯uence on both calibration and extrapolation of
tive errors and CVS will be zero. Thus, CVS were also a rainfall-runoff model
calculated to account for only one gage out of two in
each catchment, and it will be noted that their mean As mentioned earlier, the advantage of using rain
values do not change (Table 3), except in the smallest ®elds in rainfall-runoff modeling is relatively weak in
basin, where the number of stations is too small for a the case of extreme events. The error ranges from 230
signi®cant comparison. This proves that the spatial to 210% depending on the hydrological model
variability of the rainfall is correctly characterized selected and the size of the basin. For current rainfall
by the existing stations, at least for basins 1, 2 and events, the differences are clearly more signi®cant,
3. Consequently, this variability really does increase about 280 to 220%. These differences must be
with the size of the basin, and is the origin of differ- taken into account, because they can have a very
ences in both volumes and peak ¯ows as a function of strong impact on both calibration and extrapolation
either the UR and NUR patterns. of a rainfall-runoff model. Most of the time, calibra-
In addition, we studied the in¯uence of the location tion is indeed based on limited available information,
of the ®ctitious basins on the determination of the generally frequent rainfall events. Thus signi®cant
CVS. For basins 2, 3 and 4 (383, 103 and 22 km 2), errors may occur in the estimation of parameters
we regularly moved the center of the basins in the with respect to the selected pattern of spatial variabil-
study area to generate a representative sample of ity of rainfall; and consequently, the extrapolation of
100 basins of equal size. For each basin, it was the calibrated model may involve greater uncertain-
shown that the median CVS do not change Table 3), ties towards the rare events.
and that the intially selected basins are representative As an example, let us consider as `real' ¯ow values
of the spatial variability of rainfall. calculated for the largest basin (1531 km 2), using an
intermediate model (runoff production Model 3 with
STO ˆ 30 mm and V0 ˆ 1 m=s; K0 ˆ 0:7), an hourly
time step, and the NUR pattern. Next, let us calibrate the
same hydrological model to these ¯ow values, but using
the UR pattern. In order to be realistic, we also suppose
that only a limited number of observations are available
for calibration, for example those for which the return
period is less than 1 year. Under these conditions, the
calibration of the model using the UR pattern changed
the parameter values signi®cantly: among other possi-
bilities, we obtained good results with STO ˆ 10 mm;
V0 ˆ 1 m=s; K0 ˆ 0:7: As can be seen in Fig. 9, the
empirical distribution of the ¯ows obtained with UR and
Fig. 9. Empirical probability distributions of the peak ¯ows STO ˆ 10 mm coincides well with that of the ¯ows
modeled from the sample of 50 rain events, using runoff production obtained with a NUR and STO ˆ 30 mm in the range
Model 3 and the parameters linked to the type of rainfall. of the ¯ows for return periods of less than 1 year.
226 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Fig. 10. Relationship between the relative error and the spatial coef®cient of variation.

It will be noted that estimation of the STO para- generalize the degree of over-estimation, which
meter is greatly affected by the way the rainfall pattern depends to a large extent on both the hydrological
is represented. The ®rst problem consequently processes and characteristics of the region studied.
concerns the physical interpretation of the model
parameter (here STO, but it could be any other para- 3.4. In¯uence of the rain®elds characteristics
meter): how can STO be linked to physical indicators
if it is so sensitive to the rainfall pattern? The second As we have seen before, the distribution of both
problem is connected with extrapolation of the model: volumes and peak ¯ows using either UR and NUR
it should be noted that divergences from `real' ¯ows patterns is subject to variations that depend on the
appear in the case of higher return periods, and that magnitude of spatial variability and on the return
¯ows calculated with STO ˆ 10 mm and UR pattern period of the event. But up to now, these variations
are systematically over-estimated. This over-estima- have not been directly linked to any rain®elds char-
tion presents maximum relative errors of about 30% acteristics. This is the objective of this paragraph.
for the return periods ranging from 1 to 10 years, and For the largest basin (1), we compared the E 0i † with
decreases in the case of the highest return periods. the coef®cients of variation (CVi), given by Eq. (1).
