Journal of Hydrology 387 (2010) 33–45

Contents lists available at ScienceDirect

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

A simple inertial formulation of the shallow water equations for efficient

two-dimensional flood inundation modelling
Paul D. Bates a,*, Matthew S. Horritt b, Timothy J. Fewtrell a
a
School of Geographical Sciences, University of Bristol, University Road, Bristol BS8 1SS, UK
b
Halcrow Ltd., Burderop Park, Swindon, Wiltshire SN4 0QD, UK

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

Article history: This paper describes the development of a new set of equations derived from 1D shallow water theory for
Received 27 May 2009 use in 2D storage cell inundation models where flows in the x and y Cartesian directions are decoupled.
Received in revised form 19 October 2009 The new equation set is designed to be solved explicitly at very low computational cost, and is here tested
Accepted 22 March 2010
against a suite of four test cases of increasing complexity. In each case the predicted water depths com-
This manuscript was handled by
pare favourably to analytical solutions or to simulation results from the diffusive storage cell code of
K. Georgakakos, Editor-in-Chief, with the Hunter et al. (2005). For the most complex test involving the fine spatial resolution simulation of flow
assistance of Ehab A. Meselhe, Associate in a topographically complex urban area the Root Mean Squared Difference between the new formulation
Editor and the model of Hunter et al. is 1 cm. However, unlike diffusive storage cell codes where the stable
time step scales with (1/Dx)2, the new equation set developed here represents shallow water wave prop-
Keywords: agation and so the stability is controlled by the Courant–Freidrichs–Lewy condition such that the stable
Shallow water flow time step instead scales with 1/Dx. This allows use of a stable time step that is 1–3 orders of magnitude
Flood propagation greater for typical cell sizes than that possible with diffusive storage cell models and results in commen-
Inundation modelling surate reductions in model run times. For the tests reported in this paper the maximum speed up
Hydraulic modelling achieved over a diffusive storage cell model was 1120, although the actual value seen will depend on
model resolution and water surface gradient. Solutions using the new equation set are shown to be
grid-independent for the conditions considered and to have an intuitively correct sensitivity to friction,
however small instabilities and increased errors on predicted depth were noted when Manning’s n = 0.01.
The new equations are likely to find widespread application in many types of flood inundation modelling
and should provide a useful additional tool, alongside more established model formulations, for a variety
of flood risk management studies.
Ó 2010 Elsevier B.V. All rights reserved.

Introduction allowed a move away from large, irregular storage units to the dis-
cretization of the floodplain as a fine spatial resolution regular grid
Since first proposed by Zanobetti et al. (1970) methods to predict (cell areas of 102–103 km2). Here each cell within the grid is a
floodplain inundation using storage cell approaches have become storage area for which the mass balance is updated at each time
justifiably popular. Initially, such methods discretized floodplains step according to the fluxes of water into and out of each cell. Sim-
into irregular polygonal units representing large (surface areas of ilar to polygonal storage cell models, fluxes are calculated analyti-
100–101 km2) natural storage compartments and calculated the cally using uniform flow formulae but with the advantage of
fluxes of water between these according to some uniform flow for- higher resolution predictions and removal of the need for the mod-
mulae such as the weir or Manning’s equations. For many such eller to make explicit decisions about the location of storage com-
models in-channel flows are calculated using some form of the partments and the linkages between these. Numerous such
1D Saint–Venant equations, and when bankfull flow is exceeded models are now available (e.g. Estrela and Quintas, 1994; Bechteler
water is routed into and between the floodplain storage units. Most et al., 1994; Bates and De Roo, 2000) and a similar blueprint has
commercial 1D codes now include such a floodplain representation. increasingly been adopted in commercial modelling packages (e.g.
More recently the availability of increased computing power and JFLOW by JBA Ltd., FlowRoute by Ambiental and the RMS Ltd., UK
detailed descriptions of floodplain topography available through Flood Risk Model). Such models therefore solve a continuity equa-
remote sensing (e.g. LiDAR data, Marks and Bates, 2000) has tion relating flow into a cell and its change in volume:

Dh DQ
* Corresponding author. Tel.: +44 117 928 9108; fax: +44 117 928 7878. ¼ ð1Þ
E-mail address: paul.bates@bristol.ac.uk (P.D. Bates). Dt DxDy

0022-1694/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2010.03.027
34 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

