Anda di halaman 1dari 13

International Journal of River Basin Management

ISSN: 1571-5124 (Print) 1814-2060 (Online) Journal homepage: http://www.tandfonline.com/loi/trbm20

Communicating uncertainty in flood inundation


mapping: a case study
Keith Beven, Rob Lamb, Dave Leedal & Neil Hunter
To cite this article: Keith Beven, Rob Lamb, Dave Leedal & Neil Hunter (2015) Communicating
uncertainty in flood inundation mapping: a case study, International Journal of River Basin
Management, 13:3, 285-295, DOI: 10.1080/15715124.2014.917318
To link to this article: http://dx.doi.org/10.1080/15715124.2014.917318

Published online: 01 Jul 2014.

Submit your article to this journal

Article views: 215

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=trbm20
Download by: [University of Engineering & Technology Taxila]

Date: 23 November 2015, At: 01:21

Intl. J. River Basin Management Vol. 13, No. 3 (September 2015), pp. 285295
# 2014 International Association for Hydro-Environment Engineering and Research

Research paper

Communicating uncertainty in ood inundation mapping: a case study


Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

KEITH BEVEN, Lancaster Environment Centre, Lancaster University, Lancaster LA1 4YQ, UK; Department of
Geosciences, Uppsala University, Uppsala, Sweden. Email: k.beven@lancaster.ac.uk (author for correspondence)
ROB LAMB, JBA Trust, South Barn, Broughton Hall, Skipton BD23 3AE, UK. Email: rob.lamb@jbatrust.org
DAVE LEEDAL, Lancaster Environment Centre, Lancaster University, Lancaster LA1 4YQ, UK.
Email: d.t.leedal@lancaster.ac.uk
NEIL HUNTER, JBA Consulting, South Barn, Broughton Hall, Skipton BD23 3AE, UK.
Email: neil.hunter@jbaconsulting.com
ABSTRACT
An important issue in taking account of uncertainty in ood inundation mapping is the communication of the meaning of the outputs from an uncertainty
analysis. In part this is because uncertainty estimation in this domain is not a simple statistical problem in that it involves knowledge uncertainties as
well as statistical (aleatory) uncertainties in most of the important sources of uncertainty (estimated upstream discharges, effective roughness coefcients, ood plain and channel geometries and infrastructure, choice of model, fragility of defences, etc.). Thus, assumptions are required associated
with the knowledge or lack of knowledge about these different sources of uncertainty. A framework has been developed in the form of a sequence of
condition trees to help dene these assumptions. Since stakeholders in the process can potentially be involved in making and recording decisions about
those assumptions the framework also serves as a means of communicating the assumptions. Recording the decisions also serves to provide an audit trail
for later evaluation of the decisions and hence the resulting analysis. Communication can also be helped in this type of spatial problem by effective
visualization techniques and a visualization tool has been developed for both a web-based service using Google MapsTM and a desktop application
using the MatlabTM numerical package.

Keywords: Flood inundation models; epistemic uncertainty; condition tree; visualization


1

uncertainties are more problematic: by denition we do not know


the true nature of these types of uncertainty. Thus, we must resort
to assumptions about the nature of epistemic uncertainties
(which is often to assume that they can be treated as if they
were aleatory in nature).
This has meant that there is a debate about the best way to
treat different sources of epistemic uncertainty. Different
choices might be made by different analysts. This also means
that it might be quite difcult to convey to stakeholders and
decision-makers what is the meaning of the outcomes of a
modelling exercise that has involved uncertainty analysis.
Faulkner et al. (2007) suggested that a translatory discourse
was required in order to ensure good communication
between analysts and stakeholders about the assumptions that
need to be made in a particular application and the consequent
meaning of the results.

The nature of uncertainty in assessing ood risk

The importance of uncertainty in environmental modelling is


gradually being appreciated. There are many different sources
of uncertainty, and their impact may be to change the way in
which models are evaluated and decisions are made (Beven
2009, 2012, Beven and Alcock 2011). Traditionally, uncertainties have been treated statistically, as if they arise from random
variability (the aleatory uncertainties). However, uncertainties
in environmental modelling may be dominated by those that
arise from lack of knowledge about boundary conditions, processes, model structures and parameter values, and in the observational data used to evaluate models. These are the epistemic or
Knightian real uncertainties (see Rougier and Beven 2013).
The characteristics of aleatory uncertainties can generally be
determined from the sampling variability of a variable. Epistemic

Received 29 October 2013. Accepted 18 April 2014.


ISSN 1571-5124 print/ISSN 1814-2060 online
http://dx.doi.org/10.1080/15715124.2014.917318
http://www.tandfonline.com
285

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

286

Keith Beven et al.