From this example, we can conclude that calibrat- Comparison was performed using the intermediate
ing a rainfall-runoff model using the UR pattern will model: runoff production Model 3 with STO ˆ
have both a tendency to introduce a severe bias in the 30 mm; V0 ˆ 1 m=s and K0 ˆ 0:7 for the routing
interpretation of model parameters, and to over- model, and an hourly time step (Fig. 10a). The rela-
estimate rare frequency ¯ows. However, we cannot tionships between E 0i † and (CVi) are fairly good, a
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 227

Fig. 11. Relationship between event relative error (E 0 ), and respectively, CVS and PBAS, from observed and generated ®elds.

little better for volumes than for peak ¯ows. This can tant event of the sample (PBAS ˆ 120 mm; CV
be explained by the fact that peak ¯ows may be in¯u- observed ®eld ˆ 0.15, CV generated ®eld ˆ 0.67) is
enced by the position in time of the most intense storm not very sensitive to the different UR and NUR
(at the beginning or at the end of the event), whereas patterns (E 0 is very close to zero in both cases),
volumes are not. Furthermore, when considering all because the rain largely exceeds the threshold of
the basins on the same plot (Fig. 10b), it can be seen 30 mm for a large number of grid meshes. Thus, in
that the relationships do not change from a basin to spite of the CVS of the generated ®elds are much
another: this means that in a given model, the relation- higher than those of the observed ®elds, the main
ship is independent of the size of the basin. In the local factor that explains the relative error variability in
context, the coef®cient of variation can therefore be this case is no longer CV R2 ˆ 0:43†; but ln(PBAS)
considered as a strong index of error at the event R2 ˆ 0:76†: This example shows that the in¯uence of
scale: with respect to the R 2 coef®cient of determina- PBAS is predominant, and must be taken into account
tion of linear regressions, we found that the CVS of the in order to generalize the relationships between rela-
rain events explain 64% of the relative error (E 0 )s on tive errors and rainfall characteristics.
the peak ¯ow, and 92% of the relative error on runoff In order to get a more general relationship, we then
volumes, for calculated ¯ows using Model 3. carried out multiple regressions between (E 0 ), (PBAS)
It is also important to check if this kind of rela- and (CV). Finally, the same linear relationship between
tionship would change in the case of a different ln(2E 0 ), CV and ln(PBAS) was proved to be satisfac-
spatial structure of the rainfall ®elds, e.g. in other tory whether the ®elds are generated or observed:
climatic contexts. We consequently generated new
rain ®elds with the same spatial mean as the ln 2E 0 † ˆ 0:881 ln CV† 2 0:031PBAS 2 0:648 3†
observed ®eld, but with increased spatial variability.
From the (NUR) hyetographs of each event, the new Fitting of the relationship can be characterized through
rain ®elds were generated at each grid mesh by the differences between observed values (x) of the errors
multiplying rainfall with a randomized coef®cient ln(2E 0 ) and predicted values (y) by the model, these are:
that derived from the uniform distribution, varying
from 0 to 2. In this way we obtained very chaotic BV4 observed fields y ˆ 0:98x R2 ˆ 0:79
rain ®elds that may appear unrealistic, but are
convenient for our given purpose.
Such simulation was carried out on the smallest BV4 generated fieds y ˆ 0:95x R2 ˆ 0:71
basin (BV4). Runoff volumes were calculated using Applications to the observed events of the other basins
either UR or NUR rainfall patterns with the same show that Eq. (3) is still convenient, since the maximal
rainfall-runoff modeling conditions as earlier. As bias (in BV1) remains smaller than 16%:
can be seen in Fig. 11, we obtained a very different
relationship than the previous one. The most impor- BV3 observed fields y ˆ 1:00x R2 ˆ 0:85
228 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

BV2 observed fields y ˆ 1:05x R2 ˆ 0:95 conditions predominate. Thus any overestimation of
the errors due to the assumption of dry initial condi-
tions is probably slight.