and a flux equation for each direction where flow between cells is of the cell at the next time step. This limiter replaces fluxes calcu-
calculated according to Manning’s law (only the x direction is given lated using Manning’s equation with values dependent on model
here): parameters, and hence when the flow limiter is in use floodplain
!1=2 flows are sensitive to grid cell size and time step, and insensitive
5=3 i1;j i;j
h h h to Manning’s n.
Q i;j
x ¼ flow Dy ð2Þ
n Dx Flow limiters were rarely discussed in journal publications at
the time as their significance was not appreciated, but it was clear
where hi,j is the water free surface height [L] at the node (i, j), Dx from the cell sizes and time steps used in these early applications
and Dy are the cell dimensions [L], t is the time [T], n is the Man- that for many cells at each time step a flow limiter was being in-
ning’s friction coefficient [L1/3 T], and Qx and Qy describe the volu- voked. As a result flow-limited storage cell models often showed
metric flow rates between floodplain cells [L3 T1]. Qy is defined very little sensitivity to floodplain friction and their results were
analogously to Eq. (2). The flow depth, hflow, represents the depth strongly dependent on the grid size and time step selected. There
through which water can flow between two cells, and is defined is a legitimate debate over the degree of sensitivity to floodplain
as the difference between the highest water free surface in the friction one should expect in an inundation model given that flood-
two cells and the highest bed elevation. These equations are solved plain velocities are usually very small, but the almost complete
explicitly using a finite difference discretization of the time deriva- lack of such sensitivity in certain applications of flow-limited stor-
tive term: age cell models appeared to be more than could be explained in
i1;j physical terms.
tþDt i;j
h  th
i;j
t
Qx  t Q i;j t i;j1
x þ Qy  t Q i;j
¼
y
ð3Þ A solution to this problem was provided by Hunter et al. (2005)
Dt DxDy based on adaptive time-stepping. This approach seeks to remove
where th and tQ represent depth and volumetric flow rate at time t the need to invoke the flow limiter (Eq. (4)) by finding the opti-
respectively, and Dt is the model time step which is held constant mum time step (large enough for computational efficiency, small
throughout the simulation. enough for stability) at each iteration. This optimum time step is
The advantage of the storage cell formulation is that fluxes are obtained using an analysis of the governing equations and their
calculated analytically so the computational costs per time step are analogy to a diffusion system which gives the following expression
potentially much lower than in equivalent numerical solutions of for Dt:
the full shallow water equations. The method is also simple in con-  1=2  1=2 !
cept and it is therefore relatively easy to develop and maintain Dx2 2n @h 2n @h
Dt ¼ min 5=3   ; 5=3   ð5Þ
code that can perform the calculations. Such methods also inter- 4 hflow @x hflow @y
face readily with newly available remotely sensed terrain data
which typically arrives in the form of a regular grid. For this reason A scheme that uses this criterion can be implemented by
the number of research and commercial codes based on these tech- searching the domain for the minimum time step value and using
niques has proliferated over the last decade (for a review see Hun- this to update h in Eq. (3). The time step will thus be adaptive and
ter et al. (2007)). Whilst the method can only be applied to change during the course of a simulation, but is uniform in space at
gradually varied flows and does not include inertia or the ability each time step. Hunter et al. (2005) tested this new uncondition-
to capture supercritical effects, for many floodplain inundation ally stable time step formulation against analytical solutions for
problems the representation is appropriate. wave propagation over flat and planar slopes and showed a consid-
Such storage cell models were originally conceived for applica- erable improvement over the classical fixed time-step version of
tion at relatively coarse grid resolutions (25–100 m) and early the model. Moreover, the adaptive scheme was shown to yield re-
applications showed that at these scales there was a distinct com- sults that were independent of grid size or choice of initial time
putational advantage over full solutions of the 2D Saint–Venant step and which showed an intuitively correct sensitivity to flood-
equations (see for example Horritt and Bates, 2001, 2002). This al- plain friction over spatially-complex topography. Hunter et al.
lowed new applications of hydraulic models to be considered (2006) went on to test the new version of the LISFLOOD-FP model
including Monte Carlo uncertainty analysis (Aronica et al., 2002), against real world flood extent and wave travel time data for the
inclusion of hydraulic models in ensemble forecasting chains (Pap- upper River Severn in the UK for a model at 60 m spatial resolution.
penberger et al., 2005) and model applications to domain scales or- The adaptive time step model showed a better absolute perfor-
ders of magnitude larger than anything previously attempted (e.g. mance than the classical fixed time-step version at this spatial res-
Wilson et al., 2007). Despite these successes, a number of concerns olution, but at approximately six times the computational cost. In
became apparent. First, unless the constant time step used to solve particular the adaptive model appeared able to simulate floodplain
Eq. (3) was small, simulations with storage cell models quickly wetting and drying more realistically.
developed ‘chequerboard’ type instabilities as all the water in a Despite this success, the results obtained by Hunter et al. (2006)
particular cell drained into the adjacent ones in a single (large) identified a fundamental problem with Eq. (5), namely that the
time step (Cunge et al., 1980). At the next time step, this situation optimum stable time step reduces quadratically with decreasing
would reverse and all the water would flow back. To solve this grid size. For an explicit code this means that the computational
problem many modellers introduced some kind of ‘flow limiter’ cost will increase as (1/Dx)4. For applications with grid sizes in
to prevent the solution over-shooting and too much water leaving the range for which LISFLOOD-FP was originally designed (25–
a given cell in a single time step. The flow limiter sets the maxi- 100 m, Bates and De Roo, 2000) this led to a 2–10 increase in sim-
mum flow that can occur between cells and is typically a function ulation times which could be offset through advances in processor
of flow depth, grid cell size and time step. In LISFLOOD-FP, for speed. Any residual cost increases could then be justified easily as
example, the flow limiter used is: simulations were more realistic. However, for the finer resolution
! (1–10 m) grids required for application of hydraulic models to ur-
i;j i1;j
DxDyðh  h Þ ban areas (Fewtrell et al., 2008) simulation costs increased by sev-
Q i;j i;j
x ¼ min Q x ; ð4Þ eral orders of magnitude such that at these scales adaptive time
4Dt
step storage cell codes actually proved slower than full 2D solu-
This value is determined by considering the change in depth of a tions of the shallow water equations. Hunter et al. (2008) found
cell, and ensuring it is not large enough to reverse the flow in or out during benchmark testing of six 2D models applied at 2 m spatial
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 35

resolution to a 0.4 km2 area of Glasgow UK that numerical solu- the acceleration due to gravity [L T2] and all other terms are de-
tions of the full shallow water equations completed a 2 h real time fined as above. For many floodplains flows advection is relatively
simulation in around 1 h (depending on code complexity and pro- unimportant (see Hunter et al. (2007) for a discussion of the magni-
cessor architecture), whilst the storage cell codes took approxi- tude of terms in the shallow water equations) so we neglect this
mately an order of magnitude longer. term, assume a rectangular channel and divide through by a con-
Although a less serious problem, the dependence of the time stant flow width, w [L], to obtain an equation in terms of flow per
step on the water surface slope in Eq. (5) also means that the time unit width, q [L2 T1]:
step is reduced for areas with flat water surfaces, where intuitively
we would expect the governing equations to be easier to solve. In @q gh@ðh þ zÞ gn2 q2
þ þ 4=3 ¼ 0 ð7Þ
the limit of a horizontal water surface, the time step is forced to @t @x R h
zero, whereas the solution (zero flow in all directions) is trivial. For wide, shallow flows we can approximate the hydraulic ra-
In practice, the divergence of computation times in time-adaptive dius, R, with the flow depth, h. We can now discretize Eq. (7) with
storage cell models as the water comes to rest is avoided by apply- respect to the time step, Dt, to give:
ing a linearization for small surface slopes (see Hunter et al., 2005).
q  gh @ðh þ zÞ gn2 q2
Adaptive time step storage cell codes are therefore incompati- tþDt  qt t
þ t þ 7=3t ¼ 0 ð8Þ
ble with the fine spatial resolution grids increasingly required for Dt @x ht
urban flood modelling. The only solution to date is to invoke a flow
limiter, but this leads to a poor representation of flow dynamics. And rearrange to give an explicit equation for q at time t + Dt:
Whilst for fine grids full 2D models give shorter simulation times " #
@ðht þ zÞ n2 q2t
at current processor speeds, for practical applications they are still qtþDt ¼ qt  ght Dt þ 10=3 ð9Þ
only able to treat small (<1 km2) areas at the required level of de- @x ht
tail. It is clear that to allow wide area urban flood modelling at fine
This gives an equation for the unit flow at the next time step,
spatial resolution a new hydraulic model formulation is required.
qt+Dt, in terms of qt, ht and z and hence can be solved explicitly at
Development and testing of such an approach is the fundamental
a very similar cost to Eq. (2), as it contains only a single additional
aim of this paper where we describe a set of flow equations for
term. The advantage of this formulation is that since the accelera-
adaptive time step storage cell models which can overcome the
tion term is now included, the water being modelled has some
quadratic dependency on grid size in Eq. (5) yet which can be
mass, and it is therefore less likely to generate the rapid reversals
solved analytically with approximately the same computational
in flow which lead to a chequerboard oscillation. Shallow water
cost as Eq. (2). The new scheme therefore retains all the computa-
wave propagation will also be represented, rather than the diffu-
tional advantages of storage cell models over full 2D codes whose
sive behaviour typical of previous storage cell models.
equations require expensive numerical solution, yet with none of
Eq. (9) can be improved further, since instabilities may still arise
the previous disadvantages. Below we describe the derivation of
at shallow depths when the friction term becomes large. Replacing
the new set of equations from first principles, and then the new
a qt in the friction term by a qt+Dt leads to an equation linear in the
formulation is subject to a number of analytical and benchmark
unknown qt+Dt but which has some of the improved convergence
tests of increasing complexity. Finally, results are discussed and
properties of an implicit time stepping scheme:
conclusions drawn.
" #
@ðht þ zÞ n2 qt qtþDt
Derivation of an inertial formulation of the shallow water
qtþDt ¼ qt  ght Dt þ 10=3
ð10Þ
@x ht
equations
Eq. (10) can rearranged into an explicit form for calculation of
The route to a new set of equations for fast inundation model- flows at the new time step in the model:
ling in two dimensions was identified in the urban model bench-
marking study of Hunter et al. (2008). It was clear from this qt  ght Dt @ðh@xt þzÞ
qtþDt ¼   ð11Þ
comparison that the lack of mass and inertia in Eq. (2) was the 10=3
1 þ ght Dtn2 qt =ht
key reason why storage cell models required the strict time step
control implied by Eq. (5). In gradually varying shallow water flows
The enhanced stability of Eq. (11) stems from the increase in the
the effect of inertia is to reduce fluxes between cells, yet in Eq. (5)
denominator as the friction term increases, forcing the flow to zero,
flux is merely a function of gravity and friction. Eq. (5) therefore
as would be expected for shallow depths. A similar approach is
overestimates fluxes, particularly, as noted above, in areas of deep
used in Liang et al. (2006) to improve the stability of a full 2D shal-
water where there is only a small free surface gradient. Hunter
low water model.
et al. (2008) suggested that the solution was to modify explicit
Unlike (2), Eq. (11) includes shallow water wave propagation so
storage cell codes to include inertial terms (or simple approxima-
while stability is improved, it is still subject to the Courant–Freid-
tions to these) that may allow the use of a larger stable time step,
richs–Levy condition:
and hence quicker run times. In addition, inclusion of inertial ef-
fects may also be important to represent the flow physics in partic-
V Dt
ular environmental settings. Cr ¼ ð12Þ
Dx
Our starting point for derivation of such an equation is therefore
the momentum equation from the quasi-linearized one-dimen- where the non-dimensional Courant number, Cr, needs to be less
sional Saint–Venant or Shallow Water equations: than 1 for stability and V is a characteristic velocity [L T1]. In the
" # case of a shallow water flow where advection is ignored this char-
@Q @ Q2 gA@ðh þ zÞ gn2 Q 2
þ þ þ ¼0 ð6Þ acteristic velocity is:
@t
|{z} @x A @x
|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} Rffl4=3
|fflffl Affl}
{zfflffl pffiffiffiffiffiffi
|fflfflfflfflffl{zfflfflfflfflffl}
acceleration
advection
water slope friction slope gh ð13Þ
3 1
pffiffiffiffiffiffi
where Q [L T ] is the discharge, A is the flow cross section area where gh is the celerity of a long wavelength, small amplitude
[L2], z is the bed elevation [L], R is the hydraulic radius [L], g is gravity wave. Eq. (12) gives a necessary but not sufficient condition
36 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