So what form should such a translatory discourse take? This


has been an issue in a number of ood risk management research
and knowledge exchange projects in the UK including the Flood
Risk Management Research Consortium (FRMRC1) and the
Catchment Change Network (CCN2) which had ood risk as
one of its focus areas. It has also been highlighted by other
areas of practice and policy, for example in England and Wales
in the Environment Agency (EA) Mapping and Modelling
Strategy documents. As yet no consistent framework exists for
dealing with epistemic uncertainties in practice, other than a
reliance on the expertise of modellers and analysts to be aware
of possible errors in modelling, combined with processes of
challenge and debate about the interpretation of modelling in
decision-making contexts (for example, through public
enquiries).
One of the most interesting sources of epistemic uncertainty
in ood risk assessment is what is likely to happen to ood
frequencies in the future as a result of either climate change or
changes in catchment characteristics. Changes in land use and
management in a catchment should be expected to have an
impact on ood runoff generation even without any future
climate change (Di Baldassarre et al. 2010a, 2010b, De Moel
et al. 2011). Such changes are necessarily subject to epistemic
uncertainties. Despite all the research and computational effort
that has gone into climate modelling activities, they involve considerable uncertainty even in reproducing past climatic statistics,
and particularly in reproducing precipitation statistics in many
parts of the world. Various downscaling techniques that incorporate stochastic sub-models, bias corrections or simple
change factor methods have been used to make the outputs of
climate models more acceptable for use in impact studies, but
future projections are still epistemically uncertain and consequently of unknown probability. This is also the case for estimating the impacts of future land-use management on ood runoff
generation. It is then possible that the proper communication
of the nature of such uncertainties might result in decisions
being made in different ways (see Wilby and Dessai 2010,
Prudhomme et al. 2010, Beven 2011).
In fact, while the detailed assessment of uncertainty associated with model predictions is relatively new, the problem of
making decisions under uncertainty is not. Decisions with
respect to future risks are always made under uncertainty
whether that is recognized explicitly or not. In ood risk
assessments and the design of defences, this has traditionally
been handled implicitly through the coded practices of using
design oods and freeboard. This is still apparent in UK
practice, and in the requirements of the EU Floods Directive,
in ood plain zoning for planning purposes. Three zones have
been dened in the UK, using deterministic outlines for the
annual exceedance probability (AEP) 0.01 and 0.001 design
events. It is understood that such events, and their footprint on
the landscape can only be estimated with signicant uncertainty
but they represent a way of handling uncertainty implicitly. The
AEP0.01 event is expected to occur only rarely; the AEP0.001

event is expected to occur only very rarely. New events


might require re-evaluation of those (really rather uncertain)
probabilities, but we would expect that from any conditioning
process as new events occur and new evidence becomes
available.
A similar concept can be extended to the consequences component of ood risk when risk is expressed as the product of the
probability of an event of a given magnitude and the assessment
of the consequences of such an event. In the UK National Flood
Risk Assessment, risk is assessed on a national basis using the
integration of event probabilities and damage calculations up to
a certain AEP to give a Risk Exceedance Probability curve. To
allow this type of analysis to be carried out at the national scale,
the uncertainties in both the hazard footprint and damage
calculations at any site have not generally been computed
explicitly or routinely, although there has been some assessments of the uncertainty of failure of ood defences based on
prior statistical assumptions. In this case the integration to the
national scale is being used to average over local uncertainties
as if those uncertainties were purely aleatory. It is also worth
noting that this denition of risk does not take into account
the vulnerability of the areas and population affected
within the ood footprint. This has epistemic aspects that are
difcult to quantify in purely monetary terms (Tapsell et al.
2002).
Here, we concentrate on the problem of mapping ood
inundation under current conditions (see also Leedal et al.
2010). This assessment of the hazard footprint for an event
with a chosen AEP already involves a variety of sources of
both aleatory and epistemic uncertainties, from natural variability in the occurrence of oods to lack of knowledge of
the real boundary conditions for a hydraulic model (or hydraulic models, since model structure, coding and implementation
are also potential sources for epistemic uncertainty). Past
research conrms that there is signicant uncertainty associated with ood inundation predictions using hydraulic
models (Aronica et al. 1998, 2002, Romanowicz and Beven,
2003, Bates et al. 2004, Pappenberger et al. 2005, 2006,
2007a, 2007b, Merz et al. 2007, Di Baldassarre et al. 2010a,
2010b, Leedal et al. 2010, Neal et al. 2013). Thus, we
should expect the estimated extent of a 0.01 or 0.001 AEP
ood to be uncertain even if current ood maps are presented
in terms of crisp map boundaries based on deterministic model
results. In what follows we discuss the nature of different
sources of uncertainty, the way in which they impact on a
ood risk assessment, and the problem of communicating the
resulting outputs to decision-makers.
An assessment of the uncertainty in model predictions will be
most important when the outcome might affect the decision that
is made (whether that decision is concerned with mitigating
risk to life, expected annual damages (EADs) or some other
measure). This will depend crucially on the nonlinearity of a vulnerability function on the predicted depths or velocities output
from the model. In the UK, this was recognized in the

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Communicating uncertainty in ood inundation mapping

