Journal of Hydrology 448–449 (2012) 39–59

Contents lists available at SciVerse ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Influence of mesh structure on 2D full shallow water equations and SCS

Curve Number simulation of rainfall/runoff events
Daniel Caviedes-Voullième ⇑, Pilar García-Navarro, Javier Murillo
Fluid Mechanics, University of Zaragoza, Zaragoza, Spain

a r t i c l e i n f o s u m m a r y

Article history: Hydrological simulation of rain-runoff processes is often performed with lumped models which rely on
Received 6 September 2011 calibration to generate storm hydrographs and study catchment response to rain. In this paper, a distrib-
Received in revised form 6 February 2012 uted, physically-based numerical model is used for runoff simulation in a mountain catchment. This
Accepted 1 April 2012
approach offers two advantages. The first is that by using shallow-water equations for runoff flow, there
Available online 16 April 2012
This manuscript was handled by
is less freedom to calibrate routing parameters (as compared to, for example, synthetic hydrograph meth-
Konstantine P. Georgakakos, Editor-in-Chief, ods). The second, is that spatial distributions of water depth and velocity can be obtained. Furthermore,
with the assistance of Ioannis K. Tsanis, interactions among the various hydrological processes can be modeled in a physically-based approach
Associate Editor which may depend on transient and spatially distributed factors. On the other hand, the undertaken
numerical approach relies on accurate terrain representation and mesh selection, which also affects
Keywords: significantly the computational cost of the simulations. Hence, we investigate the response of a gauged
Finite volumes catchment with this distributed approach. The methodology consists of analyzing the effects that the
Shallow-water equations mesh has on the simulations by using a range of meshes. Next, friction is applied to the model and the
Mesh generation response to variations and interaction with the mesh is studied. Finally, a first approach with the well-
SCS-CN known SCS Curve Number method is studied to evaluate its behavior when coupled with a shallow-water
model for runoff flow. The results show that mesh selection is of great importance, since it may affect the
results in a magnitude as large as physical factors, such as friction. Furthermore, results proved to be less
sensitive to roughness spatial distribution than to mesh properties. Finally, the results indicate that SCS-
CN may not be suitable for simulating hydrological processes together with a shallow-water model.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction
Runoff processes at the catchment scale have been modeled
extensively with empirical and lumped methods as to generate
the hydrograph response of storm events for various applications.
It is nevertheless possible to make use of two-dimensional shal-
low-water models to perform distributed and physically-based
runoff analysis at the catchment scale. This is not done in many in-
stances because of possibly large computational time as well as the
resolution of available spatial data and difficulties when it comes
to parameter calibration. Shallow water models allow to obtain
the same type of results as lumped models, namely runoff volume
and outflow hydrographs, but also spatial distribution of water as
well as runoff flow vectors. Hence, results are physically-based and
therefore arguably better, but more complete, in the sense that
more information is obtained.
In this paper, a 2D shallow-water model is used to simulate
the response of a Pyrenean catchment during a storm event.
⇑ Corresponding author. Address: Ed. Torres Quevedo, María de Luna 3, CP 50015,
Zaragoza, Spain. Tel.: +34 976761881; fax: +34 976761882.
E-mail address: daniel.caviedes@unizar.es (D. Caviedes-Voullième).
Experimental data for the catchment exist and are used to validate
numerical results. The objective is to understand the interactions
of various model parameters with the numerical model and com-
prehend the effect that such interactions have on runoff simula-
tion. In the matter of hydrologic and hydraulic parameters
subject to the choice of modelers, it is necessary to characterize
the response of the model to each parameter in order to provide
parameter sensitivity information and a grasp of the effects of
uncertainty of parameter selection. Manning’s roughness coeffi-
cient and Soil Conservation Service Curve Number (SCS-CN)
parameters are those of particular interest in this work.
The solution of the shallow-water equations is done by means
of a first order, explicit, upwind finite volume numerical scheme
(Murillo and Garcia-Navarro, 2010; Murillo et al., 2007). It is
widely known that finite volume schemes are heavily dependant
on mesh properties such as cell shape (triangular, rectangular,
etc.) and size as well as the general mesh structure (Leveque,
2002). It is quite clear that as the number of cells increases, CPU
time increases. This becomes an important issue since the very
complex topography of mountain catchments requires a large
number of cells. Furthermore, the explicit nature of the scheme im-
poses a time step constraint so as to satisfy the CFL condition

0022-1694/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jhydrol.2012.04.006
40 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
(Murillo et al., 2009). The time step is smaller as finer meshes are
used. From another perspective, the finest mesh resolution for a gi-
ven problem is given by the available topography data, and is sel-
dom chosen by the modeler. However, cell shape and mesh
structure are modeler choices. Hence, it appears important to ex-
ploit such control in order to obtain a mesh that will allow to ade-
quately represent the complex topography and the hydraulic
phenomena of interest at the same time that a minimal number
of cells is obtained, thus generating the largest possible cells and
only refining locally, where the topography demands it. In this pa-
per, the effects of mesh resolution, combined with square and tri-
angular cells are examined. This is further studied by using
uniformly refined triangular meshes versus terrain-adapted lo-
cally-refined triangular meshes.
The methodology for this work consists of separating the anal-
ysis for each of the mentioned issues (i.e., mesh effects, SCS-CN,
Manning’s n) with a spectrum of cases, as well as combining sce-
narios in which interactions between the parameters can be seen.
With such methodology it is possible to assess the magnitude of
the variations and to which parameter or combination of parame-
ters they can be attributed.
2. Catchment description
The Arnás catchment is located in the northern Spain Pyrenees,
in the Borau valley, and has a surface of 2.84 km2, ranging in alti-
tude from around 900 to 1340 m.a.s.l. The catchment has been
gauged since 1995 which allows a vast recollection of hydrological
data. The primary channel in the catchment is the Arnás ravine
which flows into the Lubierre river. Flow measurements have been
performed at the outlet on the Arnás ravine (see Fig. 1) with 5 min
frequency. Meteorological data have been collected by a meteoro-
logical station at the outlet location, and rainfall has been regis-
tered by a tipping bucket rain gauge with a frequency of 5 min.
Geologically, the catchment lies over Eocene flysch formations
and has suffered land use and coverage changes in recent decades,
generating a mixed vegetation cover which ranges from forest
patches, dense and open shrubs, grassland cover and bare land
(Garcia-Ruiz et al., 2005). Previous studies (Garcia-Ruiz et al.,
2005; Lana-Renault, 2007; Lana-Renault et al., 2007; López-Barre-
ra, 2011; Serrano-Pacheco, 2009) have provided information such
France
42°38'24" N
0°35'03" W
Spain
0 250 500 750
Fig. 1. Arnás catchment.
as geological description and vegetation maps which are the basis
for hydrological characterization of the catchment.
From the experimental perspective, the response of the Arnás
catchment to storms has been described as fast, and the shape of
the hydrograph has been found to reflect the precipitation struc-
ture (Seeger et al., 2004). From the analysis of experimental data
it has been concluded that the response to large storms shows sea-
sonal variation, with highest runoff generation in springtime, while
in autumn and summer runoff generation is strongly limited (Lana-
Renault et al., 2007). Some of the studied storms with moderate
precipitation (total precipitation larger than 10 mm) did not gener-
ate runoff and, in general, high storm-runoff variability is observed.
Furthermore, Lana-Renault et al. (2007) conclude that the rainfall
magnitude and the type of rain influence the response, while max-
imum rainfall intensity was found to correlate poorly. They also
conclude that it is likely that runoff is generated mainly by infiltra-
tion excess during the dry season (Hortonian flow), while satura-
tion excess appears to be the primary runoff generation process
during the wet season. Garcia-Ruiz et al. (2005) characterize spring
and autumn as wet season, winter as an intermediate period and
summer as the dry season and also argue that a large spatial vari-
ability of the runoff-contributing areas exists in the catchment.
Both studies (Garcia-Ruiz et al., 2005; Lana-Renault et al., 2007)
conclude that the water table depth has limited influence on catch-
ment response to storms.
From the numerical simulation point-of-view, López-Barrera et
al. (2011) simulated the Arnás catchment using a distributed ap-
proach, based on a 2D diffusive wave model for surface flow, and
both Horton and Green-Ampt models for infiltration, finding diffi-
culties in reproducing outflow hydrographs. López-Barrera (2011)
also used the SCS-CN method to estimate infiltration losses in
the Arnás catchment, not with a distributed modeling approach
(as the aforementioned 2D diffusive wave model), but with a
lumped approach, namely the SCS Unit Hydrograph method. Ser-
rano-Pacheco (2009) simulated the catchment with a distributed
approach, but using the complete 2D shallow-water equations
for surface flow and Horton’s model for infiltration. Both studies
conclude that infiltration plays a major role in the catchment re-
sponse and very large precipitation losses occur because of infiltra-
tion. Therefore, infiltration modeling plays a major role in the
numerical response.
1000
meters Rain gauge
Flowmeter
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59 41

3. Surface flow model classified and numerically dealt with as belonging to the family
of hyperbolic systems. The mathematical properties of (1) include
In this paper, runoff is simulated using the 2D shallow-water the existence of a Jacobian matrix, Jn , of the flux normal to a
equation as shown in Eq. (1). In this section the 2D formulation direction given by the unit vector n, E  n ¼ Fnx þ Gny , (Murillo
of the shallow-water equations is presented as depth averaged and Garcia-Navarro, 2010) defined as
equations expressing volume conservation and momentum
@E  n @F @G
conservation Jn ¼ ¼ nx þ ny ð7Þ
@U @U @U
@U @FðUÞ @GðUÞ
þ þ ¼SþH ð1Þ The presented equations are solved by an explicit, first order, up-
@t @x @y
wind, finite volume scheme described in Murillo and Garcia-Navar-
where ro (2010).