BV1 observed fields y ˆ 1:16x R2 ˆ 0:95 The other simpli®ed assumptions we made (spatial
The relationship may therefore give a ®rst estimation of uniformity of the production model parameters, ®nal
the error associated with the runoff volumes, for a given in®ltration to zero, absence of percolation or evapo-
model. However, Eq. (3) must only be considered as transpiration at the event scale, linearity of the routing
indicative and a further perspective consists in checking model) on the other hand, tend to underestimate the
its validity in other contexts. relative errors. The in¯uence of these assumptions
The same procedure was applied to the peak ¯ows, requires further study. Nevertheless, it can be deduced
but no satisfactory relationship between E 0 , PBAS and that in the case of extreme events results will gener-
CV could be adjusted, probably because the position ally be little affected by these assumptions, and if this
of the most intense rainfall inside the whole event is the case, we can conclude that the UR rainfall varia-
must also be taken into account. bility pattern is acceptable for estimating the
frequency of extreme ¯oods. In such a case, design
rainfall for the estimation of extreme ¯ows could be
4. Discussion based on point rainfall distribution combined with the
areal reduction factor. In Mexico's climatic context,
It should be possible to extrapolate the results we this assumption seems to concern a rather large range
obtained to the areas with equivalent spatial and of both models and surface areas.
temporal rainfall characteristics. Both tropical and
equatorial mountainous sites may be considered.
Moreover, the relationship between the calculated 5. Conclusion
relative errors and the rainfall characteristics means
it is possible to envisage the results that would be Using a sample of 50 rain events from a dense rain
obtained in other conditions of rainfall spatial varia- gauge network around Mexico City, we studied the
bility. However, such relationship depends on both the in¯uence of the spatial variability of rainfall on the
selected runoff model and the value of its parameter. estimation of both peak ¯ows and runoff volumes.
A rainfall simulation model would be very useful to These were calculated from distributed hydrological
explore the sensitivity of the hydrological models models using different patterns of rainfall representa-
under a large array of rainfall, and this must now be tion as input data: spatially averaged UR, and NUR.
undertaken. Both calculated runoff volumes and ¯ows can differ
This study was based on some simpli®ed assump- considerably as a function of these two patterns,
tions. Although we neither consider the spatial varia- depending on the selected runoff production model,
bility of runoff production parameters, nor their the size of the catchment, and the frequency of the
temporal variability (for example related to initial event. Runoff production models based on storage
soil moisture conditions), it is possible to show what reservoirs are the most sensitive. Differences increase
effects would be caused by doing so. with the size of the catchment because there is also an
In case of the initial moisture conditions, which increase in rainfall variability of rainfall. In general,
were ®xed at zero at the beginning of each event, for any model and any basin area, the differences are
using a higher value would be equivalent to reducing less for extreme events than for current rainfall events.
the STO value for models 1 and 3. And, as was shown To sum up, relative errors vary from 230 to 210%
in Section 3.1.2, this would also reduce the relative for the former and from 280 to 220% for the latter.
errors caused by using the UR pattern instead of the There are two reasons for this: ®rstly, in the local
NUR one. However, in the tropical context, evapo- context the most signi®cant ¯ows are generated by
transpiration is generally very high between rainy widespread rainfalls, and secondly the effect of
events, and the intervals between events are usually storage associated with models 1 and 3 is strongly
long enough to ensure that low initial soil moisture mitigated for the highest rainfalls.