287

Figure 1 Sources of uncertainty in ood inundation mapping.

recommendations of the Pitt Report that followed the major


oods in 2007 (Pitt 2008). Pitt encouraged a more explicit
recognition of uncertainty in both ood risk assessments and
real-time forecasting, in part as a means of being more
transparent in decision-making while being proportionate to
the problem at hand.
2 Uncertain components of a ood risk assessment and
their interactions

of uncertainty that might need to be considered, while Figure 2


shows how those uncertainties might interact in reproducing
any observations of ood outlines. For each source of uncertainty, and for conditioning model outputs on the 2007 oods,
decisions will need to be made about how to represent the
sources of uncertainty. We will later outline the nature of those
decisions for the case study within a source-pathways-receptor
condition tree framework.
2.2

2.1

Sources of uncertainty in ood risk mapping

In what follows we will describe the process for a particular


application concerned with assessing the uncertainty associated
with predictions of ood outlines in the River Dearne valley in
the vicinity of Burton-on-Dearne and Mexborough, Yorkshire,
UK. This was an area that was ooded in summer 2007 and
where post-ood survey observations of maximum water
levels are available. We consider the possibility of assessing
uncertainty in ood risk mapping, as conditioned on the observations from 2007. In making such an assessment there is no
shortage of sources of uncertainty. Figure 1 shows the sources

Aleatory and epistemic components of uncertainty

In this type of environmental modelling problem, the assessment


of uncertainty is not just a matter of the application of statistical
methods (Beven 2002, 2009, 2012, Beven and Young 2013,
Beven et al. 2014). This is because of the epistemic nature of
some of the most important uncertainties. Aleatory uncertainties
are often described as those due to natural variability. Aleatory
uncertainties can be treated in the form of probabilities (Table 1);
epistemic uncertainties are often treated as if they can be represented as probabilities, even imprecise probabilities, but this
might lead to overcondence in uncertainty estimation if the
structure of the epistemic uncertainty is non-stationary in space

Figure 2 Interactions between sources of uncertainty in ood inundation mapping.

288

Keith Beven et al.

Table 1 A summary of aleatory and epistemic uncertainties in ood risk mapping (after Beven et al. 2014)

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Source of uncertainty

Aleatory component

Epistemic component

Design ood magnitude

Floods occur randomly

Conveyance estimates

Based on observations with random error, or statistical


estimation based on random assumptions

Rating curve extrapolation

Often based on statistical extrapolation of observations


at lower ows

Floodplain topography
Model structure

Random survey errors

Floodplain infrastructure

Random errors in specifying positions of elements,


including elevations of ood defences

Observations used in
model calibration
Future catchment change
Future climate change
Fragility of defences
Consequences/
vulnerability

Random survey errors

Realizations of weather generators for given scenario


Random nature of failures
Random natures of losses in loss classes

or time (as it often will be). Model structural error, for example, is
an epistemic uncertainty that will generally have non-stationary
characteristics. It might then be better to choose to represent
knowledge uncertainties possibilistically or as scenarios
(Rougier and Beven 2013). Possibility theory, as developed in
Fuzzy Set theory, allows associating weights to different possible
outcomes. It also allows more exible ways of manipulating
those variables (Beven 2009).
Table 1 provides examples of aleatory and epistemic uncertainties that arise in ood risk mapping. It will be immediately
apparent that it can be difcult to distinguish between them. We
might, for example, consider the occurrence of different magnitudes of events to be the result of a natural random variability
(i.e. aleatory in nature), but we lack the knowledge to be sure
about the distribution that such events might be drawn from,
and whether that distribution should be considered to be stationary or not (i.e. epistemic in nature). Table 1 gives other examples.
For epistemic uncertainties, we can never be sure that the full
range of possibilities has been considered, because there is a
lack of knowledge about what that range might be. This reinforces
the point that it is important to convey to decision-makers the
assumptions on which a model uncertainty assessment is based
whether that is based on probabilities or possibilities.

Are oods generated by different types of events?


What frequency distribution should be used for each type of
event?
Are frequencies stationary?
Will frequencies stationary into the future?
Is channel geometry stationary over time?
Do conveyance estimates properly represent momentum
losses and scour at high discharges?
Seasonal changes in channel/oodplain vegetation?
Is oodplain infrastructure, walls, hedges, culverts, etc. taken
into account?
Is channel geometry stationary over time?
Does extrapolation properly represent changes in momentum
losses and scour at high discharges?
Correction algorithms in preparing digital terrain map?
Results depend on choice of model structure, dimensions,
discretization, and numerical approximations
How to treat storage characteristics of buildings, tall
vegetation, walls and hedges in geometry
Missing features in digital elevation model (e.g. walls, culverts)
Misinterpretation of wrack marks
Systematic survey errors
Scenario errors
Scenario errors
Expectations about failure modes and parameters
Knowledge about loss classes and vulnerability
Link between vulnerability and warnings