for model stability, and is used to estimate a suitable model time whilst Test 4 was run on a single 2.8 GHz node of a quad-core Intel
step at t + Dt: Xeon Harpertown E5462 processor with 12 Mb of cache memory.
The executables for both processors were built using the Intel
Dx
Dtmax ¼ a pffiffiffiffiffiffiffi ð14Þ C++ compiler. Test 4 has also been used by Hunter et al. (2008)
ght to benchmark the performance of six 2D inundation models, by
Lamb et al. (2009) to evaluate an implementation of the JFLOW
where a is a coefficient in the range 0.2–0.7 used to produce a stable
adaptive time step diffusive storage cell code designed to run on
simulation for most floodplain flow situations. The parameter a is
massively parallel Graphics Processor Units (GPUs) and by Schu-
included because Eq. (12) is not sufficient to ensure model stability,
bert et al. (2008) to test an unstructured finite element model for
because the assumption of small amplitude in calculating the wave
urban applications. These latter studies also report data on compu-
celerity is not always valid, and because of the inclusion of friction
tational efficiency which provides important information on the
terms in the model. The stable time step is therefore often some-
likely comparative speed of models built with our new equation
what less than that indicated by the Courant–Friedrichs–Lewy con-
set.
dition, and so the parameter a is introduced to reduce the time step.
Despite these limitations in the application of the condition, Eq. (14)
represents a useful approach to time step selection for a wide range Test 1: non-breaking wave propagation over a horizontal plane
of flow conditions, subject to an appropriate choice of a.
This time step is typically 1–3 orders of magnitude larger than Hunter et al. (2005) developed a one-dimensional analytical
the stable time step for the purely diffusive scheme of Eq. (5). solution for inundation model testing where the full Saint–Venant
Moreover, within this range, proportionally larger time step differ- equations can be simplified to yield an ordinary non-linear differ-
ences become apparent as the grid size decreases, as for Eq. (14) ential equation. This can then be solved analytically to provide rig-
time step scales with 1/Dx rather than (1/Dx)2. Hence, we expect orous validation solutions. The analytical solution presented below
the new flux equation and adaptive time step constraint to be sig- is for the propagation of a non-breaking wave over a horizontal
nificantly more computationally efficient than previous storage plane which allows us to test the ability of inundation models to
cell models. The performance of this new set of equations and simulate wave movement correctly in the absence of a bed slope
the extent of this potential improvement is analysed in the follow- term (i.e. So = 0). In fact, this is not a true analytical solution to
ing section. the inertial equation solved by the model and thus we would not
expect the model results to fully converge to the analytical solution
Model testing and results although at fine grid resolutions it should be a close approxima-
tion. The derivation of the analytical solution is given in Hunter
Eqs. (11) and (14) were implemented within the LISFLOOD-FP et al. (2005) and is not repeated here, however the final equation
hydraulic model of Bates and De Roo (2000). This code has been for the water depth, h, at any point in space x or at any time, t, is:
developed extensively since conception from a simple storage cell

model written in the PC-RASTER language (Bates and De Roo, 7  7=3

hx;t ¼ C  n2 u3 ðx  utÞ ð15Þ
2000), to a flow limited code written in C++ (Horritt and Bates, 3
2001) and finally to an adaptive time step code using Eq. (5) (Hun-
ter et al., 2005, 2006). At each step previous variants were retained where u is the component of depth-averaged velocity in the x direc-
in the code and can be readily switched on or off. Finally, the LIS- tion, C is a constant of integration, which can be fixed by referring to
FLOOD-FP code has recently been placed within a version control the initial conditions of the problem (i.e. h at x = 0 and t = 0) and all
system, re-written in modular form (Fewtrell, 2009) and parallel- other terms are defined as previously. Eq. (15) can now be used to
ized using Open-MP (Neal et al., 2009). Implementing and compar- provide an analytical solution against which the model can be
ing new code variants is thus relatively straightforward given tested with boundary conditions in the form of h(t) at x = 0, and ini-
previous work on refining the code structure and bug fixing. tial conditions in the form of h(x) at t = 0. Here this solution was
The new inertial formulation of LISFLOOD-FP was assessed
against a structured sequence of numerical experiments that pro-
vide a rigorous test of its numerical and computational perfor-
mance. Three of these tests are taken from Hunter et al. (2005),
and for the final test we simulate the urban flooding problem used
in the benchmarking studies of Hunter et al. (2008) and Fewtrell
et al. (2008). Specifically these tests are:

Test 1: Non-breaking wave propagation over a horizontal plane

and comparison to an analytical solution.
Test 2: Non-breaking wave run-up on a planar beach and com-
parison to an analytical solution.
Test 3: Wetting and drying of a planar beach (i.e. a full tidal
cycle).
Test 4: Simulation of flood propagation through a complex
street and building network at fine spatial resolution.
Fig. 1. Predicted water surface elevation (z) at t = 3600 s for wave propagation over
In each case the new inertial formulation is compared to the a horizontal beach simulated at 50 m spatial resolution and n = 0.03 using: (a) a
adaptive time step diffusive model (Eqs. (2) and (5)) of Hunter fixed time step (Dt = 1 s) flow limited diffusive model (Eqs. (2)–(4), light grey
et al. (2005) in terms of root mean squared error (or difference), dashed line); (b) an adaptive time step diffusive model (Eqs. (2), (3), and (5), mid
grey dashed line); and (c) the new adaptive time step inertial model developed in
% volume error, minimum time step during the simulation and to- this paper (Eqs. (3), (11), and (14), dark grey dashed line). Each model is compared
tal computational time. Tests 1–3 were run on a single 2.66 GHz to the analytical solution of Eq. (15) (solid black line). Summary numerical results
node of a dual-core Intel Core2 Duo processor with 3 Gb of RAM, for these simulations are reported in Table 1.
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 37

Table 1
Summary of numerical and computational efficiency results from Tests 1, 2, 3 and 4.

Test case Model Root Mean Square Error (RMSE, in m) from analytical Volume error Minimum time Total Speed up
solution, or Root Mean Squared Difference (RMSD, in from analytical step during computation using
m), from diffusive model solution (%) simulation (s) time (min) inertial
version
Test 1: horizontal beach, Diffusive 0.06 1.27 0.15 0.33
Dx = 50 m, n = 0.03 Inertial 0.03 1.25 7.25 0.02 17
Test 2: planar beach, Diffusive 0.02 0.17 0.02 1.22
Dx = 50 m, n = 0.03 Inertial 0.11 1.16 4.93 0.02 61
Test 3: wetting and drying Diffusive – – 0.03 1.80
of a planar beach, Inertial See Fig. 7 – 5.59 0.03 60
Dx = 50 m, n = 0.03
Test 4: Glasgow flooding, Diffusive 0.003 155.0
2 m resolution, spatially Inertial 0.01 0.43 1.47 105
uniform friction

implemented using parameter values of u = 1 ms1, Dx = 50 m and tions from these simulations at t = 3600 s are shown in Fig. 2 and
n = 0.03 m1/3s for a simulation of duration 3600 s. summarised in Table 2. The inertial model outperformed the diffu-
Fig. 1 shows the water surface elevations predicted by the diffu- sive model in terms of RMSE and volume error at Dx = 50, 100 and
sive and inertial models at the end of the simulation (i.e. at 200 m, whereas the diffusive model was marginally better at
t = 3600 s) compared to the analytical solution (solid black line) gi- Dx = 5, 10 and 25 m. Errors are low for all resolution models apart
ven by Eq. (15). For this case only, Fig. 1 also shows the water ele- from Dx = 200 m where the solution quality for both schemes is
vations predicted by a flow limited diffusive model with a fixed dominated by the effect of the coarse grid resolution which is
time step of 1 s. The Root Mean Square Error (RMSE), volume error,
minimum time step during the simulation and the total computa-
tion time for each of these simulations are also summarised in Ta- Table 2
ble 1. These results show that the new inertial formulation is able Impact of grid resolution on RMSE and volume error for simulations of non-breaking
to match the analytical solution well, with an RMSE half that of the wave propagation over a horizontal plane with n = 0.03.

diffusive solution (0.03 m compared to 0.06 m) and similar volume Grid Model Root Mean Volume Minimum Total
errors. By contrast, the flow limited diffusive model simulates resolution Square error time step computation
wave propagation, wave front position and water depths poorly (m) Error from during time (min)
(RMSE, in analytical simulation
and therefore is not considered here further. The minimum time
m) from solution (s)
step achieved during the inertial model run is also 48 larger analytical (%)
than that for the diffusive model, and this translates to a 17 speed solution
up in computational time. The gearing of minimum time step to 5 Diffusive 0.006 0.099 0.002 514.7
computation time is due to the fact that: (a) for short simulations Inertial 0.07 2.37 0.73 2.33
the fixed costs of running LISFLOOD-FP (data input and output, 10 Diffusive 0.013 0.071 0.006 32.02
data consistency checks, etc.) actually become a relatively large Inertial 0.065 2.25 1.45 0.35
25 Diffusive 0.03 0.67 0.04 2.60
proportion of the total simulation time; and (b) timings for simu- Inertial 0.05 1.89 3.62 0.12
lations which last only a few seconds may not necessarily be suffi- 50 Diffusive 0.06 1.27 0.15 0.33
ciently precise. Hence the full speed up potential may not be seen Inertial 0.03 1.25 7.25 0.02
for this particular test case. 100 Diffusive 0.09 2.67 0.60 0.05
Inertial 0.05 0.25 14.51 0.02
To test whether these results were sensitive to grid resolution,
200 Diffusive 0.15 4.94 2.41 0.03
identical simulations were also run with the diffusive and inertial Inertial 0.11 2.95 29.04 0.01
models at Dx = 5, 10, 25, 100 and 200 m. Predicted water eleva-

Fig. 2. Predicted water surface elevation (z) at t = 3600 s for wave propagation over a horizontal beach simulated at Dx = 5, 10, 25, 50, 100 and 200 m spatial resolution
(denoted with dark to light grey lines respectively) and n = 0.03 using: (a) an adaptive time step diffusive model (dashed lines); and (b) the new adaptive time step inertial
model (dotted lines). Each model is compared to the analytical solution (solid black line).
38 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

not able to resolve the steep water surface slopes at the wave front. the inertial formulation (0.785 s compared to 0.002 s) and this re-
At resolutions greater than Dx = 25 m the improvement in compu- sulted in a 220 decrease in run time. Fig. 3 shows the evolution
tation time also becomes more apparent as expected. Notably, at of the time step Dt over the simulation and shows the quadratic
Dx = 5 m, the minimum time step achieved was 360 larger for reduction in the time step with decreasing grid size in the diffusive

Fig. 3. Time step evolution for simulations of wave propagations over a horizontal beach at Dx = 5, 10, 25, 50, 100 and 200 m spatial resolution (denoted with dark to light
grey lines respectively) and n = 0.03 using: (a) an adaptive time step diffusive model (dashed lines); and (b) the new adaptive time step inertial model (solid lines).