 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230 229

Two main conclusions can be drawn in the local Bouvier, Ch., Delclaux, F., 1996. ATHYS: a hydrological environ-
context of this study. First, given the moderate ment for spatial modelling and coupling with a GIS. Proceed-
ings of the HydroGIS 96, Vienna, Austria, pp. 19±28. IAHS
in¯uence of rainfall variability patterns on extreme
publication no 235.
events, in most cases design storms do not need to Bouvier, C., Fuentes, G., Dominguez, R., 1994. Mercedes: un
respect the whole rain ®eld, but can be derived from modeÁle hydrologique danalyse et de preÂvision de crues en milieu
statistics at a given point and combined with the heÂteÂrogeÁne. Crues et Inondations, 23 eÁmes journeÂes de lhydraulique,
probabilistic concept of the areal reduction factor. SHF, NõÃmes, 14±16 September 1994, pp. 257±260.
Bouvier, C., Ayabaca, E., Perrin, J.L., Cruz, F., Fourcade, B.,
Second, the calibration of a rainfall-runoff model
Rosario, S., Carrera, L., 1999. VariabiliteÂs temporelle et spatiale
can be severely affected when using a spatially des averses en milieu andin: exemple de la ville de Quito (Equa-
averaged rainfall instead of a rain ®eld. Generated teur). Revue de GeÂographie Alpine 3 (87), 51±66.
biases have consequences for the numerical stability Brunet-Moret, Y., Roche, M., 1966. Etude theÂorique et meÂthodolo-
of the model parameters from one event to another gique de labattement des pluies. Cahiers ORSTOM: SeÂrie
and consequently on the overall and physical interpre- Hydrologie (FRA) 3 (4), 3±13.
Chocat, B. 1997. EncyclopeÂdie de l'hydrologie urbaine et de l'as-
tation of these parameters. Furthermore, rare and sainissement, Eurydice92, Collection Technique and Documen-
extreme ¯ows are generally overestimated. In the tation, Lavoisier.
example that we presented, overestimation reached Cisneros, L., Bouvier, C., Dominguez, R., 1998. Aplicacion del
30% for return periods ranging between 2 and 5 metodo kriging en la construccion de campos de tormenta en
years, a little less for higher return periods. la Ciudad de Mexico. 18eÁ Congreso LatinoAmericano de
Hidraulica, Octubre 1998, Oaxaca, Mexico, pp. 379±388.
Although these differences did not reach excep-
Cisneros, L., Bouvier, C., Dominguez, R., 2000. Generacion de
tional values, attention must be paid to errors caused campos de precipitacion en el Valle de Mexico. 19eÁ Congreso
by not taking spatial rainfall organization into account LatinoAmericano de Hidraulica, Octubre 2000, pp. 249±258.
in the hydrological models. It should also be noted Creutin, D., Obled, C., Tourasse, P., 1980. Analyse spatiale et
that, in this study, some simpli®ed assumptions temporelle des eÂpisodes pluvieux ceÂvennols. La MeÂteÂorologie,
6 eÁme seÂrie 20±21, 233±242.
concerning the basins (spatial uniformity of the
Dawdy, D.R., Bergman, J.M., 1969. Effect of rainfall variability on
morphological and physiographical factors) and the stream¯ow simulation. Water Resour. Res. 5, 958±969.
models used (no continuous in®ltration, linear routing FaureÂs, J.M., Goodrich, D.C., Woolhiser, D.A., Sorooshian, S.,
model) may underestimate the in¯uence of the spatial 1995. Impact of small scale spatial rainfall variability on runoff
organization of rainfall. Consequently we recommend modeling. J. Hydrol. 173, 309±326.
that this kind of study be carried out for real catch- Goovaerts, P., 2000. Geostatistical approaches for incorporating
elevation into the spatial interpolation of rainfall. J. Hydrol.
ments (using other parameters or models), and also in 228 (1±2), 113±129.
other climatic regions. Gupta, V.K., Waymire, E.C., 1979. A stochastic kinematic study of
subsynoptic space-time rainfall. Water Resour. Res. 15, 637±
644.
Acknowledgements Krajewski, F.W., Lakshmi, V., Georgakakos, K.P., Jain, S.C., 1991.
A Monte Carlo study of rainfall sampling effect on a distributed
The authors thank Dr Thierry Lebel and Dr catchment model. Water Resour. Res. 27 (1), 119±128.
