You are on page 1of 15

International Journal of River Basin Management

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

A framework for uncertainty analysis in flood risk


management decisions

Jim Hall & Dimitri Solomatine

To cite this article: Jim Hall & Dimitri Solomatine (2008) A framework for uncertainty analysis in
flood risk management decisions, International Journal of River Basin Management, 6:2, 85-98,
DOI: 10.1080/15715124.2008.9635339

To link to this article: https://doi.org/10.1080/15715124.2008.9635339

Published online: 23 Aug 2010.

Submit your article to this journal

Article views: 4482

View related articles

Citing articles: 33 View citing articles

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


https://www.tandfonline.com/action/journalInformation?journalCode=trbm20
i i

“paper01” — 2008/5/26 — 20:07 — page 85 — #1


i i

Intl. J. River Basin Management Vol. 6, No. 2 (2008), pp. 85–98


© 2008 IAHR, INBO & IAHS

A framework for uncertainty analysis in flood risk management decisions


JIM HALL, Tyndall Centre for Climate Change Research, School of Civil Engineering and Geosciences,
Newcastle University, UK. E-mail: Jim.Hall@ncl.ac.uk

DIMITRI SOLOMATINE, UNESCO-IHE Institute for Water Education, Delft, Netherlands.


E-mail: D.Solomatine@unesco-ihe.org

ABSTRACT
Modern flood risk management involves responding sustainably to flood risk with portfolios of structural and non-structural measures. Under these
circumstances of multi-attribute choice between portfolios of options, the motivation for uncertainty analysis becomes more compelling than ever.
Uncertainty analysis is required in order to understand the implications for decision makers of limited data, model uncertainties, changes in the flooding
system over the long term, incommensurate scales of appraisal and potentially conflicting decision objectives. In recognition of the importance of
uncertainty analysis as an integral aspect of sustainable flood risk management, a new framework for uncertainty analysis within flood risk management
decisions has been established. The proliferation of methods for uncertainty analysis can be placed within the coherent framework. As well as
estimating the amount of uncertainty associated with key decision variables, the framework supports the decision making process by identifying the
most influential sources of uncertainty, and the implications of uncertainty for the preference ordering between options. The challenges posed by severe
uncertainty about the potential for long term changes in the flooding system are discussed and robustness analysis is advocated in response to these
uncertainties.

Keywords: Uncertainty; risk; decision making; flood risk management.

1 The requirement for uncertainty analysis in modern flooding (for example land use changes in upstream catchments)
flood risk management or reduce the consequence of flooding when it does occur, by
reducing vulnerability. The criteria for assessment of flood risk
A major shift in approaches to management of flooding is now management options are seldom solely economic, but involve
underway in many countries worldwide. This shift has been stim- considerations of public safety, equity and the environment. Fur-
ulated by severe floods, for example on the Oder (1997), Yangtze thermore, an increasing recognition of non-stationarity means
(1998), Elbe (2002), Rhône (2003), in New Orleans (2005) and that flood risk management involves explicit consideration of the
on the Danube (2006), and by a recognition of the relentless ways in which flood risk may change in future, due, for exam-
upward trend in vulnerability to flooding world-wide (Munich ple, to climate change or floodplain development. This leads to
Re Group, 2007), as well as the potential impacts of climate the notion of flood risk management being a continuous pro-
change on flood frequency. cess of adaptive management, which raises new challenges for
Whilst recent floods have been a stimulus for changing flood uncertainty analysis (Hutter and Schanze, 2008).
risk management practice, the notion of integrated risk-based There is seldom a single solution to managing flood risk.
approach to flood management is in fact well established (Hall Instead, portfolios of flood risk management measures, be they
et al., 2003; National Academy of Engineering, 2000; National “hard” structural measures such as construction of dikes, or “soft”
Research Council, 2000; Sayers et al., 2002) and methods for instruments such as land use planning and flood warning systems,
probabilistic risk analysis have been used for some years in are assembled in order to reduce risk in an efficient and sus-
the narrower context of flood defence engineering (CUR/TAW, tainable way. The makeup of flood risk management portfolios
1990; Goldman, 1997; USACE, 1996; Vrijling, 1993). Indeed is matched to the functioning and needs of particular locali-
the notion of risk-based optimisation of the costs and benefits ties and will be adapted as more knowledge is acquired and as
of flood defence was laid out in van Dantzig’s (1956) seminal systems change. Implementing this approach involves the col-
analysis. However, modern flood risk management does not rely lective action of a range of different government authorities and
solely upon engineered flood defence structures, such as dikes, stakeholders from outside government. This places an increas-
channel improvement works and barriers, but also considers a ing emphasis upon effective communication and mechanisms
host of measures that may be used to reduce the severity of to reach consensus. Some features of this approach are now

Received on January 14, 2008. Accepted on January 18, 2008.