Fig. 4. Predicted water surface elevation (z) at t = 3600 s for wave propagation over a horizontal beach simulated at Dx = 50 m using the adaptive time step diffusive model
(light grey lines) and new adaptive time step inertial model (mid grey lines) for n = 0.01, 0.03, 0.06 and 0.09. In each case the model is compared to the appropriate analytical
solution (solid black lines, note the changing range of the y axis).
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 39

models as a result of the Dx2 term in Eq. (5). Fig. 3 also shows that slope is much smaller and hence Eq. (5) leads to smaller stable time
with the diffusive model the time step continues to decrease over steps for the diffusive model. In fact the minimum time step
the simulation as depths increase and water surface slopes reduce. achieved over the simulation with the diffusive model is 20
By contrast the reduction of time step with grid size in the inertial smaller for a planar beach compared to the horizontal case at
formulation is more linear, and after an initial period of evolution 5 m resolution. This compares to a difference of only 1.5 for
stabilizes to a near uniform value as one would expect given the the inertial model at the same scale. Despite its lower accuracy,
velocity in the analytical solution is fixed at 1 ms1. the RMSE for the inertial model is still within the typical vertical
Lastly, sensitivity to friction was assessed by running the diffu- error of high resolution floodplain topographic surveys (e.g. those
sive and inertial models for n = 0.01, 0.03, 0.06 and 0.09 for a model derived using airborne laser altimetry) and likely to be less than
with Dx = 50 m. Results from these simulations are presented in uncertainties induced by boundary condition errors. It should also
Fig. 4 and Table 3 and show the inertial model outperforming the be noted that the increase in depth for this test case is relatively
diffusive scheme at frictions above n = 0.03, but the performance rapid compared to that which would be typical for most dynamic
advantage switching to the diffusive scheme for n = 0.01. Why floodplain inundation in lowland rivers. In reality in such situa-
the behaviour should change for low values of n is unknown at this tions flow evolves much more gradually and hence Test 2 actually
stage. One possible explanation is that for low Manning’s numbers represents rather a stringent case. It may therefore be that for
there is insufficient dissipation inherently in the numerical scheme many practical applications the minor increase in errors in pre-
to dissipate the energy of the flow. Hence, at n = 0.01 the accelera- dicted depth are acceptable given the large improvement in com-
tion terms in the shallow water equations start to dominate and putational efficiency, and that these can, like any other model
we get wave-like behaviour, whereas inertial LISFLOOD-FP is de- structural error, be compensated for during calibration.
signed for situations where there is a dominance of surface slope In Fig. 5 we show the impact of changing model resolution on
and friction terms. Thus for model domains dominated by very the prediction of non-breaking wave run-up. This shows predicted
low surface friction a full shallow water model may give more water surface elevations at t = 3600 s for the diffusive (dashed
accurate results. Even with this divergence from the analytical lines) and inertial models (dotted lines) at Dx = 5, 10, 25, 50, 100
solution at n = 0.01, the RMSE for the inertial model is still only and 200 m. Here there is no impact of resolution on the model re-
0.05 m and this is probably acceptable for many applications as sults and the only differences are generated by the choice of model
this is less than the vertical error in typically available terrain data formulation. All the diffusive model runs overlay the analytical
(e.g. LiDAR). For both models RMSE errors increase with increasing solution, apart from where h ? 0 where the impact of grid size
friction, to a maximum in these tests of 0.1 m for the inertial can be seen. All the inertial models also over plot, but lag the
scheme and 0.14 m for the diffusive scheme at n = 0.09. numerical solution by a short distance. This result, in combination
with the accuracy assessment in Table 4, suggests that the inertial
Test 2: non-breaking wave run-up on a planar beach models are in fact tending to a subtly different numerical solution,
as noted above. The computation time for the inertial model at
The second test case developed by Hunter et al. (2005) consists Dx = 5 m is 1120 shorter than for the diffusive model as a result
of solving Eq. (15) for a planar beach (i.e. where So – 0). Here no di- of the large difference in stable time step (0.493 s compared to
rect analytical solution exists, but an accurate numerical solution 0.0001 s) which in turn is a consequence of the shallow water sur-
to Eq. (15) with uniform velocity can be obtained using a 4th order face slope generated by this test case. Furthermore, Fig. 6 shows
Runge–Kutta scheme. Again this can be used to develop initial con- that the under-prediction of wave front position is related to fric-
ditions, boundary conditions and a numerical solution against tion and that the effect has largely disappeared when n = 0.06, at
which the model can be tested. This numerical solution was imple- and above which both diffusive and inertial models perform
mented using parameter values of u = 1 m s1, Dx = 50 m, equally well. For the simulations at n = 0.01 there is also a
So = 103 mm1 and n = 0.03 m1/3s for a simulation of duration suggestion of some minor instabilities with the inertial model.
3600 s. Again the models are compared in terms of Root Mean
Square Error (RMSE), volume error, minimum time step during
the simulation and the total computation time. These the results
are summarised in Table 1. In this case the 50 m diffusive model
has lower RMSE than the inertial model (0.02 m compared to
0.11 m), but at a significantly greater computational cost. The min-
imum time step for the diffusive model is 250 smaller than that
for the inertial model and this translates into a simulation time
that is 61 longer. The greater speed up with the inertial model
here is due to the fact that in this case the minimum water surface

Table 3
Impact of friction on RMSE and volume error for simulations of non-breaking wave
propagation over a horizontal plane with Dx = 50 m.

Friction Model Root Mean Square Error Volume error from

(Manning’s (RMSE, in m) from analytical analytical solution
n) solution (%)
0.01 Diffusive 0.02 1.15
Inertial 0.05 4.98
0.03 Diffusive 0.06 1.27
Inertial 0.03 1.25 Fig. 5. Predicted water surface elevation (z) at t = 3600 s for wave propagation up a
0.06 Diffusive 0.10 1.28 planar beach simulated at Dx = 5, 10, 25, 50, 100 and 200 m spatial resolution
Inertial 0.06 0.00 (denoted with dark to light grey lines respectively) and n = 0.03 using: (a) an
0.09 Diffusive 0.14 1.27 adaptive time step diffusive model (dashed lines); and (b) the new adaptive time
Inertial 0.10 0.31 step inertial model (dotted lines). Each model is compared to the numerical
solution (solid black line).
40 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