U ¼ ðh; qx ; qy ÞT ð2Þ
4. Simulation cases
are the conserved variables with h representing depth (m) and
qx ¼ hu and qy ¼ hv the unit discharges (m2/s), with u and v m/s A wide range of cases was proposed with the goal to observe the
the depth averaged components of the velocity vector u along the different results of simulations with varying degrees of detail and
x and y coordinates respectively. The fluxes are given by complexity. The only investigated event in this paper (referred to
 T !T as Event 1) occurred on 13/10/2005.
q2 1 2 qx qy qx qy q2y 1 2 Fig. 2 the simplest setup was to assume a no-friction, no-infil-
F ¼ qx ; x þ gh ; ; G¼ qy ; ; þ gh ð3Þ
h 2 h h h 2 tration case (F0-I0) – an ideal case of course – but which provides
interesting information. By contrasting results for such case with
where g (m/s2) is the acceleration of the gravity. The source terms of different meshes, the effect of mesh selection on the solution of
the system are split in two kind of terms. The term S is defined as the shallow-water equations for this particular problem can be
  found, mass conservation can be easily verified and a lower limit
sb;x sb;y T for time-to-peak can be obtained, as well as a maximum peak out-
S ¼ 0;  ; ð4Þ
q q flow. All meshes reported in Table 1 were used for this case.
with sb;x , sb;y (N/m2) the bed shear stress in the x and y direction The following series of simulations are those of friction effects.
respectively, with q (kg/m3) the density of the fluid. The friction A uniform average Manning roughness coefficient was used for the
losses are written in terms of the Manning’s roughness coefficient entire catchment obtained by averaging the roughness map re-
n (s m1/3): ported by Serrano-Pacheco (2009). This allows to observe the ef-
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 4=3 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 4=3 fects of friction in the same mesh, and the interrelated effects
Sfx ¼ n2 u u2 þ v 2 =h ; Sfy ¼ n2 v u2 þ v 2 =h ð5Þ between mesh shape and resolution and friction. The next step
was to simulate using the roughness map instead of the uniform
The term H is defined as
average value, to assess if spatial variability (both real and artificial
 T due to mesh properties) of the roughness coefficients was a signif-
@z @z
H ¼ 0; gh ; gh ð6Þ icant factor.
@x @y
The next set of simulations deals with infiltration, by means of
and expresses the variation of the pressure force along the bottom the SCS-CN method. Experimental data provide values of outlet
in the x and y direction respectively, formulated in terms of the bed runoff volume and effective precipitation. With these data, lumped
slopes of the bottom level z (m). estimates with SCS-CN can be made to calibrate CN and the initial
System (1) is time dependent, nonlinear, and contains source abstraction ratio a. With the lumped calibration as a starting point,
terms. Under the hypothesis of dominant advection it can be distributed SCS-CN simulations can be performed to obtain a new

0 0.200
Accumulated Precipitation (mm) in 5 minutes

0.175
0.5

0.150
1

0.125
Flow (m3/s)

1.5

0.100

2
0.075

2.5
0.050

3
0.025

3.5 0.000
0 50 100 150 200 250 300 350 400 450 500 550

Time (min)

Fig. 2. Event 1 (13/10/2005).

42 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

Table 1 for the entire catchment, as spatially distributed parameters would

Mesh properties. be impossible to calibrate and would only introduce more com-
Name Max cell area (m2) Number of cells Area factor plexity. Only a single locally refined mesh (LR-22624) was used
SS-5 25 110,010 1.000 for this procedure since it was found in the mesh analysis that it
TS-5 12.5 220,020 1.000 was accurate and efficient. Because of the wide variety of results,
SS-10 100 27,941 1.016 further testing was required than that initially planned. The addi-
TS-10 50 55,882 1.016 tional tests and the underlying reasoning is presented in Section
SS-15 225 12,478 1.0208
TS-15 112.5 24,956 1.0208
12.
SS-20 400 7083 1.0302
TS-20 200 14,166 1.0302
TU-10-40 40 106,939 1.000
5. Spatial discretization: mesh and parameters
TU-16-60 60 71,384 1.000
TU-18-100 100 43,019 1.000 The physically-based, distributed modeling approach proposed
TU-LR-16666 1640 16,666 1.000 for this catchment requires the use of computational meshes for
TU-LR-22624 1281 22,624 1.000 two reasons. First, a mesh is necessary to compute the solutions
TU-LR-31662 1247 31,662 1.000 of the finite volume approximation to the shallow-water equations
for surface flow. Second, a mesh allows to discretise the catchment
surface in order to represent the heterogeneity of the hydrological
calibration of CN and a. In all cases, since only one hydrograph is parameters of interest, such as infiltration and roughness/friction.
available, both CN and a are assumed as uniform average values The quality of finite volume approximations is known to have

Fig. 3. Local mesh refinement process.

 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59 43

7 8

6 7
Percentage of cells (%)

Percentage of cells (%)

6
5
5
4
4
3
3

2
2

1 1

0 0
10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700

Cell area (m2) Cell area (m2)

10 16

9
14
8
Percentage of cells (%)

Percentage of cells (%)

12
7
10
6

5 8

4
6
3
4
2
2
1

0 0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300

2)
Cell area (m Cell area (m2)

Fig. 4. Cell area histograms for selected meshes.

dependency on mesh selection (Leveque, 2002), in terms of both its Consider a hydrologic analysis point of view where runoff vol-
structure and resolution. Furthermore, it is expected that the afore- ume and hydrograph properties are chief interests (ignoring for
mentioned properties of the mesh should have an effect on the now the spatial distribution). Even while considering perfect repre-
solution because of its implications in the quality of the represen- sentation of infiltration and roughness parameters, mesh resolu-
tation of hydrological parameters. Hence, in order to examine the tion impacts the resulting outflow hydrograph and runoff
significance of mesh selection, several meshes were generated for volume. The latter may be affected when using coarse meshes be-
the catchment. cause high frequency variations in the topography may not be
High mesh resolution is, in terms of accuracy, always favorable. sampled. Depending on the averaging method to produce a repre-
However, in most cases mesh resolution is limited either by avail- sentative elevation of that cell from finer spatial data, local maxima
able spatial data or by the fact that it implies an important cost of or minima may be lost or generated (López-Barrera, 2011; Merw-
computational resources. Cell size and, in consequence, mesh res- ade, 2009), thus changing the available volume for surface ponding
olution, is not bounded by physical ranges (Hardy et al., 1999) and (surface storage). Similar effects may be observed when high reso-
hence is subject to the modeler choice and experience, using lution bathymetry data is not available for a river, although high
numerical convergence as a mesh quality index, and invoking the resolution topographic data may be available for the floodplain.
idea that if a finer mesh yields very similar results as a slightly The outflow hydrograph can also be affected by poor topography
coarser mesh, then the adequate resolution has been found. When representation, since volume may be artificially lost or generated
high resolution spatial data are available, by selecting an appropri- by such mechanisms. This is also true for physically-based, distrib-
ate cell size larger than the available data, the mesh resolution can uted models, and can be further evaluated by using the inundation
be appropriately selected to obtain satisfactory results without zone as a benchmarking result (Casas et al., 2006; Cook and Merw-
excessive computational cost. But other issues play a role in this ade, 2009; Hardy et al., 1999; Horritt et al., 2006). But furthermore,
selection. Complex topography such as mountainous terrains or these errors in topographic representation affect flow computa-
small details in large terrains such as levees, ditches, roads, rail- tions since terrain slopes may be poorly represented and cell shape
roads and buildings may require high resolution meshes to cor- and mesh structure may introduce numerical viscosity of – a priori
rectly represent topographic features and small scale processes – unknown magnitude. The effects of mesh resolution and struc-
which can sometimes impact large scale flow behavior (Horritt ture can be considered part of the larger problem of appropriate
and Bates, 2001; Murillo and Garcia-Navarro, 2004; Schubert representation of spatial data, which also includes issues such as
et al., 2008). uncertainty of elevation data (Wilson and Atkinson, 2005). Thus,
44 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

Fig. 5. Examples of mesh results (terrain elevation is exaggerated 1.5 times).