85

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 86 — #2


i i

86 Jim Hall and Dimitri Solomatine

becoming embedded in national government policy, for exam- options appraisal and design studies. The additional require-
ple in the UK policy document Making Space for Water (Defra, ments for analysis and computation are rapidly being (more than)
2005), the European Directive on the Assessment and Manage- compensated for by the availability of enhanced computer pro-
ment of Flood Risks, which was adopted by the European Council cessing power. However, computer processing power is only
on 18 September 2007, and progressive evolution of floodplain part of the solution, which also requires a step change in the
management in the USA (Galloway, 2005; Interagency Flood- approach to managing data and integrating the software for uncer-
plain Management Review Committee, 1994; Kahan, 2006). tainty calculations (Harvey et al., 2008). The data necessary for
Compelling as modern integrated flood risk management cer- quantified uncertainty analysis are not always available, so new
tainly is, it brings with it considerable complexity. The risk-based data collection campaigns (perhaps including time-consuming
approach involves analysing the likely impacts of flooding under expert elicitation exercises) may need to be commissioned.
a very wide range of conditions. As the systems under consider- Project funders need to be convinced of the merits of uncer-
ation expand in scope and timescale, so too does the number tainty analysis before they invest in the time and data collection it
of potentially uncertain variables. There are many potential requires.
components to a portfolio of hard and soft flood risk manage- It is not always clear how uncertainty analysis will contribute
ment measures and they can be implemented in many different to improved decision making. Much of the academic literature on
sequences through time, so the decision space is potentially huge. hydrological uncertainties (Liu and Gupta, 2007) has tended to
Communicating risks and building the consensus that is necessary focus upon forecasting problems. Providing uncertainty bounds
to engage effectively with stakeholders in the flood risk man- on a flood forecast may be intriguing, but to be meaningful needs
agement process requires special aptitude for communication, to be set within the context of a well defined decision problem
facilitation and mediation. (Frieser et al., 2005; Todini, 2008).
Uncertainty has always been inherent in flood defence engi- There is a bewildering proliferation of uncertainty meth-
neering. Traditionally it was treated implicitly through conserva- ods published in the academic literature. Pappenberger et al.
tive design equations or through rules of thumb, for example by (2006) provide a decision tree to find the appropriate method
introduction of freeboard allowances. The introduction of risk- for a given situation. Here our aim is different, in seeking to
based approaches (CUR/TAW, 1990; Meadowcroft et al., 1997; establish a framework for incorporating the main aspects of
USACE, 1996) enabled more rational treatment of natural vari- uncertainty analysis in flood risk management decision mak-
ability in loads and responses. It also paved the way for more ing. We establish this framework based upon a restatement, in
explicit treatment of uncertainty in the evidence that is used to the following section, of the conventional probabilistic represen-
support risk-based decision making. Explicit uncertainty analysis tation of a decision problem. Sources of uncertainty in flood
provides a means of analysing the robustness of flood risk man- risk management decisions are reviewed in section 3, before
agement decisions as well as the basis for targeting investment addressing the problem of uncertainty representation in section
in data collection and analysis activities that make the greatest 4. The framework for an uncertainty analysis process is pre-
possible contribution to reducing uncertainty. sented and explained in section 5. The implications and benefits
Increasingly governments are requiring a careful considera- of adopting the proposed approach are discussed in the conclud-
tion of uncertainty in major planning and investment decisions. ing section 6. Examples of methods and approaches that may
For example, USWRC (1983) (quoted in Al-Futaisi and Ste- populate this framework are provided in the remainder of this
dinger (1999)) state that: “Planners shall identify areas of risk Special Issue.
and uncertainty in their analysis and describe them clearly, so
that decisions can be made with knowledge of the degree of reli-
ability of the estimated benefits and costs and of the effectiveness 2 A brief review of risk-based decision making
of alternative plans.” The UK Department of the Environment,
Food and Rural Affairs’ (Defra) guidance on flood and coastal Modern flood risk management involves decision making at a
defence repeatedly calls for proper consideration of uncertainty range of spatial and temporal scales (Hall et al., 2003). This
in appraisal decisions. Guidance document FCDPAG1 (Defra, may be thought of as a hierarchy of decision making processes in
2001) on “Good decision making” states: “Good decisions are which high level policies at a continental, national and river basin
most likely to result from considering all economic, environ- scale provide the framework for more specific plans, which in turn
mental and technical issues for a full range of options, together provide the constraints and objectives for designs of specific flood
with a proper consideration of risk and uncertainty.” As Pate risk reduction measures and schemes. Operation of schemes, for
Cornell (1996) states, in the context of quantified risk analysis example flood warning systems, will involve an ongoing process
“Decision makers may need and/or ask for a full display of the of decision making in the light of monitoring information. Our
magnitudes and the sources of uncertainties before making an emphasis in this paper is upon the deliberative decision mak-
informed judgment.” ing associated with flood risk management plans, strategies and
However, the practice of uncertainty analysis and use of the designs rather than with the real time decision making associated
results of such analysis in decision making is not widespread, for with flood warning, or with the political process by which the
several reasons (Pappenberger and Beven, 2006). Uncertainty policies that set the overall aims of flood risk management are
analysis takes time, so adds to the cost of risk analysis, formulated.

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 87 — #3


i i

A framework for uncertainty analysis in flood risk management decisions 87

2.1 Formal decision problem question of representation of this, latter, severe uncertainty, par-
ticularly in relation to scenarios of long term future change, is
In order to situate uncertainty analysis within the decision making
dealt with in section 3.
process, we briefly review conventional decision theory. The aim
of this is three-fold: first to demonstrate the significance of uncer-
tainty through the potential for reversal of preference orderings; 2.2 Formal decision problem applied to flood risk management
secondly to note that probability is not necessarily the only repre-
In the context of flood risk management the acts d1 , . . . , dn
sentation of uncertainty in a conventional decision problem; and
are flood risk management options. They may be portfolios of
thirdly to highlight the assumptions in the conventional decision
options, i.e., different combinations of some set of basic ele-
problem that contain potential sources of uncertainty.
ments, and they may differ from one another in the sequence,
Conventionally there is a set of decision options or “acts”
though time, in which they are implemented. The decision prob-
{d1 , . . . , dn } and a set of future states of nature {θ1 , . . . , θm },
lem may involve a continuous design variable, such as the crest
defined on some space , that may materialise after the choice.
level of a dike, so the decision problem may be continuous rather
Depending on which state of nature in fact materialises, act di
than discrete.
will yield one of m possible outcomes yi,1 , . . . , yi,m (e.g., “no
The future states θ1 , . . . , θm are conventionally thought of as
flood” or a flood of a given severity). The problem of valuing
dealing at least with the unpredictable loads in nature to which
outcomes yi,1 , . . . , yi,m is a fundamental one, to which we will
flooding systems are subject e.g., fluvial flows, water levels and
return, but for the time being suppose that the net value (including
wave heights. These will seldom be discrete but will typically
both costs and benefits) associated with a given decision outcome
extend over a continuous multi-dimensional space and the state of
yi,j can be written as a scalar function v(yi,j ), then the following
the flooding system will be described by a vector of k continuous
scenarios were first identified by Knight (1921):
variables x := (x1 , . . . , xk ) : x ∈ (R+ )k . In the case of decision
(i) Decision making under certainty: The state of nature after making under risk the discrete probability distribution p(θj ) :

the decision is known i.e., m = 1. The decision maker j = 1, . . . , m :  mj=1 p(θj ) = 1 is replaced by a continuous

chooses the option with the highest value v(yi,1 ). distribution f(x) : 0 f(x)dx = 1 and the summation in Eq. (1)
(ii) Decision making under risk: Only the probabilities p(θj ) : is replaced by an integral:
m
j = 1, . . . , m : j=1 p(θj ) = 1 of occurrence of set  ∞
of states of nature {θ1 , . . . , θm } are known. Provided the v(yi (x))f(x)dx (2)
0
decision maker accepts a set of consistency and continuity
axioms (Savage, 1954) and is neutral in their attitude to risk where we now have to explicitly acknowledge that yi is a func-
then he or she should choose the option that maximises: tion of x. It is not uncommon to require a combination of discrete
and continuous variables in order to describe the state of a flood-

m
ing system comprised of multiple components (see for example
v(yi,j )p(θj ) (1)
j=1
Dawson and Hall (2006)). Here, for clarity, we will now deal
with continuous variables only.
(iii) Decision making under uncertainty: There is no informa- Typically the quantities in yi (x) will extend though time, so
tion about the probabilities of states of nature θ1 , . . . , θm . it is necessary to establish a method of aggregating a stream of
Under these circumstances there are various decision strate- annual payments or losses yi,t (x) : t = 0, . . . , T , where t denotes
gies that are in some sense rational, for example maximin the year in which the cost or risk is incurred and T is the time
utility, minimax regret, Hurwicz α, and that based on horizon. Customarily this is done by discounting to a Present
Laplace’s principle of insufficient reason (French, 1988). Value.
The assumption of risk neutrality, in which a decision maker’s Without loss of generality the expression in Eq. (2) may be
utility function u(yi,j ) is equal to the value function v(yi,1 ) is separated into terms that represent PresentValue costs and Present
made throughout this paper. The assumption of risk neutrality is Value risks. In any given year, t, the risk ri,t is given by
often advocated for government decisions (Ball and Floyd, 1998;  ∞
USWRC, 1983), though public safety decisions often illustrate ri,t = D(xt )f(xt )dxt (3)
0
aversion to low probability, high consequence events (Pasman
where D(xt ) is a damage function and we have introduced the
and Vrijling, 2003). The extension from risk neutrality to other
subscript t to signify that in general we expect x to change with
utility function is in principle straightforward (French, 1988),
time. The simplicity of Eq. (3) belies the potential complexity of
though in practice it requires elicitation of the decision maker’s
the underlying calculations in practice, which have been exten-
utilities.
sively explored elsewhere (Beard, 1997; Dawson and Hall, 2006;
Knight’s formalisation of the decision problem implicitly dis-
Stedinger, 1997). In order to estimate flood risks it is necessary
tinguishes between two types of uncertainty. Case (ii), which
to be able to:
Knight referred to as “decision making under risk”, requires a
probability distribution over the future states of nature θ1 , . . . , θm (1) Estimate probability distributions, f(xt ), for the sources of
whilst Case (iii), “decision making under uncertainty”, acknowl- flooding i.e., loading variables including extreme rainfall,
edges that this probability distribution may not be known. The water levels, marine surge tides and waves.