3 Communicating uncertainty in ood risk assessments


With so many sources of uncertainty, and particularly with the
lack of knowledge associated with dening the nature of the
epistemic uncertainties, there is clearly a problem of communicating the meaning of a set of uncertain model outcomes (in this
case an uncertain ood inundation map) to a decision-maker.
One method is to tackle this problem by developing guidelines
for good practice. This is the approach that has been taken in
the FRMRC and CCN projects (Beven et al. 2010, 2014,
Beven and Alcock 2011). The guidelines have been expressed
in terms of condition trees that frame a particular type of application. A user is guided through the process of making
decisions on assumptions about the different types of uncertainties involved. An example structure, for the assessment of ood
risk, is shown in Figure 3 where the consequence component of
the risk estimation is associated with the receptors in a sourcepathway-receptor structure. Perhaps the most important feature
of this approach is that it allows for the fact that while there
may be some default options at each level in the condition
tree, it is not possible to be prescriptive about dening epistemic uncertainties. The very act of dening the assumptions
for the different conditions can be an aid to communication

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Communicating uncertainty in ood inundation mapping

289

Figure 3 Summary of conditions in assessing uncertainty in ood risk mapping (after Beven et al. 2014).

with users of the outputs where it is possible for potential


users and stakeholders to be involved in making those
decisions. It is, however, important that the decisions made
be recorded so that they can be reviewed by any users of the
outcomes. In this way, an audit trail is provided for a particular
application as an incentive to be realistic about the assumptions
made.
For the case of assessing uncertainty in ood inundation maps
considered in this paper, the full condition tree approach is
described in Beven et al. (2014). The main sources of uncertainty
are characterized in a source-pathway-receptor framework
(Figure 3). For example, in considering the source component,
decisions must be made about design ood magnitudes and the
potential impacts of future change (climate and catchment). In
this case, it might be acceptable to decide that it is only necessary
to consider estimates of design ood magnitude under current
conditions, so that nothing is done about future change. This
should still be recorded for future reference. In other cases,
decisions might be made to simplify the analysis, or to carry
out a qualitative assessment, in order to be proportionate to the
problem under consideration. The condition tree is then providing a framework so that potential sources of uncertainty are not
overlooked, even if they are considered unimportant to a particular application. In addition to the source-pathway-receptor

uncertainty conditions, decisions about implementation of the


uncertainty analysis, interactions between sources, conditioning
on observations, collecting new observations to constrain uncertainty, and the visualization of the outputs of the analysis are also
required.

The case study of the Dearne Valley, Yorkshire

The study site (shown in Figure 1) is a reach of approximately


5 km on the oodplain of the River Dearne in Yorkshire, UK,
located immediately east of Bolton-Upon-Dearne and 2 3 km
north of Mexborough. The river is a tributary of the River
Don. Flow in the Dearne has been gauged since 1956 at
Adwick. The gauging station is one of the stations in the
HiFlows-UK network (catalogue number 27030), indicating
that it has a good quality ow record suitable for ood hydrology
studies.
The site was chosen because in addition to good ow records
there were relatively extensive high water marks recorded after
ooding in June 2007 when approximately 80 mm of rain fell
in a 16 hour period on the 25th June. A set of 87 high water
marks were located by teams of eld observers immediately
after the event and subsequently surveyed.

290

4.1

Keith Beven et al.

Completing the condition tree

A brief summary of the decisions needed for the River Dearne


case study is given below within the source-pathway-receptor
framework of Figure 3.

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

4.1.1 Decisions about source uncertainties


Flow data from the Adwick gauging site were used for the
estimation of the 1/100 AEP (100-year return period) design
ood event at the upstream boundary of the modelled ow
domain. Standard Flood Estimation Handbook (FEH) methods
were used, tting a generalized logistic distribution to the data
(Institute of Hydrology, 1999). The WINFAP-FEH software
also allows a standard error of the estimate to be calculated. In
this case the mean estimate of the 1/100 AEP event was
86.6 m3 s21 with 95% condence interval of 76.7
92.9 m3 s21. Sampling from this distribution was based on a
normal distribution. In this case it was not necessary to consider
other major tributaries where peak discharges might be correlated (but see Neal et al. 2013, for a case where this is necessary).
It was also not necessary to consider future climate or catchment
change.

4.1.4 Decision about implementation


The random variables described above were considered independent. Five hundred runs of the JFlow 2D hydrodynamic model
were made where the three uncertain parameter values were
drawn at random from their respective probability distribution
functions. After completing the model runs, it was possible to calculate the probability of inundation exceeding a chosen threshold
in each of the 10 10 m cells within the model domain.
4.1.5 Decisions about constraining uncertainty using
observations

Representing the pathways of ow on the oodplain requires a


ow routing model. In this case the depth-averaged two-dimensional model JFlow was used (Bradbrook et al. 2004, Bradbrook
2006, Lamb et al. 2009), implemented with 10 m by 10 m grid
cells, with ood plain topography being derived from a higher
resolution LIDAR survey. JFow requires an estimate of the
channel capacity before overbank ow occurs, which in the
absence of more detailed channel capacity information is taken
as the median of the annual maximum discharge distribution
(1/2 AEP event), estimated by standard FEH methods. This estimate of channel capacity was also allowed to be an uncertain
variable selected randomly between the range 12 and
54 m3 s21. JFlow also requires estimates of roughness coefcients for the oodplain. In this application a uniform random
variable was estimated to dene a global Mannings n roughness
coefcient for the model domain between upper and lower limits
of 0.05 0.2. These values were selected only with reference to
the experience of the model designer. To help focus on specic
sources of uncertainty, no account was taken of the fragility of
defences or the residual risk from defence failures in this
application.