although it is not the intention of this paper, the conclusions which of the effects that uncertainty in elevation measurements or repre-
can be gathered from introducing variations in elevation represen- sentation may have on flow results. The methodology followed
tation because of mesh selection can also aid in the understanding herein generates meshes that in one way or another may represent
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
1.04
Precipitated volume
Total simulated volume
Area factor
1.03
Normalized Volume
1.02
1.01
1 not a hindrance for result quality, but the large number of triangles
0.99
0.98
SS
SS
SS
SS
TS
TS
TS
TS
TU
TU 0
LR 00
5
10
15
20
-2
5
10
15
20
16
18
26
-6
-1
24
Fig. 6. Comparison of volumes at the end of the simulation.
poorly the real topography (this is, that topography which has been
measured accurately). But the errors which are artificially gener-
ated are similar in nature to those generated by poorly measuring
the real terrain surface.
Roughness is generally represented by means of Manning’s
coefficient. Determination of the appropriate values may be done
through reference data and improved by calibration. Nevertheless,
the calibration process may be affected by other model parameters,
in particular mesh resolution and structure, and the resulting cal-
ibrated roughness values may be mesh-dependent, thus compen-
sating numerical errors introduced by the mesh (Hardy et al.,
1999; Horritt and Bates, 2001). Therefore, it appears necessary to
evaluate the effects of different spatial representations of rough-
ness since different permutations produce different runoff flow
properties (Hardy et al., 1999). For any mesh, the simplest proce-
dure is to assign a domain uniform average roughness value. On
the other hand, it is possible to represent roughness values for
smaller areas, down to cell size, thus representing each cell with
its own particular roughness. It is therefore interesting to observe
the response of different meshes with the same roughness repre-
sentation and vice versa since sensitivity to the roughness value
may change with mesh resolution (Casas et al., 2006), which is rea-
sonable since roughness parameters intend to represent energy
loss phenomena which can be affected by mesh resolution and
model scale (Hunter et al., 2007).
6. Mesh generation
This section is devoted to test the sensitivity of the simulation
to different mesh configurations. The base data for all meshes
was a raster DEM with 5 m resolution. From this, both square (S)
and triangular (T) meshes were generated. Triangular meshes were
further subdivided into structured (TS) and unstructured meshes
(TU), and the latter were subdivided into homogeneous (TU) or lo-
cally-refined meshes (TU-LR). SS mesh generation was always per-
formed so that complete pixels were included inside square cells
(cell edge length as a multiple of pixel edge length). Once created,
the cell elevation was averaged from raster data. TS meshes were
created in the same fashion, with further division of the square
cells into two right-angled triangles. TU meshes were generated
with Triangle (Shewchuk, 1996). The cell size was controlled by
constraining the maximum cell area and the minimum interior an-
gle. The name codes for the meshes are as follows. For both square
and triangular structured meshes the edge length is indicated.
Thus, TS5 is a structured triangular mesh with 5 m catheti. The
names of unstructured triangular meshes indicate the minimum
45
and maximum cell area, thus TU-10-40 is an unstructured triangu-
lar mesh with a minimum cell area of 10 m2 and maximum area of
40 m2. The minimum area is used as an indicator of the smallest
cell, which affects the CFL condition. Locally refined meshes are
named with their total cell number, e.g, TU-LR-16666 is a locally
refined mesh with 16,666 cells.
TU meshes are more or less homogeneous, in the sense that cell
area histograms show little dispersion as shown in Fig. 4a. This is
that it implies when cells are small, is a problem for computational
time. Hence a local mesh refinement algorithm was implemented,
based on the gradient of terrain slope. Note that the local refine-
ment procedure described herein is not (strictly) an adaptive mesh
refinement process since refinement criteria is based only on ter-
rain data, not flow computations. Hence, the term local refinement
(LR) is preferred. The LR algorithm is as follows, and is illustrated in
Fig. 3. The original DEM raster data – shown in Fig. 3a – with 5 m
resolution are processed to obtain a slope vector for each pixel as
shown in Fig. 3b. Then, the gradient of the slope vector field is com-
puted thus generating a new second rank tensor field with such
gradients. The magnitude of the vector field in each pixel is then
computed as in Fig. 3c, and categories are defined to assign a gra-
dient index to a particular gradient range so that the user may
choose to have as many categories as needed. Gradient indexes
are stored in a new raster map with resolution equal to that of
the DEM raster as in Fig. 3d. A triangulation is generated using
the domain boundary with Triangle using maximum area and min-
imum angle constrains. The next step is to map gradient indices
into the triangular mesh. This is done by identifying the maximum
gradient inside a triangle, an mapping onto that triangle the index
within a gradient range classification, as in Fig. 3e. When all trian-
gles have an index, new maximum area constraints are imposed on
the triangles. The constraints are related to the gradient classes
that were previously defined. This information is then passed to
Triangle and a triangulation refinement process is done which uses
the new area restrictions. Thus, a triangulation which has more tri-
angles in regions with large slope gradients is obtained. In Fig. 3f
the new triangulation is shown together with the index raster,
where it can be seen how the indices govern the triangle genera-
tion. If the mesh is still too coarse the algorithm can be repeated
until results are satisfactory. Finally, the elevation is assigned to
the LR triangulation as the arithmetic mean elevation of those pix-
els which are contained or intersected by each triangle, as shown
in Fig. 3g. It is possible that more complex interpolation techniques
(area weighted, inverse distance weighted, Krigin) would yield bet-
ter discrete terrain surfaces, but this has not been explored in this
work. According to Murillo and Garcia-Navarro (2004) the initial
mesh which is to be refined also has an impact on the results.
The sensitivity of the results to the first triangulation was not
investigated in this work either. The resulting LR meshes have a
much wider cell area histogram than non refined TU meshes as
shown in Fig. 4. Although there is a large proportion of cells in
the smaller range, note that the maximum cell area values are sig-
nificantly greater than the average cell area values.
Table 1 shows properties of all meshes. The area factor is a mea-
sure of the catchment area error generated by using coarser
meshes, which cannot represent exactly the same catchment
boundary as the mesh with best resolution. By coarsening the
mesh, the generated squares or triangles may exist in some part
outside the original boundary, hence increasing slightly the area
of the catchment. This is accounted for in order to assess the mag-
nitude of the distortion it introduces in the computed volumes.
In order to illustrate the differences between different meshes, a
snapshot of them is shown in Fig. 5, showing a 3D rendering of the
terrain surface scaled 1.5 times in the vertical coordinate to facili-
tate its understanding. The color scale represents the norm of slope
46 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000

Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Flow (m3/s)
Simulated outflow Simulated outflow
12.50 Outlet runoff volume 60000 12.50 Outlet runoff volume 60000
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000
Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Simulated outflow Flow (m3/s) Simulated outflow

12.50 Outlet runoff volume 60000 12.50 Outlet runoff volume 60000
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000
15.00 15.00
Flow (m3/s)

Flow (m3/s)
Volume (m3)

Volume (m3)
Simulated outflow Simulated outflow
12.50 Outlet runoff volume 60000 12.50 Outlet runoff volume 60000
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

Fig. 7. Hydrographs and volumes for case F0-I0-E1 with structured meshes.

vector of each cell. A good representation of the topography re-
quires that both elevation and slope are adequately modeled and
this figure compares both. The figure shows a view from the south,
near the outlet point. The view is in the upstream direction, i.e.,
northwest, as to show the valley, in particular the main channel
topography and the eastern hills with the highest slopes and high-
est slope gradients. Fig. 5a shows the SS5 mesh with elevation data
exactly as in the original raster elevation data. This is considered
the reference and the terrain surface which best approximates
the real terrain. Fig. 5b–d shows the spatial representation by
structured meshes. Not all are shown since TS meshes are just a tri-
angular partition of the square cells of SS meshes. Note that TS10
still represents fairly well the slopes in the main channel (blue1
zones with very little slope) as well as the very high slopes in the
hills. It still keeps, in particular, the less steep zone in the middle
of the first water divide to the right. TS15 starts to loose such reso-
lution, and SS20 averages the elevation so much that the slope
changes significantly in the main channel and the less steep area
of the first water divide is also lost. The next set of meshes are the
TU meshes shown in Fig. 5e–g. All of them appear to represent fairly
1
For interpretation of the references to color in Figs. 1–20, the reader is referred to
the web version of this article.
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
22.50
20.00
17.50
15.00
Flow (m3/s)
Simulated outflow
12.50 Outlet runoff volume
Surface storage volume
10.00 Total volume
Precipitated volume
7.50
5.00
2.50
0.00
0 5000 10000 15000 20000 25000 30000
Time (s)
well the terrain, although they introduce more variability in slope
than the original SS5 representation. Nevertheless, the shapes are
accurate, and the extremes of the slope range appear to be well rep-
resented. The last group is that of locally refined meshes. These are
shown in Fig. 5h–j. It is difficult to appreciate differences between
them but, compared to TU and SS meshes, it can be seen that the
quality of the representation is as good as that of TU and sometimes
better that the SS or TS representations. The most important differ-
ence, however, is that LR representations require significantly less
cells to achieve very similar topographic representation. Note that
in all LR representations the cells near water divides, at the bottom
of the valleys and in the main channel are smaller than those cells in
the hill slopes. In the hill slopes high slopes may be present but they
do not change significantly and hence less cells are required to rep-
resent the topography correctly.
7. Mesh effects: results and discussion
In order to isolate the mesh effect, this section presents the re-
sults for the catchment considered frictionless and impervious. All
the differences in outflow hydrographs are due to mesh resolution
and cell properties. From the total volume and precipitated volume
data, it is clear that mass conservation is achieved at catchment le-
vel at all times in all cases. However, precipitated volume is
slightly higher for the coarser SS and TS meshes, as can be seen
in Fig. 6. This is expected since the increase in volume is due to
the increase in area because of the approximation of the catchment
boundary with less points. Moreover, the increase in volume is
identical to the increase in area due to the aforementioned effect,
and is in the worst case, a 3% increase in volume. Note that it only
occurs in structured meshes.
A summary of the results is presented in Fig. 7 for structured
meshes and in Fig. 8 for unstructured meshes. Both figures present
outflow hydrographs for all meshes as well as the accumulated vol-
ume from precipitation in the entire catchment. They also show
the outlet runoff volume (total volume exiting the catchment at
the outlet), the surface storage volume instantly stored in the sur-
face (volume which may be static or flowing, but in any case, inside
the domain) and the sum of both which is the total volume which
is also compared against precipitation volume.
From Figs. 7 and 8 the first conclusion is that of excellent mass
conservation by comparing total volume and precipitated volume.
However the distribution of the total volume is not the same for all
meshes. In this analysis the SS5 mesh is taken as the reference
since it is the base, unprocessed topographic with highest resolu-
47
22.50
100000 100000
20.00
17.50
80000 80000
15.00
Flow (m3/s)
Volume (m3)
Volume (m3)
Simulated outflow
60000 12.50 Outlet runoff volume 60000
Surface storage volume
10.00 Total volume
Precipitated volume
40000 40000
7.50
5.00
20000 20000
2.50
0 0.00 0
35000 0 5000 10000 15000 20000 25000 30000 35000
Time (s)
Fig. 7 (continued)
tion. Although there is no experimental result for the no-friction,
no-infiltration case, the result for SS5 is considered as the reference
solution. The response with SS5 is very quick and replicates the
shape of the hyetograph, which is expected because of the friction-
less and impervious conditions. Note also that surface storage
reaches a maximum value around the same time as peak flow,
when a large portion of the precipitated volume is static or moving
on the surface. Surface storage reduces to a residual value which is
composed of flowing water and an important proportion of pond-
ed, static water in topographic depressions, i.e., local elevation
minima. As the SS meshes become coarser the hydrograph suffers
a diffusive effect which is similar to hydrograph damping, although
for SS20 some high frequency variations appear which may be
attributed to the coarser step-wise paths that water must travel
to reach the outlet. In terms of time-to-peak, coarser meshes result
in larger times. In summary the hydrograph losses somewhat tem-
poral resolution, it is dampened and suffers lag. Another significant
result is that surface storage increases with coarser meshes. This is
due to the sum of two factors. The first is poor topographic repre-
sentation, since local minima and maxima may be poorly repre-
sented with coarse meshes, which results in static, ponded water
which cannot flow further. The second is numerical viscosity ef-
fects (and numerical diffusion) that are generated by coarser
meshes. Similar results are obtained for the TS progression. A com-
parison of SS vs TS shows that SS hydrographs appears more damp-
ened than TS solutions. This is consequent with the fact that
rectangular meshes impose a strong directionality upon the flow,
and structured triangular meshes impose less directionality. Thus,
rectangular meshes are said to be more viscous than structured tri-
angular meshes.
TU meshes are somewhat more complex to analyze. The three
TU meshes show that small dispersion of triangle sizes result in a
viscous mesh, as in the case of TU-10-40. The hydrograph is
dampened although high frequencies are still observed. Surface
storage is high, in the same magnitude as SS20. As the dispersion
grows the hydrograph is less dampened, and even higher fre-
quencies appear. In TU-16-60 high frequency variation occur
around 10,000 s that are not present in such amplitude else-
where. In TU-18-100 peak flow shows high frequency variations
which generate a significantly larger maximum instantaneous
flow. Timewise, both in TU-10-40 and TU-16-60 the second peak
is higher, thus changing significantly from SS5. The effect is evi-
dent for TU-10-40. TU-LR meshes seem to respond much better
than TU meshes. Surface storage is improved significantly and is
in the magnitude of SS5 results. These meshes have a very large
cell size dispersion, which is consistent with the observation that
48 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000

Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Flow (m3/s)
Simulated outflow Simulated outflow
12.50 Outlet runoff volume 60000 12.50 Outlet runoff volume 60000
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000
Time (s) Time (s)

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000
Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

12.50
Simulated outflow 60000 Flow (m3/s) 12.50
Simulated outflow 60000
Outlet runoff volume Outlet runoff volume
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000
Time (s) Time (s)

22.50 22.50
100000 100000
20.00 20.00

17.50 17.50
80000 80000
15.00 15.00
Volume (m3)

Volume (m3)
Flow (m3/s)

Flow (m3/s)

Simulated outflow Simulated outflow

12.50 Outlet runoff volume 60000 12.50 Outlet runoff volume 60000
Surface storage volume Surface storage volume
10.00 Total volume 10.00 Total volume
Precipitated volume Precipitated volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000
Time (s) Time (s)

Fig. 8. Hydrographs and volumes for case F0-I0-E1 with unstructured meshes.

larger dispersion results in better surface storage volume results. 8. Friction

This is due to a better representation of the terrain which better
captures local minima and slopes. In terms of hydrographs, high
frequency variations appear although in general less than in TU
meshes. In terms of time and peak flow, the first peak is always
the highest flow. Peak flow is higher in TU-LR-16666 and TU-
LR-22624, although it is instantaneous and rapidly descends to
the solution obtained with SS5. Very little lag appears. It seems
that the TU-LR meshes introduce very little numerical viscosity
hence high frequencies are observable. In general, TU-LR meshes
perform much better than TU meshes and better than most SS
and TS meshes.
Friction plays an important role in simulating surface runoff of
shallow-water flows. Conceptually, higher friction for the same
precipitation event will result in slower runoff which results in lar-
ger times to peak outflows and less intense peak outflows. In this
work, friction is modeled by means of the Glaucker–Manning
equation, using Manning’s n to represent roughness. Because fric-
tion modeling can be inaccurate when very low depth exists
(Burguete et al., 2007), as in steep hill slopes or in flows early in
the rain event it is of interest to evaluate how significant are the
effects of friction in the catchment response.
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
Fig. 9. Manning roughness coefficient map.
A roughness map generated by Serrano-Pacheco (2009) has
been used, with values ranging from 0.01 s m1/3 to 0.07 s m1/3,
with a distribution strongly weighted around the mean value of
0.0579 s m1/3. As can be seen in Fig. 9 most values are in the green
values of the scale, around 0.05 s m1/3 and 0.06 s m1/3.
Since roughness coefficients are spatially distributed, their rep-
resentation is affected by the mesh properties in the same way that
elevation (and slope) are affected. In this work three meshes have
been used to investigate the effects of mesh properties on the
roughness coefficient representation and on the outflow hydro-
graph. Comparison is made between different meshes with the
same base roughness map averaged within the larger cells in a
fashion similar to the one described for elevation averages when
generating the mesh. No other interpolation methods were stud-
ied. Because mesh effects are also investigated separately, it is pos-
sible to look for variations due to friction only. Roughness
variations could be considered as a decision parameter to control
mesh generation as is the slope gradient, but no efforts were made
in this line in the present work.
9. Friction effects: results
Results for the simulations which do consider friction, but no
infiltration, are presented in this section. A first set of simulations
was performed with uniform, average roughness for the entire
catchment. Because of this, the effects of friction in this set of re-
sults is not dependent on friction discretization. A second set of
tests was performed using spatially distributed roughness coeffi-
cients, which introduces not only real spatial variability but also
a new level of mesh dependency. The results are presented in
Fig. 10 for structured meshes and 11 for unstructured meshes. In
these figures, both for mapped friction (FM) and average friction
(FA) cases, and also the frictionless case (F0).
In general, all hydrographs generated from simulations with
friction show a much smoother behavior than frictionless simula-
tions. This is, of course physical damping due to friction. Peak flows
are lower than peak flows in the frictionless case, although the dif-
ference is less noticeable as the meshes become coarser. This indi-
cates that some damping was already generated by coarsening the
mesh. However in SS20 and TS20 high frequency variations are still
evident, and much more significant than in the rest of SS and TS.
This seems to be due to inadequate topographic representation,
worsened by friction, resulting in high frequency flow pulses prob-
ably generated by high frequency dry–wet (or nearly dry to very
49
wet) transitions in the large, poorly averaged cells of SS20 and
TS20. It is likely that it is worse in TS20 because SS20 presents
more numerical viscosity, which dampens the oscillations. In all
cases, as expected, surface storage is instantly greater with friction
(both mapped or averaged, in fact practically equal among them)
than in the frictionless case. By the end of the simulation time, sur-
face storage is practically equal regardless of friction, which indi-
cates that at such time the surface water is moving very slow or
is static.
By comparing FA results against FM results, very little, if any,
differences are found. This is reasonable since roughness distribu-
tion is concentrated around the average roughness value of 0.0579
with most (around 70%) raster roughness values being 0.05 or 0.06.
In cases of wider distributions it would be reasonable to observe
greater differences between the average roughness and the distrib-
uted roughness cases. Oscillations appear in the hydrograph, with
very high frequency but very small amplitude with respect to local
hydrograph values, although it should be noted that oscillations
also appear in average friction results and even frictionless results.
There appears to be more differences in SS than in TS meshes, and
the amplitude of oscillations is higher with coarser meshes.
Unstructured meshes show less oscillations, except in the case of
TU-18-100, where the average case generates very large oscilla-
tions. It should be noted that this mesh also generates large oscil-
lations in the frictionless case. It seems that the mesh represents
poorly the terrain and thus affects outflow. LR meshes show quite
good behavior. LR hydrographs are much smoother, quite similar
to the SS5 and TS5 hydrographs. The oscillations seen in the fric-
tionless LR simulations are dampened significantly and a better re-
sponse is obtained.
In terms of time-to-peak, friction introduces a delay, as ex-
pected. It is interesting to note that peak flow in SS5-F0 case occurs
at 3536 s with a value of 18.79 m3/s, while SS20-F0 shows peak
flow at 4881 s with a value of 17.57 m3/s. SS5-FA shows peak flow
at 4864 s and a value of around 15.67 m3/s. Meanwhile, SS20 with
friction shows a peak flow of 17.00 m3/s at 5241 s. The noteworthy
observation here is that the SS5-FA case has a time-to-peak delay
of 1328 s with respect to the SS5-F0 (around 38% more time) while
the delay between SS5-F0 and SS20-F0 is of 1345 s, or a 38% in-
crease in time-to-peak. In other words, the delay introduced by
mesh coarsening from SS5 to SS20 is practically the same as the de-
lay introduced by friction in the SS5 mesh. Furthermore, the delay
between SS20-FA and SS5-FA is 377 s, much less than the 1345 s
delay between SS20-F0 and SS5-F0 or the 1328 s between SS5-FA
and SS5-F0. It is similar to the 360 s delay introduced by friction
with SS20. These results show that time-to-peak delays due to fric-
tion and delays due to mesh coarsening interact in a complex way
and do not respond linearly to mesh coarsening. TS meshes shows
similar response, although TS20 does not delay as much the (first)
peak flow relative to TS5.
TU meshes involve time-to-peak error evident in the fact that
the second peak becomes higher than the first except in TU-18-
100 which in turn has large oscillations. Friction smoothes the re-
sults but the second peak remains more important in TU-10-40
and TU-16-60. TU-18-100 responds with significant oscillations
when combined with average roughness coefficients, which are
probably due to a combination of poor topographic and roughness
representation near the outlet boundary condition. LR meshes are a
significant improvement over TU meshes. For example LR-22624-
F0 has a delay of 120 s with respect to SS5-F0. Considering friction,
LR-22624-FA has a delay of only 3 s with respect to SS5-FA. Recall
that the delay due to friction in SS5 is 1328 s. Is is evident that LR-
22624 generates very small errors in terms of time-to-peak. In the
cases of TS and SS meshes the delay due to coarsening was even
larger than the delay due to friction (more than twice with
SS20). In the case of LR-22624 it is not. In fact the coarsening delay
50 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

22.50 22.50 F0
F0 PF* = 20.38 m3/s
3
PF = 18.79 m /s 100000 TP* = 3521 s 100000
20.00 TP = 3536 s 20.00 F0
PF = 20.49 m3/s
17.50 TP = 5772 s
17.50 FM FM
PF = 15.73 m3/s 80000 PF = 15.64 m3/s
80000
TP = 4728 s TP = 4795 s

Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Flow (m3/s)
FA
FM outflow FM outflow
FA
PF = 15.67 m3/s F0 outflow PF = 15.63 m3/s
F0 outflow
12.50 TP = 4864 s FA outflow 60000 12.50 TP = 4744 s FA outflow 60000
FM Surface storage volume FM Surface storage volume
F0 surface storage volume F0 surface storage volume
10.00 FA surface storage volume
10.00 FA surface storage volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 22.50
F0 100000 F0 100000
3 3
20.00 PF = 17.81 m /s 20.00 PF = 18.30 m /s
TP = 3865 s TP = 3911 s
FM FM
17.50 PF = 16.28 m3/s 17.50
80000 PF = 15.89 m3/s 80000
TP = 4766 s TP = 4681 s
Volume (m3)