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 88 — #4


i i

88 Jim Hall and Dimitri Solomatine

(2) Relate given values of loading variables to probabilities of Whilst the decision problem in section 2 was introduced in
flooding at locations where flooding may cause damage. This the context of a scalar value function, practical flood risk man-
may involve hydrological and hydraulic modelling as well agement decisions are fraught with valuation problems (Egorova
as analysis of the reliability of flood defence structures and et al., 2008; Finkel, 1990). Whilst the primary decision criterion
pumping stations, and the operation of reservoirs. In urban may be based upon comparing the economic value of risk avoided
areas flood risk analysis will involve analysis the effects of with the costs of implementing the given decision option, there
the sewer network and pumped systems as potentially major will almost invariably be other significant criteria in the decision
modifier of flooding behaviour, as well as analysing overland problem. Flood risk management decisions are motivated by the
flows. desire to improve public safety, and reduce the health and social
(3) Calculate the damage that is caused by floods of a given impacts of flooding. Engineering works on rivers and coasts
severity. will usually involve environmental costs and benefits. Uncertain-
ties associated with valuation enter the decision problem either
Steps 2 and 3 are together contained in D(xt ). These three steps
if an economic valuation approach is adopted for dealing with
typically involve sequences of models and analysis processes. For
these non-market risks and costs, or if an explicit multi-criteria
systems with significant time-dependency at the sub-annual scale
approach is adopted. In the former, these originate in the prices
(for example hydrological response of catchments), accurate sim-
assigned to non-market goods and services, whilst in the latter
ulation of flooding will involve additional explicit treatment of
the uncertainties are associated with the value functions used to
the temporal dimension.
transform (uncertain) predicted outcomes to aggregate utilities.
The Present Value risk PV(ri ) is:
If the estimates of r0 , ri or ci are uncertain then the preference

T
ri,t ordering between options could be switched. Uncertainty is of
PV(ri ) = (4) relevance to decision makers because of its potential influence
t=0
(1 + q)t
on preference orderings. In the case of continuously variable
where q is the discount rate. The Present Value cost PV(ci ) is options (e.g., the crest height of a flood defence) any variation in
defined similarly. the risk or cost estimates will alter the choice of design variable.
In the case of purely economic decision making, com- The joint probability density function f(xt ) in Eq. (3) already
monplace decision criteria are Net Present Value (NPV) and represents random variability. In calculating risks we acknowl-
Benefit-Cost Ratio (BCR) in which case it is necessary to intro- edge that quantities of relevance to decision making vary. In
duce the notion of a “base case” that involves no investment. flood risk analysis f(xt ) is, as a minimum, used to represent ran-
The Present Value risk corresponding to the base case is r0 and dom variation in loading variables such as water levels and wave
expected to be unacceptably high, which is why investment in heights that vary naturally through time. It is usually extended
risk reduction is being contemplated. The Net Present Value to include natural variability in space of variables such as soil
NPVi = PV(r0 )−PV(ri )−PV(ci ), whilst the Benefit-Cost Ratio properties (National Research Council, 2000). Natural variabil-
is BCRi = (PV(r0 ) − PV(ri ))/PV(ci ). If the preference order- ity (which is sometimes also referred to as “inherent uncertainty”
ing between risk reduction options is established on the basis of or “aleatory uncertainty”) is thought of as a feature of nature and
NPV then if NPVi > NPVj > NPVl , the preference ordering is cannot be reduced (Vrijling and Van Gelder, 2006).
denoted i  j  l, and similarly for BCR. Epistemic uncertainties (knowledge uncertainties) require
more careful consideration, and are the main focus of attention
in this paper. Epistemic uncertainties include model uncertain-
3 Sources of uncertainty in flood risk ties and statistical uncertainties due to observation error and small
management decisions samples of random phenomena. Model uncertainties mean that
the function D(xt ) is uncertain. Statistical uncertainties mean
The familiar theoretical framework reviewed in the previous sec- that the function f(xt ) may not be an accurate description of the
tion depends upon a number of assumptions. Typically there are variation in xt (National Research Council, 1999). This may be
uncertainties in: due to limitations in the number of statistical samples, ambiguity
in the choice of potential statistical models, or inappropriate sta-
(1) the system characterisation in terms of k variables in xt ;
tistical assumptions, such as statistical stationarity through time.
(2) the specification of the joint probability density function
We have acknowledged that many of the quantities of interest
f(xt ), which describes the variation in xt ;
in a flood risk calculation will change in a systematic way over
(3) the function D(xt ) which embodies all of the modelling to
extended timescales, and this extends to the statistical properties
relate given values of variables xt to flood damage, as well
of f(xt ), for example due to non-stationary climate. The nature of
as the problem of damage valuation (including non-market
that change will also be uncertain due to epistemic uncertainties.
goods and services);
By way of example, Table 1 sets out the main uncertain vari-
(4) the cost ci ; and
ables in a flood risk analysis for a flood risk analysis in the Thames
(5) the choice of discount rate q.
estuary in the UK (Newcastle University and Halcrow, 2006).
There may also be uncertainties due to numerical approximations Most of the uncertain quantities listed in Table 1 are potentially
in the integration in Eq. (3). subject to change in the future. Following the approach adopted

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 89 — #5


i i

A framework for uncertainty analysis in flood risk management decisions 89

Table 1 Summary of uncertainties in the Thames Estuary flood risk analysis.

Variable or function Sources of uncertainty Driver of change in the future

Variable: Water level at outer boundary of Statistical uncertainties in estimating extreme Mean sea level rise
estuary water levels Changing storm surge frequency
Variable: Shape of tide/surge Single representative surge used in analysis Changing storm surge duration
Variable: Discharge at upstream fluvial Statistical and hydrological uncertainty in Changing rainfall
boundary of estuary estimating discharge Changing land use and runoff
Function: Water level at dikes in the estuary Tidal/river model uncertainty Natural and anthropogenic morphological
changes in estuary
Variable: Dike crest level Scarcity/accuracy of measurements. Settlement
Variable: Probability of breaching of dikes Scarcity of information about defences Deterioration
(conditional probability distribution) Limitations of quantified knowledge of failure
modes
Variables: Breach location, Breach width, Scarcity of information about defence materials None
Rate of breach growth Limitations of quantified knowledge of breaching
processes
Variable: Probability of failure to close Uncertain reliability of mechanical and human Technological changes enabling improved
barriers and gates systems reliability in future
Variable: Digital Elevation Model (DEM) Effective elevations at model grid-cell scale Land use change, especially building size and
location
Variable: Floodplain roughness Effective parameterisation of surface roughness Land use changes that influence surface
roughness
Function: Water level at points in the Model grid size
floodplain Roughness parameterisation
Representation of buildings and other
obstructions in the floodplain
Hydraulic model solution method
Variable: Location and type of properties Classification and aggregation of properties Land use, demographic and behavioural
and people in the floodplain Census data and numbers of commuting people changes
Variable: Property threshold levels Threshold levels not same as DEM levels Potential use of flood proofing
Function: Depth-damage relationships Aggregation uncertainties for property types Changing wealth and household contents
Regional and local variations
Function: probability of loss of life and Vulnerability of people Changing public vulnerability with different
other harms to people in the floodplain Velocity, arrival time and other hydraulic socio-economic scenarios
properties of flood.
Variable: reduction factors for warning, Effectiveness of warning, evacuation etc. Changing public behaviour with different
evacuation etc. socio-economic scenarios