Where observation data are available for model evaluation then


one of the most important sets of decisions that can be made is
how to assign a likelihood to each model run based on how well
those data are reproduced. This is because the evaluation can be
very important in constraining the range of uncertainty that will
generally arise from prior assumptions about the other sources of
uncertainty. It can also compensate for other sources of error that
have not been considered. This can be treated as a statistical
error problem (Romanowicz et al. 1996, Hall et al. 2011) but, as
noted earlier, many of the sources of uncertainty will not be statistical in nature, but epistemic (Beven 2009, 2012, Rougier and Beven
2013). Thus, a number of other measures for evaluating ood inundation models have been suggested in the past (Aronica et al. 1998,
Romanowicz and Beven, 2003, Bates et al. 2004, Hunter et al.
2005, Pappenberger et al. 2005, 2007a, 2007b).
As noted earlier, observations of maximum water levels after the
June 2007 oods were available for the Mexborough site. The
observation points are shown as green dots in Figure 1, together
with the predicted pattern of depths for one run of JFlow that
matched the observations to within an acceptable error. Such observations can be used to constrain the uncertainty in the inundation
predictions. A posterior likelihood function was formed for the
model parameter space conditioned on the t of the 2007 inundation
data based on the sum of absolute errors. The peak ow for this
event was estimated from rating curve information to be within
the 95% condence bound of 71.173.5 m3 s21.
Due to the complex interaction between model parameters,
and input and calibration data uncertainty; the model performed
within the range considered by the modeller to be acceptable
over a considerable subset of the parameter space. The resulting
parameter space and inow distribution can then be interrogated
and used to produce a range of inundation extents at specied
design ood return periods and levels of likelihood conditioned
on the model calibration exercise.

4.1.3 Decisions about receptors

4.1.6 Decisions about visualization of the results

In this application it was only necessary to consider uncertainty


in mapping the footprint of the ood for the chosen design discharge, given existing ood defences. Thus, decisions about
the damages that might result from overtopping or failure were
not considered. Importantly, these choices must be documented
during the reporting process.

Given the posterior likelihood of each model run, a likelihoodweighted probability of a model cell being inundated for the
design event could be calculated. For example, if a cell
became inundated in all 500 model runs then that 10 10 m
region of the oodplain would be ascribed an effectively 100%
probability of being ooded during a ood of magnitude equal

4.1.2 Decisions about pathway uncertainties

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Communicating uncertainty in ood inundation mapping

291

Figure 4 Visualized likelihoods of inundation for the AEP0.01 event for Mexborough. Map data source # 2012 Google Maps, DigitalGlobe, GeoEye,
Getmapping plc, Infoterra Ltd & Bluesky, The GeoInformation Group.

to the design event; if a cell was inundated with a cumulative


likelihood of 0.5 over all the ensemble members then the corresponding cell of the oodplain was ascribed a probability of
ooding of 50% and so on. In this way a map can be developed

of the likelihood of inundation for the 1/100 AEP event. The


results can be displayed in a visualization tool that has been developed as part of the FRMRC and CCN programs. This is shown in
Figure 4 where the predicted likelihoods of inundation are shown

Figure 5 Zoom on to Bolton-upon-Dearne Sewage Treatment Works with overlay of extreme ood outline for AEP0.01 event. Map data source
#2012 Google Maps, DigitalGlobe, GeoEye, Getmapping plc, Infoterra Ltd & Bluesky, The GeoInformation Group.

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

292

Keith Beven et al.

Figure 6 Matlab version of the visualization tool showing the likelihood distribution of depths predicted for the AEP 0.01 ood at the user-selected
point indicated by the red circle. The map image for the MATLABw version of the visualization tool for Mexborough is reproduced from the Explorer
1:25000 scale map by the permission of Ordnance Survey on behalf of the Controller of Her Majestys Stationery Ofce # Crown Copyright. Lancaster
University, Bailrigg, Lancaster. O.S. Licence No. A281220.