Volume (m3)
15.00 FA 15.00 FA
Flow (m3/s)

Flow (m3/s)
PF = 16.06 m3/s
FM outflow FM outflow
F0 outflow PF = 15.83 m3/s F0 outflow
12.50 TP = 5230 s 60000 12.50 TP = 4620 s 60000
FA outflow FA outflow
FM Surface storage volume FM Surface storage volume
10.00 F0 surface storage volume 10.00 F0 surface storage volume
FA surface storage volume FA surface storage volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 F0 22.50
PF = 17.51 m3/s
TP = 4500 s
100000 F0 100000
20.00 20.00 PF = 17.49 m3/s
FM TP = 4096 s
PF = 17.59 m3/s FM
17.50 TP = 4830 s 17.50
80000 PF = 15.90 m3/s 80000
FA TP = 5125 s
15.00 PF = 17.23 m3/s
15.00
Volume (m3)

Volume (m3)
Flow (m3/s)

Flow (m3/s)

FM outflow FA FM outflow
TP = 4844 s F0 outflow 3
PF = 15.87 m /s F0 outflow
12.50 FA outflow 60000 12.50 TP = 5285 s FA outflow 60000
FM Surface storage volume FM Surface storage volume
10.00 F0 surface storage volume 10.00 F0 surface storage volume
FA surface storage volume FA surface storage volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

Fig. 10. Hydrograph comparison of friction cases with structured meshes.

is just 9% of the friction delay of SS5. This shows that LR meshes
outperform structured mesh coarsening and homogeneous
unstructured meshes.
These observations lead to the conclusion that despite mesh
coarsening may be properly done, as the mesh is coarsened numer-
ical friction due to the larger cell size is introduced and this effect
must be considered. Roughness coefficients, though properly se-
lected according to standard charts may add an excessive amount
of friction to the model since the mesh may already generate sim-
ilar effects. It is clear then, that roughness coefficients must be cal-
ibrated with each mesh that is to be used in order to compensate
for numerical, mesh generated friction.
10. Computational efficiency
The wide range of simulations performed in this study are
aimed to understand the sensitivity of results to mesh selection
both by direct reasons as well as deeper interactions with other
phenomena, such as friction and its representation. However, com-
putational efficiency is also of interest and must be considered
when deciding on mesh selection. Following this, a summary of
the most important benchmarking variables is presented in
Fig. 12 together with CPU time for all the studied meshes. The re-
sults have been normalized by the corresponding SS5 results in
each case for comparison.
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59 51

22.50 FM
22.50
PF = 19.83 m3/s F0
100000 PF = 17.02 m3/s 100000
20.00 TP = 5309 s 20.00 TP = 4571 s
F0
PF = 17.57 m3/s FM
17.50 TP = 4881 s 17.50 PF = 16.55 m3/s
80000 TP = 5561 s 80000
FA

Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Flow (m3/s)
PF = 17.00 m3/s FA
3
TP = 5241 s FM outflow PF = 16.44 m /s FM outflow
F0 outflow TP = 5454 s F0 outflow
12.50 60000 12.50 60000
FA outflow FA outflow
FM Surface storage volume FM Surface storage volume
10.00 F0 surface storage volume 10.00 F0 surface storage volume
FA surface storage volume 40000 FA surface storage volume 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

Fig. 10 (continued)

Frictionless cases show only sensitivity to the mesh. It is clear
from Fig. 12a that time to peak and peak surface storage grow rap-
idly with coarser structured meshes. Peak outflow is mostly under-
estimated (compared to SS5) save for TS5, TU-18-100 and all LR
meshes, with LR-31662 practically reproducing the same peak
flow, though slightly later. Time to peak surface storage is either
well represented or underestimated by approximately 30%. With
these results most SS and TS meshes can be discarded for inaccu-
racy as well as TU-10-40 and TU-16-60. The best choices are TS5,
TU-18-100 and LR meshes. However, when also considering CPU
time, LR-16666 and LR-22624 appear as best choices since TS5
has a higher computational cost as does LR-31662. Within LR-
16666 and LR-22624, all variables except time to peak flow are bet-
ter represented by 22624 with very little additional computational
effort. Thus LR-22624 is considered the best choice.
Fig. 12b and c confirms the observations although errors are
dampened significantly because of friction. Surface storage errors
are still the most significant because they depend most on mesh
generation techniques. Both roughness representations confirm
that LR-22624 is the best choice for this terrain.
11. SCS Curve Number
From its origins, SCS-CN was developed for agricultural water-
sheds (Woodward et al., 2002). Since then, it has been applied to
a variety of watersheds and a wide range of engineering problems.
In engineering practice it is common to have very little, if any, data
for calibration of hydrologic models, in particular, infiltration data.
Complex, physically based solutions for infiltration can be formu-
lated by solving Richards’ equation. However, the required data
are complex and difficult to obtain, heterogeneity is difficult to
deal with, and computational time is significant. Simpler models
such as Horton and Green-Ampt models are much less demanding
in computational resources, and require significantly less field data
to obtain results. Even so, in many applications, Horton or Green-
Ampt are still far too demanding of field data. Because of this,
SCS-CN is still widely used in engineering practice with ungauged
catchments, and because of this widespread usage, exploring the
effects of such a simple and limited model is necessary.
The SCS-CN is formulated to describe runoff per unit area of
catchment, or effective precipitation Pe (mm) as
( tive precipitation is computed by subtracting the accumulated
ðPt Ia Þ2
for Pt > Ia
Pe ¼ P t Ia þS
0 for Pt 6 Ia
where Pt (mm) is total accumulated precipitation, Ia (mm) is the ini-
tial abstraction which infiltrates before runoff begins and S (mm) is
the potential maximum retention, estimated by means of the Curve
Number following
25; 400
S¼  254 ð9Þ
CN
The initial abstraction and potential storage are related by
Ia ¼ aS ð10Þ
Inserting (10) in (8) yields
ðPt  aSÞ2
Pe ¼ ð11Þ
P t þ Sð1  aÞ
where a is a coefficient sometimes called initial abstraction ratio and
frequently assumed as 0.2 from the original SCS-CN method (Ponce
and Hawkins, 1996; SCS, 1964). Nevertheless, various studies have
suggested that a may be in a different range, rarely satisfying
a = 0.2 (Beck et al., 2009; Mishra and Singh, 1999; Mishra et al.,
2003) with particular results as low as a = 0.004 in the study of a
Greek catchment (Baltas et al., 2007) and between 0.01 and 0.154
in the Three Gorges area in China with an optimal value of 0.05
(Shi et al., 2009).
It should be noted that the SCS-CN method was not designed to
consider time (Woodward et al., 2002). It was thought as to com-
pute the total runoff volume from a watershed basin and this
may be a reason of why it does not adjust easily to particular tem-
poral rain patterns, or to catchments which have a significant evo-
lution of runoff in time. Furthermore, it has been reported that CN
appears to work better in agricultural watersheds, followed by
range lands and worst results are obtained in forested watersheds
(Woodward et al., 2002). These considerations, together with the
empirical evidence that a  0.2 allow to question the validity of
the method. The intention in this paper, is to calibrate the SCS-
CN method with two parameters, CN and a in order to simulate
runoff volume correctly.
However, the chosen methodology is not a lumped method, but
a distributed and time-advancing method. Hence, the SCS-CN
method is not applied once to the entire catchment. Effective pre-
cipitation is computed with the SCS-CN for every cell in every time
step, using the accumulated precipitation Pt since the beginning of
the storm to that particular time. Therefore, an accumulated effec-
tive precipitation is obtained for that time. The increment in effec-
effective precipitation of the previous time step. Further time ef-
ð8Þ
fects can be noticed when considering that all precipitation infil-
trates while the accumulated rainfall Pt(t) is less than the initial
abstraction Ia. Ia is dependent on both CN and a values. As both val-
ues grow total runoff volume is reduced, but also the time after the
52 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

22.50 22.50
100000 100000
20.00 20.00 F0
F0 PF = 17.38 m3/s
PF = 16.33 m3/s TP = 5836 s
17.50 TP = 5807 s 17.50
80000 FM 80000
FM PF = 14.99 m3/s
15.00 PF = 14.33 m3/s 15.00

Volume (m3)

Volume (m3)
TP = 6107 s
Flow (m3/s)

Flow (m3/s)
TP = 6020 s FM outflow
FA F0 outflow
12.50 FA FM outflow 60000 12.50 PF = 15.09 m3/s 60000
FA outflow
PF = 14.31 m3/s F0 outflow TP = 5836 s
TP = 5969 s FA outflow FM Surface storage volume
10.00 FM Surface storage volume 10.00 F0 surface storage volume
F0 surface storage volume FA surface storage volume
FA surface storage volume 40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 F0 F0
PF = 22.17 m3/s 22.50 3
PF = 23.96 m /s
100000 100000
20.00 TP = 3917 s TP = 3605 s
FM outflow 20.00
F0 outflow
17.50 FM FA outflow FM
PF = 15.36 m3/s 80000 17.50 80000
FM Surface storage volume PF = 15.63 m3/s
TP = 4742 s
Volume (m3)

Volume (m3)
15.00 F0 surface storage volume TP = 4800 s
Flow (m3/s)

Flow (m3/s)
FA FA surface storage volume 15.00 FM outflow
FA
12.50 PF = 15.39 m3/s 60000 PF = 15.59 m3/s
F0 outflow
60000
TP = 4939 s 12.50 TP = 4869 s FA outflow
FM Surface storage volume
10.00 10.00 F0 surface storage volume
FA surface storage volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

22.50 22.50 F0
F0
PF = 22.69 m3/s 100000 PF = 19.06 m3/s 100000
20.00 TP = 3656 s 20.00 TP = 3691 s

17.50 FM 17.50 FM
PF = 15.72 m3/s 80000 PF = 15.66 m3/s 80000
TP = 4996 s TP = 4859 s
Volume (m3)

Volume (m3)
15.00 15.00
Flow (m3/s)

Flow (m3/s)