by Evans et al. (2004), one can consider the potential impact potential approaches:
of the drivers of change, which are listed in the right hand col-
umn of Table 1. A brief summary of the main uncertainties in (1) In the Bayesian approach any variable that is thought of
future changes and their implications is provided in Table 2. These as being uncertain for any reason (natural or epistemic) is
two tables illustrate the scope of variables that may need to be incorporated in xt and the uncertainty about it encoded in
incorporated in a comprehensive uncertainty analysis. f(xt ) (Apostolakis, 1990). If necessary the damage function
D and the cost ci can be parameterised and the associated
parameter uncertainties incorporated in f(xt ). With these new
4 Representation of uncertainty in the decision problem
uncertainties incorporated, the calculation of utilities can be
repeated for all decision options which may, or may not,
4.1 Representing epistemic uncertainties
yield a new preference ordering in the light of the new infor-
In the previous section we recognised that natural variability in mation about uncertainties. In other words all uncertainties
the flooding system is represented in the joint probability density are incorporated in f(xt ) and then “integrated out”, and in so
function f(xt ). The outstanding question is how epistemic uncer- doing may alter the preference ordering based upon expected
tainties are to be represented in Eq. (3). There are broadly two utilities. This is the approach described by Kanning and van

i i

i i
i
i

i
i

Table 2 Uncertainties in future changes in flood risk in the Thames Estuary. 90

Variable Nature of future change in variable Main uncertainties in predicted future changes

Relative sea level Changes in mean sea level due to climate change. Uplift/subsidence Predictions of mean and local sea level rise
of land Future greenhouse gas emissions
Waves Changes in wave height and direction GCM predictions of wind. Downscaled wave modelling
Future greenhouse gas emissions
Surges Changing surge heights GCM predictions of wind. Downscaled surge modelling
Future greenhouse gas emissions
Precipitation Changes in quantity, spatial distribution of rainfall and intensity. GCM predictions of rainfall. Downscaling to Thames catchment
Jim Hall and Dimitri Solomatine

Rain/snow proportion Future greenhouse gas emissions


Runoff and river conveyance Changes in land surface (e.g., construction of impermeable surfaces Changes in land surface cover in the Thames catchment
and storm water drainage systems) Future engineering works in the catchment that may alter
Changes (including engineering changes) in river morphology and/or downstream fluvial discharges
vegetation that influence flood storage and flood conveyance
Estuarial morphology Changes in the nearshore seabed, shoreline and estuaries. May be the Morphological evolution in the outer and inner Thames estuary,
consequence of anthropogenic activities such as dredging, including magnitude of and response to human interventions e.g.,
reclamation and coastal protection dredging
Flood defence systems Changes in the condition and reliablity of the flood defence system Rates of deterioration of the flood defence system e.g., material
(including fixed and moveable components) deterioration, settlement of embankments and deterioration of
mechanical systems
Buildings and contents Changes in the cost of flood damage to domestic, commerical and Rate of increase of domestic wealth
other buldings and their contents Vulnerability of domestic and commerical buildings and contents
Urban vulnerability Changes in the number and distribution of domestic, commerical and Current and future planning policies
other buildings in floodplains Economic wealth in the Thames Estuary
Social vulnerability Changes in social vulnerability to flooding, for example due to Direction of social and culural change: community v individualistic.
changes in health and fitness, equity and systems of social provision National wealth
Infrastructure and business Systems of communication (physical and telecommunication), Technological change
“paper01” — 2008/5/26 — 20:07 — page 90 — #6

vulnerability energy distribution, etc. Changes in the extent to which society is Nature of business processes. Degree of dependence upon
dependent on these systems vulnerable infrastructures
Agriculture Changes in the intensity and seasonality of agriculture, including Agricultural and environmental policy (global, EU and UK)
removal of agricultural land from production and hence changes
vulnerability to flood damage
Construction industry Changes in technology, construction process and Technological change
national/international construction markets that determine the price Labour costs
of construction works Cost of raw materials, including energy
Discount rates Valuation of streams of costs and benefits through time Market interest rates
Ethics of long-term trade-offs of costs and risks e.g.,
intergenerational issues

i
i

i
i
i i

“paper01” — 2008/5/26 — 20:07 — page 91 — #7


i i

A framework for uncertainty analysis in flood risk management decisions 91

Gelder (2006) in their analysis of the PC-RING software,


which calculates the probability of failure of dike rings in the Interval
representation
Netherlands. The approach can, however, result in communi-