as colour coded likelihoods. This will differ from the more usual
ood map, which is usually based on a single run of the model
with best estimate inputs and parameters, and is referred to
here as a deterministic output (in the sense that although the
event may be dened in terms of AEP of the peak ow, the resulting ood outline is modelled deterministically). In particular, some
areas that are not ooded in the deterministic case are predicted as
having a likelihood of being ooded in the uncertain case. This is
the case for example for the Yorkshire Water Sewage Treatment
Works on the oodplain to the north of the river. This is defended
by embankments that are not overtopped in the deterministic case
but have a likelihood of being over topped in the uncertain case (as
shown in Figure 5). The tool also allows the likelihood for a
chosen depth of inundation to be displayed, or for any grid cell
in the model, a graph of the cumulative likelihood of depths of
inundation. This is shown on the Matlab desktop version of the
tool in Figure 6.
4.2

Making use of uncertain ood risk maps

The visualization tool shows that for some areas where the ood
plain is constrained by embankments or steeper natural topography the incorporation of uncertainty has little effect on the extent

of inundation, but may have an effect on the predicted depths at


those points. In other areas, where slopes are lower, the extent of
inundation is much more uncertain.
This information could be fed back into the planning process.
At present three ood plain planning zones are recognized in the
UK. Those are dened as being within the 0.01 AEP design event
outline, between the 0.01 AEP and 0.001 AEP outlines and
beyond the 0.001 AEP outline, with the outlines being dened
as crisp lines derived from deterministic model runs. There is
advantage in having such crisp lines in making planning
decisions efciently and the mapping of uncertain ood inundation boundaries would appear to undermine that process with a
danger of allowing more appeals to decisions by developers.
However, at a CCN workshop on uncertainty in ood risk maps
that was held for planners and consultants in collaboration with the
Royal Town Planning Institute, the opinion was also expressed that
the consideration of uncertainty might also allow more exible and
better decision-making. This might be interpreted in the following
way. For some proposed developments on ood plains, it might be
sensible to be more accepting of risk or averse to risk in either siting
of a development, foundation levels or level of defences. This is
equivalent to choosing a particular outline from the likelihood
scale for the uncertain ood outlines. A 50% outline would be

Communicating uncertainty in ood inundation mapping

expected to be close to the deterministic best estimate ood inundation map. A lower likelihood would indicate a more risk-averse
stance; a higher likelihood would indicate a more risk accepting
stance. Different likelihoods might be more or less appropriate
for different types of use.

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

The consequences part of a risk assessment

The information can also be fed back into a more complete risk
analysis. At present the map shows only the probability of the
hazard at the AEP0.01 level. A more complete analysis of
risk also involves the assessment of any resulting consequences.
In the estimation of EADs this also requires the integration of
both hazard and consequences over the distribution up to a
given AEP level. Consequences, in terms of damages or
other measures, will have their own uncertainties (both aleatory
and epistemic) that will also depend on the uncertainty
assessments for the patterns of depths and velocities in the
hazard footprint.
Standards for ood damage assessments in the UK are
provided by the manual for economic appraisal (often known as
the multi-coloured manual) produced by the Flood Hazard
Research Centre at Middlesex University (PenningRowsell et al. 2013). Damage assessments are based on predicted
ood depths and in the past have been generally applied purely
deterministically, though recent versions of the book give upper
and lower bounds for the damages at different depths of inundation based on the range of empirical estimates available from
past ood events that have been used to construct the synthetic
depth/damage curves for different types of property at risk.
The uncertainty in the damage estimates, of course, has a
direct impact on the EADs and cost benet analysis for any
particular ood mitigation scheme. The uncertainties, which
should include the possibility of future land-use patterns over
the lifetime of a project, might be large and have been generally
thought to be too large to present to decision-makers, especially
if the resulting cost benet ratio might range from less than
that required to justify a project to much greater
(Penning-Rowsell 2014). Thus, an institutionalized approach
to uncertainty has been taken, dened by the rules with
which a deterministic cost benet analysis is carried out
(Environment Agency 2010).
As with the denition of crisp ood risk zones, this institutionalization of the damage calculations is one, rather convenient,
way of dealing with uncertainty. In that the cost and benets
should be prepared in standardized ways for different proposed
schemes, then the ranking of the schemes in deciding on investments can be presented as an objective way of deciding on
investment priorities.
The question therefore is whether a more complete uncertainty assessment, which would be more expensive for every
project proposed, might result in better decision-making. This
perhaps can already be seen from Figures 4 and 6 where,

293

without consideration of uncertainty, some areas would be


dened as not at risk. Less obvious is the effect of the uncertainty
on depths of inundation, as shown in the inset in Figure 6. Where
the dependence of damages on depths (and perhaps predicted
velocities) of ow is highly nonlinear, then the uncertainty in
depths might have a signicant effect on the EADs when integrated over the uncertainty (Penning-Rowsell 2014). This will
particularly be the case where, for example, the consequences
of ooding of infrastructure, previously thought not to be at
risk to the design level, will be high (as in Figure 5).