FA FM outflow FA FM outflow
PF = 15.52 m3/s F0 outflow PF = 15.61 m3/s F0 outflow
12.50 TP = 4871 s FA outflow 60000 12.50 TP = 4769 s FA outflow
60000
FM Surface storage volume FM Surface storage volume
10.00 F0 surface storage volume 10.00 F0 surface storage volume
FA surface storage volume FA surface storage volume
40000 40000
7.50 7.50

5.00 5.00
20000 20000
2.50 2.50

0.00 0 0.00 0
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

Fig. 11. Hydrograph comparison of friction cases with unstructured meshes.

beginning of the storm when runoff begins to appear is longer.
Hence, the outflow hydrograph can be affected in a complex way
by these parameters. In fact, it is possible that because of the tem-
poral effects of CN and a, the numerical response that could be
attributed to friction (both physical and numerical) is in part gen-
erated by the infiltration method. This is why it is interesting to
evaluate the response to SCS-CN parameters independently from
friction parameters. To address these issues, lumped estimates
for the catchment are performed using runoff volume as the target
variable, with both CN and a as calibration parameters in order to
compare calibration results with those obtained from the distrib-
uted simulations.
12. Curve Number effects: results and discussion
Traditionally, the SCS-CN method is calibrated by setting a = 0.2
and, with an initial guess based on land cover and use, performing
multiple simulations so as to obtain the Curve Number which best
satisfies the runoff volume. However, because of the considerations
mentioned in Section 11, a was also set as a calibration parameter.
To establish the reference values, the experimental runoff vol-
ume and the runoff coefficient were computed and are presented
in Table 2. Total precipitation was 36 mm for the entire storm
event, which results in a total volume of 99065.5 m3. Table 2 also
presents the results for eight lumped estimates using the SCS-CN
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59 53

2.5 100 Table 2

Time to peak outflow
Peak outflow Volumes for event E1.
Time to peak surface storage
Peak surface storage Case a CN Eff. precip. Runoff vol. Runoff
CPU time (mm) (m3) coef.

Normalized CPU time

2 10
Normalized Data

Experimental – – 0.579 1595.04 0.0161

Lumped 0.2 65.4 0.580 1596.4 0.0161
1.5 1 0.1 51.5 0.580 1597.9 0.0161
0.05 38.4 0.577 1588.4 0.0160
0.01 19.1 0.578 1592.3 0.0161
0.005 15.2 0.578 1591.1 0.0161
1 0.1 0.001 11.4 0.576 1586.6 0.0160
0.0005 10.9 0.579 1593.2 0.0161
0.0000 10.4 0.583 1603.4 0.0162

0.5 0.01
SS

SS

SS

SS

TS

TS

TS

TS

TU

TU -40

TU -60

TU -10

TU R-

TU R-
5

10

15

20
5

10

15

20

-1

-1

-1

-L 0

-L 166

-L 226
0

M
calibrating both parameters, and it can be only thought of as a cal-

R 24
-3
16
ibration parameter with some sort of meaning which relates to
66

62
permeability.
It is clear that by calibrating a some time distribution effect
2 10
Time to peak outflow should be reflected on the effective hyetograph. In order to observe
Peak outflow
Time to peak surface storage such differences, the SCS-CN was applied in a time-distributed
Peak surface storage
CPU time fashion, though still spatially-lumped. This is, computing effective
Normalized CPU time

precipitation at each time of the total precipitation hyetograph. Re-

Normalized Data

1.5 1
sults are shown in Fig. 13.
Note that, although all combinations of CN and a generate prac-
tically the same effective rainfall, the hyetographs can be drasti-
cally different. The two extreme cases are a = 0.2 and a = 0. Note
1 0.1 that the hyetograph shows non-zero values around 6000 s for
a = 0.2, while for a = 0 around 1000 s there exists effective precip-
itation already. All cases show similar tendency. Note that as
a ! 0, the effective hyetograph resembles the total precipitation
0.5 0.01 hyetograph, though in significantly different magnitude. Further-
more, it can be seen that hyetograph distribution is less sensitive
SS

SS

SS

SS

TS

TS

TS

TS

TU

TU -40

TU -60

TU -10

TU R-

TU R-
5

10

15

20
5

10

15

20

-1

-1

-1

-L 0

-L 166

-L 226

to changes in a as a approaches zero. The conclusion that can be

0

M
R 24
-3

drawn from this results is that, although any number of combina-

16
66

62

tions might be found for a and CN that satisfy runoff volume, the
choice may affect significantly the time distribution of effective
2 10 precipitation, and hence, of runoff generation. High values of a
Time to peak outflow
Peak outflow indicate a large proportion of losses in the beginning of the storm,
Time to peak surface storage
Peak surface storage
and less incremental losses throughout the storm, while low values
CPU time of a indicate that infiltration does not occur when rainfall begins,
Normalized CPU time
Normalized Data

but that it is a slow process. Hence, a has the capability of signifi-

1.5 1
cantly deforming the hyetograph.
The second level of analysis is that of the distributed model.
Simulations were performed with mesh LR-22624, infiltration
modeled by SCS-CN but without friction to observe clearly the ef-
1 0.1
fects of SCS-CN. Recall that the calibrated values obtained for
lumped estimates only ensure accurate runoff volume, not runoff
hydrographs nor peak flow, since SCS-CN cannot account for catch-
ment morphology and cannot transform rain into flows. In this
0.5 0.01 study, the shallow-water model is intended to account for catch-
SS

SS

SS

SS

TS

TS

TS

TS

TU

TU -40

TU -60

TU -10

TU R-

TU R-

ment morphology in a physically-based way. Hence a number of

5

10

15

20
5

10

15

20

-1

-1

-1

-L 0

-L 166

-L 226
0

simulations were performed as to calibrate both a and CN not only

R 24
-3
16

to obtain accurate runoff volumes, but also accurate runoff hydro-

66

62

graphs. This demands quite a lot more from the SCS-CN model as
will be shown.
Fig. 12. Comparison of benchmarking variables and CPU time for all meshes. The results from 40 different a and CN configurations did not
provide satisfactory calibration of the parameters. Neither volume,
method. The first simulation uses SCS-CN with the empirical value hydrograph peak flow or time to peak were adequately reproduced
of a = 0.2. This allows indeed for a calibration of CN which repre- by the simulations. The configurations are shown in Fig. 14. The
sents accurately the losses for event E1. However, the other seven very wide range of configurations is due to the fact that no ade-
simulations use arbitrary values of a ranging between 0.0 and 0.1 quate results were seen with the values selected based on lumped
and calibrating CN to achieve the same runoff coefficient. Note that estimates. The two best results in terms of outflow volume were
calibration is still possible and accurate values of runoff volume are obtained with CN = 52, a = 0.08 and CN = 30, a = 0.01. Note that
obtained in all cases, even with a = 0 which implies Ia = 0. Of the first case (CN = 52) is similar, though not exactly the same
course, the physical meaning of CN becomes quite diffuse when as one combination presented in the lumped estimates, though
54 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

-5
4.5⋅10
-5 CN 65.4, α=0.2 CN 38.4, α=0.05
4.0⋅10 CN 51.5, α=0.1 CN 19.1, α=0.01

Effective precipitation (mm)

-5
3.5⋅10
-5
3.0⋅10

2.5⋅10-5

2.0⋅10-5
-5
1.5⋅10
-5
1.0⋅10
-6
5.0⋅10
0
0.0⋅10
-5
4.5⋅10
CN 15.2, α=0.005 CN 10.9, α=0.0005
Effective precipitation (mm)

-5
4.0⋅10 CN 11.3, α=0.001 CN 10.4, α=0.0000
-5
3.5⋅10
-5
3.0⋅10
-5
2.5⋅10
-5
2.0⋅10
-5
1.5⋅10

1.0⋅10-5
-6
5.0⋅10

0.0⋅100
0 5000 10000 15000 20000 25000 30000 35000 0 5000 10000 15000 20000 25000 30000 35000

Time (s) Time (s)

Fig. 13. Effective hyetographs from lumped SCS-CN.

80
smaller than in the previous case, around 2500 s, but still the sim-
70 ulated hydrograph is severely different from the experimental
results.
60
From all the simulations shown in Fig. 14, a mean value of
50 surface storage of 1657.9 m3 at the end of the simulation was
obtained, and the values ranged from zero to 2002.0 m3. For the
CN

40
30
20
10
0
0 0.01 0.02 0.03 0.04
Fig. 14. CN–a combinations for target Vout = 1595 m3.
generating more effective precipitation (because of larger CN and
lower a) than the second case of lumped estimates. The second
configuration (CN = 30) has no similar in the lumped estimates.
For the CN = 52 case, both hydrograph and volume comparisons
are shown in Fig. 15. Note that the simulated outflow volume is
around 80% of the experimental outflow volume. However, the
simulated peak flow is around 30% greater than the measured peak
flow. The rising limb of the hydrograph is severely distorted and
the simulated time to peak is dramatically greater. There is a lag
of the order of 7500 s between the simulated and experimental
peak flows. A great deal of the delay is due to a as can be seen in
Fig. 13 where a value of CN = 51.5 and a = 0.1 was used, similar
to the case currently discussed.
Analogous results are presented in Fig. 16. In this case the sim-
ulated outflow volume is around 30% greater than the experimen-
tal volume, and the peak outflow is 130% greater than the
experimental measure. The time lag between peak flows is
LR-22624 mesh without any infiltration or friction, the surface
storage volume at the end of the simulation was 2026.8 m3.
Although this volume might contain moving water, an important
part of it is water ponded because of local topographic minima.
This information is significant because experimental runoff
volume, as shown in Table 2, is in the order of magnitude of the
0.05 0.06 0.07 0.08 0.09 0.1 final surface storage volume, and indeed quite close. The problem
α is that ponded volume is constituted of runoff volume from some
other area of the catchment, and SCS-CN is blind to such effects.
Hence, calibration of a and CN needs to account for the fact
that the volume that is produced as runoff may not show as
outflow volume at the outlet. In other words, some distortion of
runoff volume – as understood by SCS-CN – is introduced by the
topography. This can lead to the conclusion that with an effective
precipitation small enough to generate a runoff volume smaller
than the surface storage volume, no runoff will be observed at
the outlet. This is not necessarily true, as some part of such effec-
tive precipitation falls on the surface downstream of the largest
topographic local minima, as in the runoff path that such rain
describes, there may be no significant storage zones.
To test this hypothesis a synthetic event consisting of an instan-
taneous pulse of precipitation was generated. The rainfall pulse oc-
curs in the first instant of the simulation. The volume of the pulse
was set equal to the experimental runoff volume. Infiltration was
not simulated at all, setting an impervious catchment with no fric-
tion. After 32,100 s the surface storage volume was 1590 m3, and
the runoff volume at the outlet was 2.35 m3. Results are shown
in Fig. 17. This synthetic rain case allows to see that, even when
the correct effective precipitation volume is imposed, the runoff
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59 55

0.20 1600
Simulated outflow
0.18 Experimental outflow
Simulated outlet volume 1400
Experimental outlet volume
0.16
1200
0.14

1000

Volume (m3)
Flow (m3/s) 0.12

0.10 800

0.08
600

0.06
400
0.04

200
0.02

0.00 0
0 5000 10000 15000 20000 25000 30000 35000
Time (s)

Fig. 15. Results for CN = 52, a = 0.08.