Table 4 mentation of a sinusoidal wave boundary condition at x = 0 (see

Impact of grid resolution on RMSE and volume error for simulations of non-breaking for example Leendertse and Gritton, 1971; Falconer and Chen,
wave run-up on a planar beach with n = 0.03.
1991). In this case we use a wave of amplitude 4 m, period 4 h,
Grid Model Root Mean Volume Minimum Total and So = 103 m m1 for a simulation of duration 7200 s. This gives
resolution Square error time step computation boundary conditions comparable to those used in Test 2. As far as
(m) Error from during time (min)
(RMSE, in analytical simulation
the authors are aware, there is no analytical solution to this prob-
m) from solution (s) lem so in this case we simply look at differences between the dif-
analytical (%) fusive and inertial formulations. These are shown in Fig. 7 for a
solution simulation using Dx = 50 m and n = 0.03. This shows smaller differ-
5 Diffusive 0.030 0.464 0.0001 4834.8 ences between the inertial and diffusive formulations than for Test
Inertial 0.107 1.069 0.493 4.3 2. For example, the RMSD at the end of flow advance at 3600 s is
10 Diffusive 0.031 0.467 0.001 303.5
only 0.074 m, compared to 0.11 m for Test 2 at the same time.
Inertial 0.107 1.083 0.986 0.65
25 Diffusive 0.036 0.536 0.004 9.12 Overall, Fig. 7 shows wave front position during wetting and drying
Inertial 0.109 1.148 2.465 0.07 simulations to differ only marginally between the two formula-
50 Diffusive 0.040 0.553 0.018 0.73 tions, and for wave shape to show noticeable differences only at
Inertial 0.108 1.158 4.931 0.05 the start of flood wave recession (maximum RMSD of 0.137 m at
100 Diffusive 0.047 0.590 0.072 0.1
4500 s). Compared to Test 2 the differences between the inertial
Inertial 0.118 1.349 9.876 0.02
200 Diffusive 0.060 0.674 0.296 0.05 and diffusive formulations at lower frictions are also less pro-
Inertial 0.127 1.494 19.78 0.02 nounced. This is shown in Fig. 8 where we compare the output
from dynamic wetting and drying simulations at t = 1800, 3600,
5400 and 7200 s for n = 0.01, 0.03, 0.06 and 0.09 for both inertial
Test 3: wetting and drying of a planar beach and diffusive models. Maximum differences occur at t = 3600 s
and 4500 s for n = 0.01 and n = 0.03, whilst at other times and for
Tests 1 and 2 have tested the ability of the new model formula- other frictions the results are only marginally different. This sug-
tion to simulate wave propagation over flat and sloping flood- gests that the differences between diffusive and inertial formula-
plains, however correct simulation of floodplain inundation over tions shown in Figs. 5 and 6 may be a worst case scenario and
whole events also requires the accurate representation of flow that differences may be less marked for more realistic cases.
reversals and inundation front recession during floodplain drying. Fig. 8 also confirms the presence of small instabilities during the
The ability of the model to do this can be evaluated by extending initial phase of the simulation with the inertial model when
Test 2 to represent wave run-up and drying through the imple- n = 0.01, however these die out as the simulation proceeds.

Fig. 6. Predicted water surface elevation (z) at t = 3600 s for wave propagation up a planar beach simulated at Dx = 50 m using the adaptive time step diffusive model (light
grey lines) and new adaptive time step inertial model (mid grey lines) for n = 0.01, 0.03, 0.06 and 0.09. In each case the model is compared to the appropriate numerical
solution (solid black lines, note the changing range of the y axis).
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 41

Fig. 7. Predicted water surface elevation (z) during wetting and drying of a planar beach simulated using Dx = 50 m and n = 0.03 for: (a) an adaptive time step diffusive model
(grey lines); and (b) the new adaptive time step inertial model (black lines).

In terms of computational cost, Test 3 should be relatively 60 shorter. Fig. 9 shows the time step evolution for the diffusive
expensive for the diffusive model to solve as during flow reversal and inertial simulations for Test 3 and highlights the fact that the
and the start of wave front recession the water surface profile be- minimum stable time step differs by 1–2 orders of magnitude over
comes near horizontal. Fig. 7 clearly shows this happening be- the majority of the simulation. Moreover, as expected the time step
tween 4500 and 5400 s. Flow reversals are a necessary feature of evolution for the inertial model is quasi-linear and after an initial
any dynamic flood simulation and cause the minimum time step period of evolution stabilizes to a near uniform value over the
in a diffusive model to become very small because of the presence whole simulation.
of the free surface gradient terms in Eq. (5). Hence we would ex-
pect greater computational savings with the inertial model for Test Test 4: fine spatial resolution simulation of urban inundation
3 than for Tests 1 or 2. This is clearly shown in Table 1 where the
inertial scheme results in a minimum time step 186 larger than Tests 1–3 have demonstrated the numerical and computational
that for the diffusive model and a total simulation time that is performance of the inertial formulation in a series of idealised

Fig. 8. Predicted water surface elevation (z) during wetting and drying of a planar beach simulated using Dx = 50 m and n = 0.01, 0.03, 0.06 and 0.09 (black to light grey lines
respectively) for: (a) the new adaptive time step inertial model (top four panels) and (b) an adaptive time step diffusive model (bottom four panels).
42 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

Flooding at this site is caused, at least in part, by a small (1 m

wide) stream that enters near the north-east corner of the domain
(located at point XQ on Fig. 10a) and almost immediately enters a
culvert that runs under the entire site. Flooding has been observed
to occur here as a result of flow exceeding the capacity of the cul-
vert and spilling into the street network. Once the capacity of the
culvert is exceeded water flows along two main east–west oriented
streets before converging and ponding in low-lying areas in the
southern part of the domain.
The flow event simulated is based on a real flood that occurred
at this site on 30 July 2002. The inflow boundary condition con-
sisted of the hydrograph shown in Fig. 11, which was imposed as
a point source internal to the model domain at location XQ. This
hydrograph represents the water volume overflowing the culvert
and lasts <60 min, but as in Hunter et al. (2008) simulations were
continued for 120 min to allow water to come to rest and pond in
depressions. All external boundaries for each model were closed as
mass flux across the external boundary is negligible. Lastly, friction
was represented as in Fewtrell et al. (2008) using a single compos-
ite value of n = 0.035. This was selected based on the spatial distri-
Fig. 9. Time step evolution during wetting and drying of a planar beach simulated
using Dx = 50 m and n = 0.03 for: (a) an adaptive time step diffusive model (dashed
bution of land use within the domain as determined from the OS
line); and (b) the new adaptive time step inertial model (solid line). MastermapÒ data.
Simulations were run with these data using the adaptive time
step storage cell formulation of Hunter et al. (2005) and the new
inertial formulation developed in this paper. Fig. 12 shows the pre-
cases of increasing complexity. However, the critical test of the
dicted water depths at the end of the simulation (120 min) for both
model is whether it is able to simulate flood propagation over com-
formulations and a map of the absolute differences in water depth
plex topography and reduce the long run times of storage cell
(diffusive minus inertial) at this time. Overall flood extents differ
codes when applied at fine spatial resolution. To test this we used
only marginally and the Root Mean Squared Difference in water
the new inertial formulation of LISFLOOD-FP to simulate the urban
inundation test case used in Hunter et al. (2008) and Fewtrell et al.
(2008). This comprises a 1.0  0.4 km domain in the Greenfields
area of Glasgow, UK which has been observed to flood in response
to heavy rainfall in the small (5 km2) upstream catchment. For
this site digital elevation data at 1 m resolution were available
from an airborne laser altimetry survey that can be used to build
high resolution inundation models. This was supplemented by
Hunter et al. (2008) with Ordnance Survey (OS) MastermapÒ digi-
tal map data that defined building locations, the road network and
land-use type as vector layers. The LiDAR data acquired for this
study had already been filtered to remove vegetation and building
features to leave a ‘bare earth’ digital elevation model (DEM) with
horizontal and vertical accuracies less than 50 cm and 15 cm Root
Mean Square Error (RMSE) respectively. For hydraulic modelling
Hunter et al. (2008) aggregated the ‘bare earth’ LiDAR data to
2 m and reinserted buildings, kerbs and roads based on their loca-
tions in the digital map layer. Fig. 10 shows: (a) the road and build-
ing layout at this study site overlaid onto the surface height (z)
from the benchmark DEM and (b) a high resolution aerial photo Fig. 11. Event hydrograph simulated in Test 4. The vertical dashed lines at 30 and
of the same area. 60 min represent instances for which model results are presented in later figures.