Probability density
cation problems with decision makers who think of flood risk
as being a characteristic of a flooding system that is modified
only by changes in the physical system (e.g., dike deteriora-
tion). These decision makers have been surprised and puzzled Probabilistic
representation
by the notion that flood risk can be reduced merely by col-
lecting more information. Of course to a Bayesian decision
theorist there is nothing surprising here, and indeed if this
were not the case then there would be no economic value
4.5E+05 4.7E+05 4.9E+05 5.1E+05 5.3E+05 5.5E+05
in collecting information. However, it raises particular prob- NPV
lems where (as in the Netherlands) probability of flooding is Option A Option B Option C
used as a regulatory criterion or (as in the UK) an estimate
Figure 1 Probability distributions and interval bounds of Net Present
of risk is used for year-on-year monitoring of how effective
Value, representing epistemic uncertainties.
investment in flood defences has been at reducing the risk to
which people are exposed. If uncertainties are incorporated
in risk calculations alongside and on an equal footing with “robust” is used in the sense of a robust decision, which is a
natural variability, decision makers may legitimately ask of a choice that continues to be desirable under a wide range of plau-
calculated reduction in flood risk: “Is the flood risk to which sible future conditions. The requirement for robustness becomes
people are exposed being reduced or have we merely been more acute when uncertainties are great and not readily quantified
able to reduce the uncertainties in our risk assessment?” in probabilistic terms.
(2) A number of researchers (Helton and Burmaster, 1996;
Pate-Cornell, 1996) have argued that epistemic uncertain-
4.2 Decision making under severe uncertainty
ties should be separated from natural variability, so that
rather than “integrating out” all of the uncertainties, flood When random factors influence the future performance of design
risk estimates and other decision variables of interest should options or policy alternatives, then decision makers can read-
be represented as uncertain quantities, with accompanying ily use probabilistic decision analysis. Unfortunately, some of
information on uncertainties in the form of probability dis- the most serious decisions confronting flood risk managers are
tributions or interval bounds (Fig. 2). The approach has subject to severe uncertainties that do not succumb to the con-
been adopted for flood damage reduction studies in the USA ventional methods of probabilistic decision analysis. Situations
(USACE, 1996). The probabilistic representation allows cal- of severe uncertainty are characterised by the distinct possibil-
culation of the probability that the preferred option, on the ity of surprise, of situations that were unforeseen or unexpected.
basis of expected utilities, may not be preferred, through This is more than a question of analysing the extreme tails of a
the convolution of the tails of the distributions that overlap. probability distribution – it has to do with fundamental and unex-
Additional insights may be gained, for example, through cal- pected shifts in the functioning of the system. In the face of this
culating the probability that the NPV of a given option is less type of uncertainty, analysts may attempt to construct a probabil-
than zero (Stedinger, 1997). Stedinger (1997) also proposes ity distribution over the future states of nature, but there will be
incremental analysis in order to scrutinise the marginal con- precious little evidence upon which to base such a distribution.
tribution of project components in a portfolio in order to Experts may have intuitions about what is probable, but can be
remove components that make highly uncertain contribution hard pressed to quantify the rare but possible.
to the net benefit. This raises fundamental questions about the representation
of uncertainty, which have, especially since the 1970s, fuelled
Interval representation leads to a partial ordering of prefer- a proliferation of mathematizations of uncertainty that depart
ences, so for example in Fig. 2 we can say that A  B but there is in one way or another from Kolmogorov’s axioms of probabil-
no definite ordering of preference between A and B or between ity, including fuzzy set theory (Zadeh, 1965), evidence theory
B and C. (Shafer, 1976) and the theory of imprecise probabilities (Walley,
Use of the types of uncertainty representation shown in Fig. 2 1991). The latter, which is the most general approach, involves
leads naturally to an assessment of robustness. When values of the dealing with sets of probability distributions rather than having
decision variable of interest e.g., NPVi are close to one another, to define one unique “true” distribution. The outputs of imprecise
relative to the uncertainty in the estimates of NPVi , then rela- probability analysis are typically presented as lower and upper
tively small variations may lead to a swap of preference orderings. probabilities on an event or outcome of interest.
Where they are well separated, then the preference orderings are Connected, and more fundamental, is the question of how
thought of as being robust. Various notions of robustness have in practice agents should go about making decisions in the face
been proposed by government (Defra, 2005) and in the liter- of this type of uncertainty. In situations of severe uncertainty,
ature (Rosenhead and Mingers, 2001). In this paper the term should they seek to optimise the outcome or would it be wiser to

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 92 — #8


i i

92 Jim Hall and Dimitri Solomatine

make a decision that delivers a satisfactory outcome under a wide interdependence


range of possible conditions? Ben-Haim (2006) refers to the latter
of these two approaches as “robust satisficing”. To “satisfice” world global
means to “achieve an acceptable outcome” (Simon, 1956), as markets sustainability
opposed to achieving the best possible outcome. Governments
are waking up to the importance of robust solutions in some of conventional
development
the most important policy areas. For example the UK policy on individual community
flood risk management (Defra, 2005) calls for “adaptability to
climate change through robust and resilient solutions”. Yet much
less is said about how these solutions might actually be identified national local
enterprise stewardship
and analysed.
There is always a trade-off between optimising and robust sat-
isficing. Optimal designs, which maximally exploit every feature autonomy
and resource of the system, are minimally robust to conditions Figure 2 The Foresight Futures scenario grid of four scenarios (Office
that depart from the assumptions under which the optimisation of Science and Technology, 2002).
was conducted. The quest for optimisation of flood defence sys-
tems should therefore be pursued with some caution, lest the
redundancy that contributes to safety in unforeseen situations be remains, however, of specifying the range of that space of pos-
eliminated in pursuit of economic efficiency. In the face of severe sibilities. Actually, if decision makers and their analysts know
uncertainty it becomes necessary to trade off some outcome qual- anything, it is about the central tendencies, not the bounds of
ity against the confidence (robustness to uncertainty) in attaining variation. A similar problem of precisely specifying the bounds
that outcome. The decision maker invests some resources in man- of possibility besets the use of the theory of Imprecise Probability
aging uncertainty, at the expense of garnering tangible benefit (Hall, 2003; Walley, 1991), which was discussed briefly above.
directly from those resources.
4.4 Info-gap theory for robust decision making
4.3 Scenario analysis
An alternative approach is to start an uncertainty analysis at
Scenario analysis is increasingly used for analysing long term the best estimate of the future and then examine how decision
uncertainties that are not readily quantifiable. In an uncer- options perform as conditions depart increasingly from expecta-
tainty analysis there may be a small number of variables, often tions, without identifying a worst case. If an option continues to
associated with the difficulty of predicting human and societal be preferred, even at a very large horizon of uncertainty, then it
behaviour, particularly over the long term, that stand out as is thought to be robust to uncertainty. In order to do this type of
requiring special treatment because of their severe uncertainty. analysis it is necessary to have some understanding of the ways in
The UK Government’s Foresight Flood and Coastal Defence which the future may depart from expectations. If there are many
project (Evans et al., 2004) used socio-economic scenarios to uncertain variables that together determine the performance of
look 30 to 100 years into the future at flood risk in the UK the decision options, then insights into how they co-vary can
in an attempt to find robust strategies for flood risk manage- be used to construct the uncertainty model. However, it is not
ment. Whilst there are many versions of scenario analysis, they necessary to distribute a probability measure or for that matter
all tend to be based around construction of a small number of any other measure over the space of possibilities. Like Lempter
contrasting yet internally consistent narratives about the future. et al.’s (2003) robust decision making method, uncertainty is rep-
As well as the UK Foresight programme (Fig. 2) (Office of resented in terms of sets of possible futures. However, in info-gap
Science and Technology, 2002) a narrative scenario approach theory, the boundaries of those sets are unknown and vary up to
underpins the Intergovernmental Panel on Climate Change’s Spe- a horizon that is measured by an unknown positive scalar α.
cial Report on Emissions Scenarios (IPCC, 2000). The restriction Figure 3 illustrates an instance of this approach applied to
to a small number (typically four) of possible futures is motivated long term uncertainties in climate (including sea level, storm
by the need to keep the analysis manageable and intelligible for surge frequency and rainfall in the Thames catchment) and
human participants, but overlooks the capacity to use computer economic growth (including sectoral, demographic and techno-
simulations to explore much larger spaces of possible futures. logical changes) in the Thames Estuary where London is situated.
In recent years Lempert, Popper and Bankes (2003) have In the face of uncertainty about both of these axes, an info-gap
developed computer-intensive simulation models for analysing analysis permits examination of how each of a set of options
the possible outcomes of policy decisions over large spaces of for long term flood risk management performs as the horizon of
possible futures. The approach recognises the deficiencies in any uncertainty (scaled by α) expands from the best estimate ũ. Hine
model of a complex system and does not attempt to represent the and Hall (2007) provide an illustration of the use of info-gap
uncertainties in probabilistic terms. Rather, the approach is based theory for dealing with uncertainties in flood frequency estima-
upon identifying options that perform acceptably well over the tion and hydraulic model uncertainties in design of flood defence
widest subset of the space of possible futures. The problem still crest level. Ben-Haim (2006) also describes how info-gap theory