0.40 2200

2000
0.35
1800

0.30
1600

0.25 1400

Volume (m3)
Flow (m3/s)

1200
0.20
Simulated outflow 1000
Experimental outflow
Simulated outlet volume
0.15 Experimental outlet volume 800

600
0.10

400
0.05
200

0.00 0
0 5000 10000 15000 20000 25000 30000 35000
Time (s)

Fig. 16. Results for CN = 30, a = 0.01.

volume is significantly less than expected from experimental data
or SCS-CN lumped estimates calibrated directly from experimental
data. It is clear that, in order to calibrate SCS-CN properly for its use
with the shallow-water simulator, the resulting effective precipita-
tion must be greater than the one computed from experimental
data in order to account for ponded water. Clearly, this numerical
issue can have a wide range of magnitudes, depending both on
infiltration, rain volume and distribution as well as topography.
Hence, how much effective precipitation should be obtained from
SCS-CN is difficult to know without actually simulating with the
shallow-water model at least to know how much runoff may be
lost to surface storage. Given that surface storage for this catch-
ment is practically equal to the experimental runoff volume, the
calibration of SCS-CN should produce an effective precipitation of
around twice that volume, this is approximately 3200 m3.
To support this conclusion, an additional synthetic pulse event
was simulated, also instantaneous, but designed to generate
3200 m3. Surface volume results are shown in Fig. 18, where it
can be seen that, once the rain pulse flows out of the catchment,
the instantaneous surface storage decreases rapidly to an asymp-
totic value of around 1800 m3, 200 m3 more than expected from
the previous discussion. On the other hand, although Fig. 18 shows
that the simulated and experimental hydrographs are quite differ-
ent – which is reasonable since the hyetograph is drastically differ-
ent –, the simulated outflow volume is around 165 m3 less than the
experimental value, which represents an error of around 10% less
volume. Both results are consistent and support the conclusions
of the previous discussion that the SCS-CN effective precipitation
volume must also account for ponded volume which cannot be ob-
served in the outflow hydrograph.
56 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59

1594.0
Surface volume Because the synthetic pulse confirmed the hypothesis, addi-
1593.5
Total volume tional tests were performed to generate a rain of approximately
3200 m3, using the hyetograph for event E1 and calibrating CN
1593.0 and a for such volume.
A configuration with CN = 20.1 and a = 0 which generates a vol-
Volume (m3)

1592.5
ume in the aforementioned range was used to test the hypothesis
1592.0 together with the hyetograph. The results are shown in Fig. 19. In
terms of volume, the results are fairly satisfactory. The simulated
1591.5 volume is 3.5% smaller than the measured experimental volume.
1591.0
The simulated hydrograph still produces a larger peak flow which
is almost twice the experimental peak flow and has a delay in the
1590.5 order of 2000 s. The hydrograph is not simulated accurately.
Although a is zero, a large delay still exists which appears to be
1590.0
0 5000 10000 15000 20000 25000 30000 35000 unsolvable with the chosen methodology.
Time (s) Another result worth of mention is the one obtained with
CN = 44.4 and a = 0.05, and shown in Fig. 20. In this case the sim-
Fig. 17. Volume results for synthetic rain pulse simulation. ulated volume is around 15% less than the experimental volume.
The simulated peak flow is around 1.5 times the experimental peak

1.40 3200
Simulated outflow
Experimental outflow 3000
Simulated outlet volume
Experimental outlet volume 2800
1.20
Surface volume
2600
2400
1.00
2200
2000

Volume (m3)
Flow (m3/s)

0.80 1800
1600

0.60 1400
1200
1000
0.40
800
600
0.20
400
200
0.00 0
0 5000 10000 15000 20000 25000 30000 35000
Time (s)

Fig. 18. Results for 3200 m3 synthetic rain pulse simulation.

0.30 1600

1400
0.25

1200

0.20
Simulated outflow 1000
Volume (m3)
Flow (m3/s)

Experimental outflow
Simulated outlet volume
Experimental outlet volume
0.15 800

600
0.10

400

0.05
200

0.00 0
0 5000 10000 15000 20000 25000 30000 35000
Time (s)

Fig. 19. Results for CN = 20.1, a = 0.