Conclusions

Figures 1 and 2 reveal the many sources of uncertainty, and their


interactions, that are involved in producing ood inundation
maps. Not all of those sources of uncertainty are easily quantied
a priori because of lack of information. They involve epistemic
as well as aleatory uncertainties, as summarized in Table 1.
This means that subjective decisions must be made in any evaluation of this type, including when comparing model outputs with
observations in constraining the prediction uncertainties. This
can be seen, for example, in the changes in posterior roughness
parameter distributions that arise from calibrating against observations for events of different magnitudes (Romanowicz and
Beven 2003). This may be the result of a physical change in
roughness with discharge, but might also be compensating for
a lack of knowledge of other uncertain factors (including any
unknown unknowns that have been left out of the analysis).
That is why a framework for making decisions about how to
treat different sources of uncertainty can be useful in both providing guidance based on past experience, and in providing an audit
trail for a later review. It can also serve as a basis for providing
the translatory discourse between analysts and stakeholders in
explaining the signicance of the output uncertainties.
Is it worth the cost of the additional analysis required relative to
making deterministic predictions of the 0.01 and 0.001 AEP
oods? As noted earlier, the choice of those design exceedance
probabilities is already an implicit way of compensating for uncertainties in dening rare and very rare oods. Further account of
uncertainty is also taken into account implicitly in the design of
ood defence levels by standard procedures for estimating freeboard (Kirby and Ash 2000) which can also be assessed with
uncertainty (Brandimarte and Di Baldassarre 2012). Attempts to
quantify the impact of uncertainty more explicitly will certainly
be more scientically honest, but a full analysis may not be proportionate for all applications. It will, however, usually reveal
risks (such as for the Sewage Treatment Works at Mexborough
in Figure 5) that it might be valuable to address where the consequences might be sufciently high. Thus, a consideration of uncertainty might change the decision that is made towards providing
greater protection in some cases where it might be important to
be risk averse. In such cases, the relative cost of the additional
effort might easily be justied.

294

Keith Beven et al.

Acknowledgements
Jeff Neal is thanked for carrying out the model likelihood evaluations.
We are grateful to the EA for agreement to make use of the case study
data.

Funding

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

This work was funded by the UK Flood Risk Management Research


Consortium, and the NERC Catchment Change Network Knowledge
Transfer and CREDIBLE projects.

Notes
1. www.oodrisk.org.uk
2. www.catchmentchange.net

References
Aronica, G., Bates, P.D., and Horritt, M.S., 2002. Assessing the
uncertainty in distributed model predictions using observed
binary pattern information within GLUE. Hydrological
Processes, 16 (10), 2001 2016.
Aronica, G., Hankin, B.G., and Beven, K.J., 1998. Uncertainty
and equinality in calibrating distributed roughness coefcients in a ood propagation model with limited data.
Advances in Water Resources, 22 (4), 349 365.
Bates, P.D., et al., 2004. Bayesian updating of ood inundation
likelihoods conditioned on ood extent data. Hydrological
Processes, 18 (17), 3347 3370.
Beven, K.J., 2002. Towards a coherent philosophy for environmental modeling. Proceedings of the Royal Society of
London A, 458 (2026), 2465 2484.
Beven, K.J., 2009. Environmental modelling: an uncertain
future? London: Routledge.
Beven, K.J., 2011. Distributed models and uncertainty in ood
risk management. In: G. Pender and H. Faulkner, eds. Flood
risk science and management. Chichester: Wiley-Blackwell,
291 312.
Beven, K.J., 2012. Causal models as multiple working hypotheses about environmental processes. Comptes Rendus
Geoscience. doi:10.1016/j.crte.2012.01.005
Beven, K.J. and Alcock, R., 2011. Modelling everything everywhere: a new approach to decision making for water management under uncertainty. Freshwater Biology, 56. doi:10.1111/
j.1365-2427.2011.02592.x
Beven, K. and Young, P., 2013. A guide to good practice in
modeling semantics for authors and referees. Water Resources
Research, 49 (8), 5092 5098. doi:10.1002/wrcr.20393
Beven, K.J., Leedal, D.T., and Alcock, R., 2010. Uncertainty
and good practice in hydrological prediction. Vatten, 66 (3),
159 163.
Beven, K., Leedal, D., and McCarthy, S., 2014. Framework for
assessing uncertainty in uvial ood risk mapping. CIRIA