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 93 — #9


i i

A framework for uncertainty analysis in flood risk management decisions 93

Climate change
therefore iterative, so that successive cycles refine uncertainty
estimates and progressively focus upon the uncertainties that are
V Hi most significant to decision making.
Figure 4 illustrates in outline the proposed uncertainty analysis
framework. The main steps in this process are now described.
The framework is being applied to the major long term flood risk
Hi management decisions in the Thames Estuary, UK, which will
be reported in a future paper.
α u~ (1) Establish purpose and scope of uncertainty analysis
Med The starting point for uncertainty analysis is definition of the
decisions that the analysis is intended to support and the vari-
ables upon which the decision will be based. As we have seen,
Low flood risk management is both adaptive, in that it is a process of
ongoing control of a dynamic system, and tiered, in that there is
1% 2% 3% a multi-layered process of policy, planning, strategic, design and
Economic growth p.a. operational decision making. An uncertainty analysis may seek to
Figure 3 Illustration of an info-gap analysis dealing with two severe inform more than one cycle or tier in the flood risk management
long term uncertainties in the Thames Estuary: the rate of economic process, but this added complexity reinforces the need for clear
growth (and hence vulnerability and exposure to flooding) and the rate definition of decision points and decision variables. Definition
of climate change. of the purpose of uncertainty analysis and the strategy for com-
munication of results requires early engagement of stakeholders
to ensure the legitimacy of the resulting decisions (Kleindorfer
may be applied to dealing with the valuation uncertainties in
et al., 1993; Sluijs et al., 2003).
multi-attribute decision problems.
(2) Identify and define uncertainties
This step involves identifying any factors or processes that are
5 The uncertainty analysis process
believed to influence decision variables. These are identified in a
brainstorming session, with the help of past reviews (e.g., Tables
Analysis of flood risks and flood risk management decisions pro-
2 and 3) and checklists, followed by consolidative work to struc-
vides the theoretical framework for dealing with uncertainty, but
ture the list of uncertainties and provide short descriptions of
equal attention should be applied to the process of conducting
each. The agreed list of uncertainties can then be exposed to peer
uncertainty analysis and communicating to stakeholders (Fun-
review, with further insights from peer review being fed back into
towicz and Ravetz, 1990; IPCC, 2005; Sluijs et al., 2005; Sluijs
the list of uncertainties. At this stage the aim should be for com-
et al., 2003; Walker et al., 2003). In this section a process for
pleteness, irrespective of the availability of evidence upon which
uncertainty analysis is introduced, which has been designed with
to base uncertainty estimates. Whilst factors that obviously do
major flood risk management planning and investment decisions
not influence the decision variables of interest can be excluded,
in mind. Whilst the methods, for example for estimation or prop-
at this stage any factor or process for which there is a possibility
agation of probabilities may differ (and are not discussed here)
that it may noticeably alter the output(s) of interest should be
the logical structure is intended to be generically applicable to
included.
this class of decision problems, though it may need to be adapted
to the characteristics of a specific situation. The approach is as far (3) Assemble evidence about uncertainties
as possible quantified, by using probability distributions where
Having identified uncertain factors and processes, the next step it
these can be credibly generated and using intervals or sets of
to assemble evidence that can be used to understand and quantify
probability distributions where probability distributions cannot
those uncertainties. Any evidence of potential relevance should
be justified. Uncertainties are propagated through to key decision
be included at this stage. It may take the form of:
outputs (e.g., metrics of net benefit in terms of risk reduction) and
results are presented as distributions and maps. Sensitivity anal- • measurements of variables of interest either at the site in ques-
ysis is used to understand the contribution that different factors tion (e.g., water levels) or in analogous situations (e.g., unit
make to total uncertainty. The effect of uncertainty on choices is rates for construction costs);
analysed using robustness analysis. • model calibration and/or validation data (e.g., from benchmark
One of the principles of risk-based decision making is that studies);
the amount of data collection and analysis should be proportion- • multiple model outputs (e.g. from ensembles of climate model
ate to the importance of the decision (DETR et al., 2000). This runs);
also applies to the uncertainty analysis process, which should be • observations of output variables of interest that may be com-
tiered and scalable so that it can be used from preliminary screen- pared with the result of uncertainty calculations (e.g., records
ing through to more detailed options appraisal. The process is of flood extent and/or flood damage).

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 94 — #10


i i

94 Jim Hall and Dimitri Solomatine

Establish purpose and scope of


Identify and define Peer
uncertainty analysis through
uncertainties review
interaction with stakeholders

Assemble evidence about


Document
uncertainties

Review uncertainty Construct functions quantifying Identify severe uncertainties that


Peer
quantification and if uncertainties at present and under require scenario, robustness and/or
review
necessary acquire more data conditions of future change info-gap analysis

Propagate uncertainties Propagate Analyse effect of uncertainties Store


through to measures of uncertainties on assessment of environmental results in
flood risks through to costs and other impacts a database

Examine, using plausible variations in Analyse the relative contribution of the


model response functions, the input and model uncertainties to the total
potential effects of model uncertainties uncertainty in outputs of interest

Document and
Rank uncertainties in
discuss results
order of importance
and implications

Where appropriate use


Analyse robustness
optimisation (under uncertainty)
to severe
techniques in supervised search
uncertainties
for preferred options

Examine the effects of Document and


uncertainties on option discuss results
choices, based upon agreed and
decision criteria implications

Figure 4 Uncertainty analysis framework.

Identification of relevant evidence can take place in a workshop obtained from well known physical limits. However, in almost
meeting of domain experts, which will have to be followed by all cases there will be a considerable element of expert judgement
data collection efforts that will typically involve contacting sev- involved in constructing mathematical uncertainty representa-
eral organisations. The assembled evidence should be logged and tions. The translation of evidence into quantified uncertainty
documented in a summary that can then be circulated for review representations should therefore be thoroughly documented so
by any experts not previously involved who might be aware of that it can be exposed to peer review (IPCC, 2005). Furthermore,
other relevant data. it is expected that these quantifications will be revisited during
iterations in the analysis, with the most critical one being perhaps
(4) Construct appropriate functions quantifying uncertainties subject to several iterative refinements. For critical uncertainties
The proposed uncertainty framework is based, as far as possi- that are highly dependent upon expert judgement it may be neces-
ble, upon quantified methods. Qualitative discussion provides an sary to employ formal expert elicitation methods (Cooke, 1991;
important commentary, particularly on limitations of the analysis, Vick, 2002). Some uncertainties may be highlighted as being
but a quantified approach is required in order to understand the particularly severe and so more appropriately subject to scenario
magnitude of uncertainties and to focus resources and analysis and/or info gap analysis (see section 3).
upon the key uncertainties. A critical step is therefore to derive It will not be necessary to generate quantified estimates for
quantified estimates of uncertainty based on the available evi- every uncertainty previously identified. For some it will be clear
dence. The rigour of this step and use of appropriate peer review from review of available evidence and analysis of the relevant
is of utmost importance to the credibility of the analysis. processes that the influence of the uncertainty is insignificant
These mathematical uncertainty representations will be in the so the variable can be neglected. These choices should also be
form of (in decreasing order of informativeness) (i) probability documented.
distributions, (ii) sets of probability distributions, (iii) interval If variables are not independent then it is necessary to consider
bounds. In some cases probability distributions will be able to that effect of dependence between variables. In cases where there
be estimated statistically from data or interval bounds may be are joint measurements of the variables of interest (for example