(a) (b)

Fig. 10. The Greenfields study site (Test 4): (a) building and road topology derived from Ordnance Survey MastermapÒ data with the surface height (z) from the benchmark
DEM shown as a grey scale and (b) high resolution aerial photo of the study site. All map plots are in Cartesian coordinates where east–west is oriented along the x axis and
north–south along the y axis. Dimensions are in m.
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 43

depth is only 0.01 m (see Table 1). Maximum differences in water olution unstructured grid, full shallow water model and Lamb et al.
depth are ±0.1 m, but these are very localized and over the major- (2009) who report a 9 min run time for a structured grid, adap-
ity of the domain water depth differences are close to the RMSD. tive time step diffusive model (JFLOW) run on a massively parallel
The largest difference occurs at x = 180, y = 250 where the flood ex- Graphics Processor Unit (GPU). Whilst these are not controlled
tent is slightly greater in the diffusive version and fills up a small tests conducted with identical processors and compilers (for exam-
(0.1 m deep) depression not flooded by the inertial model. Other ple the processors used for the simulations reported by Hunter
areas of difference, such as the higher predicted depths in the iner- et al. are now relatively old), these results do suggest that the
tial model at x = 700, y = 200 at the end of the simulation, may be a new inertial formulation would be faster for this case than any pre-
result of the additional physics in this scheme obtained by includ- viously applied code. This is a significant advance on diffusive stor-
ing the acceleration terms from the full shallow water equations. age cell models which were shown by Hunter et al. (2008) to be an
The run time for this test is 105 shorter for the inertial model order of magnitude slower than full shallow water models at this
(1.47 min compared to 155 min for the diffusive model) which is resolution because of the quadratic dependency on grid size in
in line with theoretical expectations. The total computation time Eq. (5). Moreover, when one couples the speed up achieved here
for the inertial model also compares favourably with the run times with the 5–6 speed up in LISFLOOD-FP run times achieved by
reported for other classes of model applied at this test site. These Neal et al. (2009) using Open-MP parallelization on an 8 core
include Hunter et al. (2008) who report run times of around processor, the potential for a 600 reduction in run times for
60 min for various structured grid, full shallow water models, this test case (i.e. down to 0.25 min) becomes a realistic
Schubert et al. (2008) who report a 18 min run time for a 2 m res- expectation.

Fig. 12. Predicted water depths for at the end of the simulation for Test 4 (120 min) using a grid resolution of Dx = 2 m for: (a) an adaptive time step diffusive mode and (b)
the new adaptive time step inertial model. Panel (c) shows the difference in predicted water depths (diffusive minus inertial) between (a) and (b).
44 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45