C721. London: Construction Industry Research and


Information Association.
Bradbrook, K., 2006. JFLOW: a multiscale two-dimensional
dynamic ood model. CIWEM Water and Environment
Journal, 20 (2), 79 86.
Bradbrook, K.F., et al., 2004. Two dimensional diffusion wave
modelling of ood inundation using a simplied channel
representation. International Journal of River Basin Management, 2 (3), 211 223.
Brandimarte, L. and Di Baldassarre, G., 2012. Uncertainty in
design ood proles derived by hydraulic modelling. Hydrology Research, 43 (6), 753 761.
De Moel, H., Aerts, J.C.J.H., and Koomen, E., 2011.
Development of ood exposure in the Netherlands during
the 20th and 21st Century. Global Environmental Change,
21 (2), 620 627.
Di Baldassarre, G., et al., 2010a. Floodplain mapping: a critical
discussion on deterministic and probabilistic approaches.
Hydrological Sciences Journal, 55 (3), 364 376.
Di Baldassarre, G., et al., 2010b. Flood fatalities in Africa: from
diagnosis to mitigation. Geophysical Research Letters, 37,
L22402. doi:10.1029/2010GL045467
Environment Agency, 2010. Flood and coastal erosion risk management appraisal guidance (FCERM-AG) [online]. Bristol:
Environment Agency. Available from: https://www.gov.uk/
ood-and-coastal-defence-appraisal-of-projects [Accessed 23
May 2014].
Faulkner, H., et al., 2007. Developing a translational discourse to
communicate uncertainty in ood risk between science and the
practitioner. Ambio, 16 (7), 692 703.
Hall, J.W., Manning, L.S., and Hankin, R.K.S., 2011. Bayesian
calibration of a ood inundation model using spatial data.
Water Resources Research, 47, W05529. doi:10.1029/
2009WR008541
Hunter, N.M., et al., 2005. Utility of different data types for
calibrating ood inundation within a GLUE framework.
Hydrology and Earth System Sciences, 9 (4), 412 430.
Institute of Hydrology. 1999. Flood estimation handbook
(5 volumes). Wallingford: Institute of Hydrology.
Kirby, A.M., and Ash, J.R.V., 2000. Fluvial freeboard guidance
note. Bristol: Environment Agency (Environment Agency
R&D Technical Report W/187).
Lamb, R., Crossley, A., and Waller, S., 2009. A fast twodimensional oodplain inundation model. Proceedings of
the ICE Water Management, 162 (6), 363 370.
Leedal, D.T., et al., 2010. Visualization approaches for communicating real-time ood forecasting level and inundation
information. Journal of Flood Risk Management, 3 (2),
140 150.
Merz, B., Thieken, A.H., and Gocht, M., 2007. Flood risk
mapping at the local scale: concepts and challenges.
Flood risk management in Europe: innovation in policy
and practice, 13, 231 251, Dordrecht, The Netherlands:
Springer.

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Communicating uncertainty in ood inundation mapping

Neal, J., et al., 2013. Probabilistic ood risk mapping including


spatial dependence. Hydrological Processes, 27, 1349 1363.
doi:10.1002/hyp.9572
Pappenberger, F., et al., 2005. Uncertainty in the calibration of
effective roughness parameters in HEC-RAS using inundation
and downstream level observations. Journal of Hydrology,
302 (1), 46 69.
Pappenberger, F., et al., 2006. Inuence of uncertain boundary
conditions and model structure on ood inundation predictions. Advances in Water Resources, 29 (10), 1430 1449.
doi:10.1016/j.advwatres.2005.11.012
Pappenberger, F., et al., 2007a. Grasping the unavoidable subjectivity in calibration of ood inundation models: a vulnerability
weighted approach. Journal of Hydrology, 333 (2), 275 287.
Pappenberger, F., et al., 2007b. Fuzzy set approach to calibrating
distributed ood inundation models using remote sensing
observations. Hydrology and Earth System Sciences, 11 (2),
739 752.
Penning-Rowsell, E.C., 2014. A realistic assessment of uvial
and coastal ood risk in England and Wales. Transactions of
the Institute of British Geographers. doi:10.1111/tran.12053
Penning-Rowsell, E., et al., 2013. Flood and coastal erosion risk
management: a manual for economic appraisal. Routledge,
448 pp.

295

Pitt, M., 2008. Learning lessons from the 2007 oods. London:
Cabinet Ofce.
Prudhomme, C., et al., 2010. Scenario-neutral approach to
climate change impact studies: application to ood risk.
Journal of Hydrology, 390 (3), 198 209.
Romanowicz, R. and Beven, K.J., 2003. Bayesian estimation of
ood inundation probabilities as conditioned on event inundation maps. Water Resources Research, 39 (3), W01073.
doi:10.1029/2001WR001056
Romanowicz, R., Beven, K.J., and Tawn, J., 1996. Bayesian
calibration of ood inundation models. In: M.G. Anderson,
D.E. Walling, and P.D. Bates, eds. Floodplain processes.
Chichester: Wiley, 333 360.
Rougier, J. and Beven, K.J., 2013. Model limitations: the sources
and implications of epistemic uncertainty. In: J. Rougier, S.
Sparks, and L. Hill, eds. Risk and uncertainty assessment
for natural hazards. Cambridge: Cambridge University
Press, 40 63.
Tapsell, S.M., et al., 2002. Vulnerability to ooding:
health and social dimensions. Philosophical Transactions
of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 360 (1796), 1511 1525.
Wilby, R.L. and Dessai, S., 2010. Robust adaptation to climate
change. Weather, 65 (7), 180 185.

Downloaded by [University of Engineering & Technology Taxila] at 01:21 23 November 2015

Anda mungkin juga menyukai