Kuczera, G., Williams, B.J., 1992. Effect of rainfall errors on accu-
Murugesu Sivapalan for providing helpful sugges-
racy of design ¯ood estimates. Water Resour. Res. 28 (4), 1145±
tions for improving the manuscript. This work has 1153.
been performed within a project developed by IRD Lebel, T., Laborde, J.P., 1988. A geostatistical approach for areal
(Institut de Recherche pour le DeÂveloppement, rainfall statistics assessment. Stochastic Hydrol. Hydraul. 2,
France) and UNAM (Universidad Nacional Auton- 245±261.
Maidment, D., 1993. Handbook of Hydrology. McGraw-Hill, New
oma de Mexico, Instituto de Ingenieria), and was
York.
funded by CONACYT in Mexico. Marshall, R.J., 1980. The estimation and distribution of storm
movement and storm structure, using a correlation analysis tech-
nique and rainfall-gauge data. J. Hydrol. 48, 19±39.
References Morin, J., Rosenfeld, D., Amitai, E., 1995. Radar rainfall ®eld
evaluation and possible use of its high temporal and spatial
Bell, T.L., 1987. A space-time stochastic model of rainfall for satel- resolution for hydrological purposes. J. Hydrol. 172, 275±
lite remote-sensing studies. J. Geophys. Res. 92 (8), 9631±9643. 292.
230 P. Arnaud et al. / Journal of Hydrology 260 (2002) 216±230

Niemczynowicz, J., 1987. Storm tracking using raingauge data. Surkan, A.J., 1974. Simulation of storm velocity effects on ¯ow
J. Hydrol. 93, 135±152. from distributed channel networks. Water Resour. Res. 10 (6),
Niemczynowicz, J., 1991. On the storm movement and its applica- 1149±1160.
tions. Atmos. Res. 27, 109±127. Troutman, B., 1983. Runoff predictions errors and bias in para-
Obled, Ch., Wendling, J., Beven, K., 1994. The sensibility of hydro- meters estimation induced by spatial variability of precipitation.
logical models to spatial rainfall patterns: an evaluation using Water Resour. Res. 19 (3), 791±810.
observed data. J. Hydrol. 159, 305±333. Waymire, E., Gupta, V.K., Rodriguez-Iturbe, I., 1984. A spectral
Rodriguez-Iturbe, I., Eagleson, P.S., 1987. Mathematical models of theory of rainfall intensity at meso-beta scale. Water Resour.
rainstorm events in space and time. Water Resour. Res. 23, Res. 20, 1453±1465.
181±190. Wilson, C.B., Valdes, J.B., Rodriguez-Iturbe, I., 1979. On the in¯u-
Rodriguez-Iturbe, I., Mejia, J.M., 1974. On the transformation of ence of the spatial distribution rainfall on storm runoff. Water
point rainfall to areal rainfall. Water Resour. Res. 10, 729±735. Resour. Res. 15 (2), 321±328.
Sivapalan, M., BloÈschl, G., 1998. Transformation of point rainfall to Wotling, G., Bouvier, Ch., Danloux, J., Fritsch, J.-M., 2000. Regio-
areal rainfall: intensity±duration±frequency curves. J. Hydrol. nalization of extreme precipitation distribution using the princi-
204 (1±4), 150±167. pal components of the topographical environment. J. Hydrol.
Sivapalan, M., Wood, E.F., 1987. A multidimensional model of 233, 86±101.
nonstationary space-time rainfall at the catchment scale. Yoo, C., Valdes, J.B., North, G.R., 1996. Stochastic modeling of
Water Resour. Res. 23, 1289±1299. multidimensional precipitation ®elds considering spectral
Smith, J.A., Karr, A.F., 1985. Parameter estimation for a model of structure. Water Resour. Res. 32 (7), 2175±2187.
space-time rainfall. Water Resour. Res. 21 (8), 1251±1257.

View publication stats