 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
0.25
Simulated outflow
Experimental outflow
Simulated outlet volume
Experimental outlet volume
0.20
0.15
Flow (m3/s)
0.10
0.05
0.00
0 5000
flow and a delay in the order of 2500 s. Further simulations with
larger values of a resulted in larger errors both in volume and hyd-
rograph results. For CN = 68.3 and a = 0.2 the outflow volume was
2.57 m3 and the peak outflow was in the order of 0.0005 m3/s
although the expected volume computed from lumped estimates
was 3258 m3 and the simulated surface storage volume at the
end of the simulation was 1899 m3.
The evidence gathered from the wide range of simulations per-
formed leads to several conclusions concerning the use of SCS-CN
together with distributed surface flow modeling by means of shal-
low-water simulation. Firstly, SCS-CN method may generate an
effective hyetograph which severely differs from the registered
hyetograph. The effect observed in simulated runoff is of signifi-
cant delays in peak flow as well as hydrographs which do not
resemble the experimental hydrographs. Second, calibration of
SCS-CN must include not only CN but also a since the effective hye-
tograph may be dramatically different with different values of a.
The third conclusion is that SCS-CN calibration cannot be per-
formed only based on the outled hydrograph and the outflow vol-
ume obtained from such hydrograph. Surface storage of ponded
water must be considered as well. Hence, SCS-CN must be cali-
brated to generate an effective precipitation volume which is the
sum of the outlet volume and the ponded water in the catchment.
This volumes of surface storage may be previously obtained by
simulating an impervious catchment and observing the volume
that remains ponded in the catchment after the storm. In lumped
methods, which make use of synthetic unit hydrographs, this effect
may be masked by calibration of the unit hydrograph parameters
but, in distributed, physically-based models, no calibration free-
dom is allowed, since topography and surface flow are modeled
physically and explicitly. Furthermore, in this particular case there
appears to be no possible calibration with SCS-CN that leads to an
adequate hydrograph or peak flow. It was only possible to calibrate
runoff volume. It appears that the use of shallow-water solvers for
rain-runoff transformation requires a better simulation of tempo-
ral distribution of infiltration and consequently effective
precipitation.
13. Conclusions
This study was intended to understand the possible effects that
different model parameter can have on simulation results. The
57
1600
1400
1200
1000
Volume (m3)
800
600
400
200
0
10000 15000 20000 25000 30000 35000
Time (s)
Fig. 20. Results for CN = 44.4, a = 0.05.
chosen methodology was simply to perform a range of simulations
with changing parameters, without the presence of other parame-
ters allowing to value them in their own right as well as to observe
their interactions.
The results show that mesh properties are one of the most sig-
nificant factors. Square and right-angled triangular structured
meshes with different resolutions were tested, as well as homoge-
nous unstructured triangular meshes and locally refined unstruc-
tured triangular meshes. Resolution and cell size are of great
importance. Square structured meshes are the most numerically
viscous and show flow directionality aligned with the principal
direction of the mesh. Triangular structured meshes lessen these
effects. However, high resolution structured meshes require large
computational efforts, and coarsening of structured meshes results
in a great loss of topographic representation quality, and because
of this, a significant loss in flow result quality. Unstructured trian-
gular meshes respond in complex ways, with mixed results, but
computational time can be similar to that of structured meshes.
Locally refined meshes by an automatic algorithm which adds cells
to represent correctly changes in terrain slope (slope gradient cri-
teria) outperform most other meshes. By adding more and smaller
cells were terrain complexity is high (high slope gradient) an
accurate representation of the terrain surface is achieved, while
keeping a low cell count, and thus, reducing the required computa-
tional effort. High precision is also achieved with locally refined
meshes.
Friction representation by means of Manning’s roughness is
perhaps the most commonly used representation. In this study,
the effects of friction on catchment flow is studied by comparing
frictionless results against results with friction. Furthermore, two
roughness representations were studied: a roughness map based
on soil and vegetation properties and an equivalent average rough-
ness value for the entire catchment. Both representation yielded
practically identical results since the distribution of roughness val-
ues is heavily weighted near the average value. High frequency
oscillations were observed when using the roughness map, but
their amplitude was small enough to be negligible. When compar-
ing frictionless results to results with friction the expected behav-
ior is seen. Friction dampens the outflow hydrograph, delays peak
flow and smoothes oscillations. In a sense, by introducing friction,
mesh dependant oscillations are eliminated. Although it would ap-
pear that by doing so mesh selection would be less important, this
58 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
is not so. Mesh coarsening proved to have a similar effect on the
outflow hydrograph as friction. As the mesh becomes coarser, the
hydrograph is somewhat dampened and peak flow is delayed. In
fact, the delays can be larger than those generated by physical fric-
tion. It becomes easy to mix the effects of physical friction and
numerical, mesh generated friction if a frictionless analysis is not
performed. This also implies that roughness coefficients must be
calibrated with every particular mesh that is used since each mesh
introduces its own amount of numerical friction. But it also implies
that a coarse mesh may be inappropriate because, on its own right
without friction, it can dampen and delay a hydrograph beyond the
effects of physical friction, making roughness calibration impossi-
ble. On a side note, roughness discretization was performed only
by mapping high resolution raster data to different meshes by
arithmetic averaging. It appears that, because of roughness distri-
bution in this particular case no significant error is introduced.
However, where important variations in roughness are present,
more complex averaging techniques, or even considering rough-
ness variations as criteria for mesh generation (in a fashion similar
to that of slope gradient) could result in better solutions.
For infiltration computations the SCS-CN method was selected
in this study, to test its response when coupled with a physically
based, distributed model for surface flow. The results were poor.
It was not possible to calibrate the SCS-CN parameters (CN and a)
to achieve an accurate representation of the hydrograph. In many
cases runoff volume was not achieved, and when runoff volume
was achieved, the hydrograph shape was poorly reproduced and
peak flow was delayed far beyond the experimental results. In
any case, it appears necessary to calibrate a as well as CN since
it was important effects on time distribution of infiltration. Fur-
thermore, when calibrating SCS-CN parameters using catchment
outflow hydrographs, care must be taken to remember that the
measured outflow volume is not equal to non-infiltrated volume.
Some of the runoff generated in a particular position of the catch-
ment may become ponded elsewhere, thus never reaching the
outlet. Hence, if SCS-CN is calibrating based only on outflow vol-
ume from experimental outflow hydrographs, the simulated re-
sults will always yield less outflow volume since some runoff
will be lost to ponding. This has a simple solution, which is to
measure the available ponding volume in the catchment by flood-
ing the catchment with the same storm (precipitation volume)
and allowing enough simulation time to only have static water
in the catchment. In this way, ponding volume is known and
can be added to outflow volume to calibrate SCS-CN parameters.
In this way SCS-CN will compute runoff volume, understood as
non-infiltrated volume, which must be equal to outflow volume
and ponded volume. This behavior may become less important
as catchment scale changes. In this particular case, ponding vol-
ume is practically the same as outflow volume and errors grow
rapidly and are very significant. In larger catchments, with less
infiltration, or higher precipitated volumes and lower ponding
volumes it may not be as significant. Still, it is easy to consider.
However, the results show that even when considering this ef-
fects, resulting in correct outflow volumes, hydrographs are
poorly simulated. In this case, SCS-CN is not appropriate to couple
with a distributed model for runoff computations. In a lumped
approach, SCS-CN may be coupled with synthetic unit hydrograph
techniques, and SCS-CN may still be valid, since it can be accu-
rately calibrated to simulate outflow volume, and the synthetic
unit hydrograph calibration parameters may have enough free-
dom to generate a hydrograph similar to the field measurements.
With a physically-based distributed model, such freedom is lost,
and the infiltration method must accurately compute results in
time, not only lumped volumes. Because SCS-CN appears to be
a poor choice, in future work, Horton and Green-Ampt methods
will be tested.
Acknowledgement
This work was partially supported and funded by the Spanish
Ministry of Science and Technology under research project
CGL2008-05153-C02-02.
References
Baltas, E.A., Dervos, N.A., Mimikou, M.A., 2007. Technical note: determination of the
SCS initial abstraction ratio in an experimental watershed in Greece. Hydrology
and Earth System Sciences 11, 1825–1829.
Beck, H.E., de Jeu, R.A.M., Schellekens, J., van Dijk, A.I.J.M., Bruijnzeel, L.A., 2009.
Improving Curve Number based storm runoff estimates using soil moisture
proxies. IEEE Journal of Selected Topics in Applied Earth Observations and
Remote Sensing 2, 250–259.
Burguete, J., Garcia-Navarro, P., Murillo, J., Garcia-Palacin, I., 2007. Analysis of the
friction term in the one-dimensional shallow-water model. Journal of Hydraulic
Engineering-ASCE 133, 1048–1063.
Casas, A., Benito, G., Thorndycraft, V., Rico, M., 2006. The topographic data source of
digital terrain models as a key element in the accuracy of hydraulic flood
modelling. Earth Surface Processes and Landforms 31, 444–456.
Cook, A., Merwade, V., 2009. Effect of topographic data, geometric configuration and
modeling approach on flood inundation mapping. Journal of Hydrology 377,
131–142.
Garcia-Ruiz, J., Arnaez, J., Begueria, S., Seeger, M., Marti-Bono, C., Regues, D., Lana-
Renault, N., White, S., 2005. Runoff generation in an intensively disturbed,
abandoned farmland catchment, Central Spanish Pyrenees. Catena 59, 79–92.
Hardy, R.J., Bates, P.D., Anderson, M.G., 1999. The importance of spatial resolution in
hydraulic models for floodplain environments. Journal of Hydrology 216, 124–
136.
Horritt, M.S., Bates, P.D., 2001. Effects of spatial resolution on a raster based model
of flood flow. Journal of Hydrology 253, 239–249.
Horritt, M.S., Bates, P.D., Mattinson, M.J., 2006. Effects of mesh resolution and
topographic representation in 2D finite volume models of shallow water fluvial
flow. Journal of Hydrology 329, 306–314.
Hunter, N.M., Bates, P.D., Horritt, M.S., Wilson, M.D., 2007. Simple spatially-
distributed models for predicting flood inundation: a review. Geomorphology
90, 208–225.
Lana-Renault, N., 2007. Respuesta hidrológica y sedimentológica en una cuenca de
montaña media afectada por cambios de cubierta vegetal: la cuenca
experimental de Arnás, Pirineo Central. Ph.D. thesis, Universidad de Zaragoza.
Lana-Renault, N., Latron, J., Reguees, D., 2007. Streamflow response and water-table
dynamics in a sub-Mediterranean research catchment (Central Pyrenees).
Journal of Hydrology 347, 497–507.
Leveque, R., 2002. Finite Volume Methods for Hyperbolic Problems. Cambridge
University Press.
López-Barrera, D., García-Navarro, P., Brufau, P., 2011. Sources of uncertainty in the
validation of a coupled hydrological-hydraulic simulation model with sediment
transport. La Houille Blanche.
López-Barrera, D., 2011. Desarrollo de un modelo hidrológico/hidrodinámico de
simulación de procesos de erosión y sedimentación. Ph.D. thesis, Universidad de
Zaragoza.
Merwade, V., 2009. Effect of spatial trends on interpolation of river bathymetry.
Journal of Hydrology 371, 169–181.
Mishra, S., Singh, V., 1999. Another look at SCS-CN method. Journal of Hydrological
Engineering 4, 257–264.
Mishra, S., Singh, V., Sansalone, J., Aravamuthan, V., 2003. A modified SCS-CN
method: characterization and testing. Water Resources Management 17, 37–68.
Murillo, J., Garcia-Navarro, P., 2004. A new topography adapted mesh refinement for
2D unsteady shallow water flow simulations over dry beds. In: Liong, S.-Y.,
Phoon, K.-K., Babovic, V. (Eds.), Proceedings of the 6th International Conference
on Hydroinformatics, Singapore.
Murillo, J., Garcia-Navarro, P., 2010. Weak solutions for partial differential equations
with source terms: application to the shallow water equations. Journal of
Computational Physics 229, 4327–4368.
Murillo, J., Garcia-Navarro, P., Burguete, J., 2009. Time step restrictions for well-
balanced shallow water solutions in non-zero velocity steady states.
International Journal for Numerical Methods in Fluids 60, 1351–1377.
Murillo, J., Garcia-Navarro, P., Burguete, J., Brufau, R., 2007. The influence of source
terms on stability, accuracy and conservation in two-dimensional shallow flow
simulation using triangular finite volumes. International Journal for Numerical
Methods in Fluids 54, 543–590.
Ponce, V.M., Hawkins, R.H., 1996. Runoff Curve Number: Has it reached maturity?
Journal of Hydrologic Engineering 1, 11–19.
Schubert, J.E., Sanders, B.F., Smith, M.J., Wright, N.G., 2008. Unstructured mesh
generation and landcover-based resistance for hydrodynamic modeling of
urban flooding. Advances in Water Resources 31, 1603–1621.
SCS, 1964. SCS National Engineering Handbook. Soil Conservation Service, USA
Department of Agriculture.
Seeger, M., Errea, M., Beguería, S., Martí, C., García-Ruiz, J., Arnáez, J., 2004.
Catchment soil moisture and rainfall characteristics as determinant factors for
discharge/suspended sediment hysteretic loops in a small headwater
catchment in the Spanish Pyrenees. Journal of Hydrology 288, 299–311.
 D. Caviedes-Voullième et al. / Journal of Hydrology 448–449 (2012) 39–59
Serrano-Pacheco, A., 2009. Simulación numérica bidimensional de procesos
hidrológicos e hidráulicos sobre lecho irregular deformable. Ph.D. thesis,
Universidad de Zaragoza.
Shewchuk, J.R., 1996. Triangle: engineering a 2D quality mesh generator and
Delaunay triangulator. In: Lin, M.C., Manocha, D. (Eds.), Applied Computational
Geometry: Towards Geometric Engineering. Springer-Verlag, vol. 1148 of
Lecture Notes in Computer Science, pp. 203–222. From the First ACM
Workshop on Applied Computational Geometry.
Shi, Z.H., Chen, L.D., Fang, N.F., Qin, D.F., Cai, C.F., 2009. Research on the SCS-CN
initial abstraction ratio using rainfall-runoff event analysis in the three gorges
area, china. Catena 77, 1–7.
59
Wilson, M., Atkinson, P., 2005. Prediction uncertainty in floodplain elevation and its
effect on flood inundation modelling. In: Atkinson, P.M., Foody, G.M., Darby, S.E.,
Wu, F. (Eds.) Proceedings of the 7th International Conference on
GeoComputation, Southampton, England, 2003, pp. 185–202.
Woodward, D., Hawkings, R., Hjelmfelt, A., Van Mullen, J., Quan., Q., 2002. Curve
Number method: origins, applications and limitations. In: Second Federal
Interagency Hydrologic Modeling Conference, Las Vegas, Nevada.