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 95 — #11


i i

A framework for uncertainty analysis in flood risk management decisions 95

surge water levels and fluvial flows) then it is possible to esti- et al., 2000). These have been extended to deal with sets of
mate joint probability distributions (Hawkes et al., 2002). Other probability distributions and intervals (Hall, 2006). The result
situations are more problematic though it is usually possible to of this variance-based sensitivity analysis are sensitivity indices
estimate conservative bounds on joint probability distributions that quantify the relative contribution of different factors to the
(Ferson et al., 2004). total output uncertainty. These sensitivity indices can be used to
rank the uncertainties. The top ranked uncertainties should be
(5) Propagate uncertainties through to outputs of interest
subject to further scrutiny whilst variables that are shown to have
Uncertainty propagation involves calculating the uncertainty dis- a negligible effect can be neglected.
tribution of one or more outputs of a model, given the quantified The results of variance-based sensitivity analysis rely upon
uncertainties on the inputs. In this analysis the key output vari- the input uncertainty functions being accurate. If the initial
ables of interest are flood risk indicators, costs, and specified uncertainty estimate for an input quantity of interest is grossly
environmental impacts. For complex and non-linear models, inaccurate then the ranking of uncertainties will be misleading.
uncertainty propagation is based on some version of a sampling There is therefore a requirement to test the robustness of the
methodology (see for example Apel et al. (2008) for a flood uncertainty ranking, subject to plausible variations in the distri-
risk analysis example), which may call flooding models directly, butions of the input variables. A process of iterative refinement
employ a pre-processed database of results or a model emulator. of the input uncertainties should converge upon a robust final
In Monte Carlo analysis the results of the analysis are presented ranking of uncertainties. This process of iteration should stop
as frequency plots (histograms) for the output quantities of inter- when key uncertainties are sufficiently confidently defined that
est. Non-parametric methods (Silverman, 1986) are then used to the ranking of top uncertainties does not vary significantly.
construct smooth and normalised probability density functions
from these plots. (8) Examine the effects of uncertainties on option choices
In propagating input uncertainties it is (often implicitly) The focus of the uncertainty analysis is upon examining how sen-
assumed that the model used for propagating uncertainties is sitive choices are to ambient uncertainties, now and in the future.
true. This of course is not the case, so at the same time the effect Analysis of the effect of uncertainties upon choices requires well
of plausible variations in model behaviour should be examined. defined decision criteria. In section 3, different versions of robust-
Some insights can be gained for example by changing model cal- ness analysis, including the methods of Lempert et al. (2003)
ibration parameters or refining grid resolution. It is at this stage and info-gap analysis (Ben-Haim, 2006), were discussed. It is at
in the analysis that the approaches for conditioning on observed this stage that they can be used to inform the choice of options,
data are of relevance (Beven, 2006; Duan et al., 2003; Liu and based upon analysis of the most severe uncertainties previously
Gupta, 2007; Montanari and Brath, 2004; Shrestha and Solo- identified.
matine, 2008). Alternative representations of physical processes Choosing between options is an iterative process, with some
may be explored (Georgakakos et al., 2004). Comparison with options being screened out at a relatively early stage while oth-
validation data will give insights into the magnitude of model ers are progressively refined. Each of these choices is subject
structural uncertainty, though this data may not be available for to uncertainty, particularly at the early stages in analysis, so
the most important, extreme, events (van Gelder et al., 2008). for example, decisions to exclude options from further analysis
(6) Store the results in a database should be robust even given the large uncertainties at early stages
The uncertainty analysis yields very large amounts of data. Fur- in the analysis. At later stages in the decision making process
thermore, it is possible to illustrate system response conditional options will be developed in more detail and it is at this stage that
upon a wide range of specified events, for example a water level approaches to optimisation under uncertainty show most promise,
of given return period, so the data used to generate these results subject to the limitations discussed above.
should be stored in a database. Capacity to visualise the large
(9) Report and discuss results
quantity of outputs in the database is essential for assuring the
quality of analysis. The uncertainty analysis requires careful interpretation in order
Key results should be documented in written reports, but the to understand the meaning and significance of the results. It is
full result set should be retained in a database that can be inter- through this process of scrutiny and discussion that the most
rogated for specific options, variables and locations and can be useful insights for decision makers are obtained.
used to present results as graphs, maps and time series. Furthermore, the conduct of uncertainty analysis provides new
insights into model behaviour that will need to be discussed and
(7) Uncertainty-based sensitivity analysis agreed with the experts responsible for models that are employed
Important insights can be gained by attributing output uncer- in the analysis. Indeed the process of scrutiny that uncertainty
tainty to input variables and model uncertainties in order to find analysis provides is an additional benefit. These insights also
out which variables and models make the greatest contribution need to be well documented (USACE, 2006).
to the uncertainty in the outputs of interest. There is a well estab- Experience is now growing in the communication of uncer-
lished set of approaches to decomposing the variance in an output tainty to decision makers and members of the public, for example
variable into the contributions from all of the input variables, act- in the context of environmental risks (Sluijs et al., 2003) and
ing individually or in combination (Hall et al., 2005; Saltelli climate change (IPCC, 2005). Sluijs et al. (2003) stress the

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 96 — #12


i i

96 Jim Hall and Dimitri Solomatine

importance of engaging stakeholders from this early stage, iden- Acknowledgements