Discussion and conclusions Friedrichs–Lewy condition and the approach of Hunter et al.
(2005). This approach has been shown to be reasonable through
The results outlined in the preceding section show the new the test cases presented in this paper, for models of typical natural
inertial formulation of the shallow water equations developed in floodplains. While the absence of a rigorous analysis of stability is a
this paper to produce flow predictions that compare favourably drawback compared to other numerical schemes, in some circum-
with analytical solutions for non-breaking wave propagation over stances this is outweighed by the advantages of simplicity, which
horizontal and planar beaches and to the results from an adaptive make implementation on parallel and non-standard architectures
time step diffusive storage cell code. Like the diffusive model, the as discussed above far easier. However, for model domains charac-
inertial code is relatively insensitive to grid resolution and displays terised by very low surface friction a full shallow water model may
an intuitively correct response to changing friction. give more accurate results.
The inertial code performs slightly better than the diffusive Future research should test the new inertial formulation for fur-
code for a horizontal beach and slightly worse for a planar beach ther test cases and seek to benchmark model performance (in
when compared to the analytical solution. However for either code terms of both predictions and computational times) against a vari-
the Root Mean Squared Error is always less than the typical vertical ety of other model types in controlled experiments such as those
error (0.1 m) in high resolution terrain data (such as airborne la- described by Hunter et al. (2008). Further work should also be con-
ser altimetry or LiDAR) used for inundation modelling. Water ducted to try to improve stability of the new equation set at low
depth errors are independent of grid scale but do vary with friction, friction or examine the possibility of developing models capable
with the inertial model tending to perform worse than the diffu- of using different physical formulations in different parts of the
sive model when n = 0.01. Differences between the inertial and dif- model domain depending on changing flow dynamics. A spatially
fusive codes become less marked for more realistic cases involving varying time step may also improve model efficiency, although
dynamic wetting and drying and for higher frictions, and for a real previous work on this for 1D models shows that the potential
test case involving the fine spatial resolution simulation of flow in speed up is reduced significantly by other computational over-
a topographically complex urban area the Root Mean Squared Dif- heads associated with changing the time step for different parts
ference is 1 cm. Whilst water depths predicted by the new for- of the model (Wright et al., 2003). Despite this need for ongoing re-
mulation are similar to benchmark solutions, these results are search, this paper has demonstrated the utility of the new inertial
achieved at a significantly reduced computational cost because formulation against a series of stringent tests of increasing com-
the minimum stable time step scales with Dx, rather than with plexity. The new equations are therefore likely to find widespread
(1/Dx)2 as would be the case for a purely diffusive scheme. The ex- application in many types of flood inundation modelling and
act speed up over a diffusive code will depend on grid resolution should provide a useful additional tool, alongside more established
and water surface gradients within the flow domain but from the- model formulations, for a variety flood risk management studies.
oretical considerations is likely to be 1–3 orders of magnitude. The
maximum speed up achieved for the tests reported here is 1120
and as a result the inertial code will most likely be faster than Acknowledgements
either diffusive or full shallow water models at any given spatial
resolution. Moreover, as the explicit equation we here describe is Timothy Fewtrell’s work on this paper forms part of his Re-
relatively easy to code it should be simple to parallelize using search Fellowship funded by the Willis Research Network (http://
Open-MP techniques (e.g. Neal et al., 2009) or software tools such www.willisresearchnetwork.com). The University of Bristol Blue-
as NVIDIA’s Kuda which allow models to run on massively parallel Crystal high performance computer was used to undertake the
GPU cards (e.g. Lamb et al., 2009). This will pave the way to further simulations reported in this paper.
substantial reductions in run times and make possible a whole
range of new applications of hydraulic models. These could include
References
whole city risk analyses at the native spatial resolution of LiDAR
data (i.e. 0.25–1 m), real time two-dimensional inundation fore- Aronica, G., Bates, P.D., Horritt, M.S., 2002. Assessing the uncertainty in distributed
casting using ensemble data assimilation, multi-year simulations model predictions using observed binary pattern information within GLUE.
Hydrological Processes 16 (10), 2001–2016.
of flows at continental scales (e.g. the Amazon River basin, Wilson
Bates, P.D., De Roo, A.P.J., 2000. A simple raster-based model for flood inundation
et al., 2007), and explorations of model uncertainty using Monte simulation. Journal of Hydrology 236 (1–2), 54–77.
Carlo analyses with orders of magnitude more realisations than Bechteler, W., Hartmann, S., Otto, A.J., 1994. Coupling of 2D and 1D models and
has hitherto been possible. integration into Geographic Information Systems (GIS). In: White, W.R., Watts,
J. (Eds.), 2nd International Conference on River Flood Hydraulics. John Wiley
Care should be taken when using the new inertial formulation and Sons Ltd., Chichester, pp. 155–166.
for domains where large areas of low friction land use dominate Cunge, J.A., Holly, F.M., Verwey, A., 1980. Practical Aspects of Computational River
(i.e. where n is equal to 0.01), as here predicted water depth er- Hydraulics. Pitman Publishing, London. 420 pp.
Estrela, T., Quintas, L., 1994. Use of a GIS in the modelling of flows on floodplains. In:
rors increase and small instabilities can creep into the solution. White, W.R., Watts, J. (Eds.), 2nd International Conference on River Flood
The instability seen at low Manning’s n results from the hybrid nat- Hydraulics. John Wiley and Sons Ltd., Chichester, pp. 177–190.
ure of the model, and shortcomings in the stability criteria applied. Falconer, R.A., Chen, Y., 1991. An improved representation of flooding and drying
and wind stress effects in a 2-D tidal numerical model. Proceedings of
The Courant–Friedrichs–Lewy condition is more usually applied to Institution of Civil Engineers: Research and Theory 91 (12), 659–678.
advective problems where upwinding is applied, and thus may not Fewtrell, T.J., 2009. Development of simple numerical methods for improving two-
be strictly applicable to the centred difference approach adopted dimensional hydraulic models of urban flooding. Unpublished Ph.D. Thesis,
University of Bristol.
here. Hunter et al. (2005) showed that without the inertia terms,
Fewtrell, T.J., Bates, P.D., Horritt, M., Hunter, N., 2008. Evaluating the effect of scale
stable time step is proportional to n (Eq. (5)), and hence will be dri- in flood inundation modelling in urban environments. Hydrological Processes
ven to zero if the friction is neglected. Friction is acting to stabilise 22, 5107–5118.
Horritt, M.S., Bates, P.D., 2001. Predicting floodplain inundation: raster-based
the scheme in this respect, with stability increased further by addi-
modelling versus the finite-element approach. Hydrological Processes 15 (5),
tion of the inertia term as described in this paper. A rigorous anal- 825–842.
ysis of the stability of a non-linear model including both friction Horritt, M.S., Bates, P.D., 2002. Evaluation of 1-D and 2-D numerical models for
and wave propagation behaviour will be complex, if not impossi- predicting river flood inundation. Journal of Hydrology 268 (1–4), 87–99.
Hunter, N.M., Horritt, M.S., Bates, P.D., Wilson, M.D., Werner, M.G.F., 2005. An
ble, but the analysis presented here is a pragmatic approach taking adaptive time step solution for raster-based storage cell modelling of floodplain
into account the limiting behaviours understood by the Courant– inundation. Advances in Water Resources 28, 975–991.
 P.D. Bates et al. / Journal of Hydrology 387 (2010) 33–45 45

Hunter, N.M., Bates, P.D., Horritt, M.S., Wilson, M.D., 2006. Improved simulation of Neal, J.C., Fewtrell, T.J., Trigg, M.A., 2009. Parallelisation of storage cell flood
flood flows using storage cell models. Proceedings of the Institution of Civil models using OpenMP. Environmental Modelling and Software 24 (7),
Engineers, Water Management 159 (1), 9–18. 872–877.
Hunter, N.M., Bates, P.D., Horritt, M.S., Wilson, M.D., 2007. Simple spatially- Pappenberger, F., Beven, K.J., Hunter, N., Bates, P.D., Gouweleeuw, B., Thielen, J., De
distributed models for predicting flood inundation: a review. Geomorphology Roo, A.P.J., 2005. Cascading model uncertainty from medium range weather
90, 208–225. forecasts (10 days) through a rainfall–runoff model to flood inundation
Hunter, N.M., Bates, P.D., Neelz, S., Pender, G., Villanueva, I., Wright, N.G., Liang, D., predictions within the European Flood Forecasting System (EFFS). Hydrology
Falconer, R.A., Lin, B., Waller, S., Crossley, A.J., Mason, D., 2008. Benchmarking and Earth System Science 9 (4), 381–393.
2D hydraulic models for urban flood simulations. Proceedings of the Institution Schubert, J.E., Sanders, B.F., Smith, M.J., Wright, N.G., 2008. Unstructured mesh
of Civil Engineers, Water Management 161 (1), 13–30. generation and landcover-based resistance for hydrodynamic modeling of
Lamb, R., Crossley, A., Waller, S., 2009. A fast 2D floodplain inundation model. urban flooding. Advances in Water Resources 31 (12), 1603–1621.
Proceedings of the Institution of Civil Engineers, Water Management 162 (1), 1– Wilson, M.D., Bates, P.D., Alsdorf, D., Forsberg, B., Horritt, M., Melack, J., Frappart, F.,
9. Famiglietti, J., 2007. Modeling large-scale inundation of Amazonian seasonally
Leendertse, J.J., Gritton, E.C., 1971. A Water Quality Simulation Model for Well- flooded wetlands. Geophysical Research Letters 34 (Paper No. L15404).
mixed Estuaries and Coastal Seas, vol. 2. Computation Procedures. The Rand Wright, N.G., Whitlow, C., Crossley, A., 2003. Local time stepping for modeling open
Corporation, Santa Monica, Report R-708-NYC. channel flows. American Society of Civil Engineers, Journal of Hydraulic
Liang, D., Falconer, R.A., Lin, B., 2006. Improved numerical modelling of estuarine Engineering 129 (6), 455–462.
flows. Proceedings of the Institution of Civil Engineers, Maritime Engineering Zanobetti, D., Longeré, H., Preissmann, A., Cunge, J.A., 1970. Mekong delta
159 (1), 25–35. mathematical model program construction. American Society of Civil
Marks, K., Bates, P.D., 2000. Integration of high-resolution topographic data with Engineers, Journal of the Waterways and Harbors Division 96 (WW2),
floodplain flow models. Hydrological Processes 14, 2109–2122. 181–199.