tifying the target audiences and then using appropriate language
to communicate uncertainties. Alongside numerical results and The work described in this publication was supported by the
their implications for decision makers, the limitations of data European Community’s Sixth Framework Programme through
sources and analysis methods should be made clear and areas of the grant to the budget of the Integrated Project FLOODsite,
ignorance should be highlighted. Contract GOCE-CT-2004-505420 and through the UK Environ-
ment Agency’s Thames Estuary 2100 programme through project
SD7. This paper reflects the authors’ views and not those of the
6 Conclusions European Community or the Environment Agency. Neither the
European Community nor any member of the FLOODsite Con-
Whilst much published research dwells upon technical method- sortium nor the Environment Agency is liable for any use of the
ologies for uncertainty analysis, this paper has focused upon the information in this paper. The advice of Prof Yakov Ben-Haim is
process of uncertainty analysis, with particular reference to how gratefully acknowledged.
uncertainty analysis can inform decision making. The basis for
risk-based decision making is well established, but the expanding
scope of flood risk management to embrace a broader range of References
non-structural flood risk management measures alongside more
conventional structural measures raises new challenges for uncer- 1. Al-Futaisi, A. and Stedinger, J.R. (1999). “Hydrologic
tainty analysis that require careful choice of methods within a and Economic Uncertainties in Flood Risk Project Design,”
rigorous overall framework. The scope of uncertainty analysis set Journal of Water Resources Planning and Management,
out in this paper encompasses all of the natural, human and tech- 125(6), 314–324.
nological processes that may influence flood risk, both at present 2. Apel, H., Merz, B. and Thieken, A.H. (2008). “Quantifica-
and in the future under scenarios of change. It is designed to sup- tion of Uncertainties in Flood Risk Assessments,” Journal
port strategic analysis on a broad scale to examine portfolios of of River Basin Management (This issue).
interventions that may be implemented in a phased way across 3. Apostolakis, G. (1990). “The Concept of Probability in
whole catchments or coastlines. These measures will be operated Safety Assessments of Technological Systems,” Science,
in combination and can be expected to be modified as conditions 250, 1359–1364.
and preferences change and as more information becomes avail- 4. Ball, D.J. and Floyd, P.J. (1998). Societal Risk, Health &
able. Furthermore, sensitivity analysis is also required to explore Safety Executive, Risk Assessment Policy Unit, London.
the valuation problem associated with aggregating economic, 5. Beard, L.R. (1997). “Estimated Flood Frequency and Aver-
public safety, social and environmental impacts of flooding. Each ageAnnual Damages,” Journal of Water Resources Planning
of the considerations mentioned above depends upon the context and Management, 123(2), 84–88.
of particular flood risk management decisions, and the approach 6. Ben-Haim, Y. (2006). Information-Gap Decision Theory:
to uncertainty analysis will need to be modified accordingly. Decisions Under Severe Uncertainty, Academic Press, San
However, insofar as a flood risk management is now emerging Diego.
as the modern paradigm for dealing with flooding, it is timely 7. Beven, K. (2006). “A Manifesto for the Equifinality Thesis,”
to consider how uncertainty should in general be incorporated in Journal of Hydrology, 320, 18–36.
that paradigm. 8. Cooke, R.M. (1991). Experts in Uncertainty: Opinion and
Strategic analysis of flooding systems and portfolios of man- Subjective Probability in Science, Oxford University Press,
agement options is involving new high resolution datasets and New York.
coupling of computer models that simulate a variety of flooding 9. CUR/TAW (1990). Probabilistic Design of Flood Defences,
processes. Whilst these integrated modelling systems are pro- Centre for Civil Engineering Research and Codes (CUR),
viding unprecedented insights they are also, because of their Technical Advisory Committee on Water Defences (TAW),
complexity, increasingly difficult for any individual to inter- Gouda, the Netherlands.
pret, so require systematic uncertainty analysis in order to 10. Dantzig, D. van (1956). “Economic Decision Problems
build confidence in their outputs. Decision makers are increas- for Flood Prevention,” Econometrica, 24(3), 276–287.
ingly demanding information about the uncertainties in the 11. Dawson, R.J. and Hall, J.W. (2006). “Adaptive Impor-
evidence upon which they are basing decisions. A number of tance Sampling for Risk Analysis of Complex Infrastructure
challenges associated with the more comprehensive use of uncer- Systems,” Proc. Royal Soc. A: Mathematical, Physical and
tainty analysis have been highlighted in this paper, including Engineering Sciences, 462(2075), 3343–3362.
the communication of uncertainties to decision makers and the 12. Defra (2001). Flood and Coastal Defence Project
representation of severe uncertainties. Whilst an example of Appraisal Guidance: Overview (Including General Guid-
a comprehensive analysis that incorporates all of the issues ance), FCDPAG1, Department of the Environment, Food
addressed in this paper has yet to be published, it is intended that and Rural Affairs, London.
the framework set out in this paper will pave the way for a new 13. Defra (2005). Making Space for Water: Taking Forward a
generation of practical decision-oriented uncertainty analysis. New Government Strategy for Flood and Coastal Erosion

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 97 — #13


i i

A framework for uncertainty analysis in flood risk management decisions 97

Risk Management in England: First Government Response 29. Hall, J.W., Tarantola, S., Bates, P.D. and M.S., H.
to the Autumn 2004 Making Space for Water Consultation (2005). “Distributed Sensitivity Analysis of Flood Inunda-
Exercise, Department of the Environment, Food and Rural tion Model Calibration,” Journal of Hydraulic Engineering,
Affairs, London. 131(2), 117–126.
14. DETR, Environment Agency and Institute of Environmental 30. Harvey, D.P., Peppé, R. and Hall, J.W. (2008). “Reframe:
Health (2000). Guidelines for Environmental Risk Assess- A Framework Supporting Flood Risk Analysis,” Journal of
ment and Management, The Stationery Office, London. River Basin Management (This issue).
15. Duan, Q., Gupta, H., Sorooshian, S., Rousseau, A.N. 31. Hawkes, P.J., Gouldby, B.P., Tawn, J.A. and Owen, M.W.
and Turcotte, R. (2003). Calibration of Watershed Mod- (2002). “The Joint Probability of Waves and Water Levels in
els. Water Science and Applications, American Geophysical Coastal Engineering,” Journal of Hydraulic Research, 40,
Union. 241–251.
16. Egorova, R., Noortwijk, J.M.V. and Holterman, S.R. 32. Helton, J.C. and Burmaster, D.E. (1996). “Treatment of
(2008). “Uncertainty in Flood Damage Estimation,” Journal Aleatory and Epistemic Uncertainty in Performance Assess-
of River Basin Management (This issue). ments for Complex Systems,” Reliability Engineering and
17. Evans, E.P. et al. (2004). Foresight Flood and Coastal System Safety, 54, 91–258.
Defence Project: Scientific Summary: Volume I, Future 33. Hine, D. and Hall, J.W. (2007). “Analysing the Robustness
Risks and their Drivers, Office of Science and Technology, of Engineering Decisions to Hydraulic Model and Hydro-
London. logical Uncertainties,” Harmonising the Demands of Art and
18. Ferson, S. et al. (2004). Dependence in Probabilistic Mod- Nature in Hydraulics: Proc. 32nd Congress of IAHR, IAHR,
eling, Dempster-Shafer Theory, and Probability Bounds Venice.
Analysis, SAND2004-3072, Sandia National Laboratory, 34. Hutter, G. and Schanze, J. (2008). “Learning How to Deal
Albuquerque, NM. with Uncertainty of Flood Risk in Long-Term Planning,”
19. Finkel, A.M. (1990). Confronting Uncertainty in Risk Man- Journal of River Basin Management (This issue).
agement: A Guide for Decision-Makers, Center for Risk 35. Interagency Floodplain Management Review Committee
Management, Resources for the Future, Washing, DC. (1994). Sharing the Challenge: Floodplain Management
20. French, S. (1988). Decision Theory: An Introduction to the into the 21st Century, Report to the Administration Flood-
Mathematics of Rationality, Ellis Horwood, Chichester. plain Management Task Force, U.S. Government Printing
21. Frieser, B.I., Vrijling, J.K. and Jonkman, S.N. (2005). Office, Washington, DC.
“Probabilistic Evacuation Decision Model for River Floods 36. IPCC (2000). Special Report on Emissions Scenario, Inter-
in the Netherlands,” 9th International Symposium on governmental Panel on Climate Change.
Stochastic Hydraulics, IAHR, Nijmegen, the Netherlands. 37. IPCC (2005). Guidance Notes for Lead Authors of the IPCC
22. Funtowicz, S.O. and Ravetz, J. (1990). Uncertainty and Fourth Assessment Report on Addressing Uncertainties,
Quality in Science for Policy, Kluwer, Dordrecht. Intergovernmental Panel on Climate Change.
23. Galloway, G.E. (2005). “Corps of Engineers Responses 38. Kahan, J.P., Wu, M., Hajiamiri, S. and Knopman, D.
to the Changing National Approach to Floodplain Man- (2006). From Flood Control to Integrated Water Resource
agement Since the 1993 Midwest Flood,” Journal of Management: Lessons for the Gulf Coast from Flooding in
Contemporary Water Research and Education, 130, Other Places in the Last Sixty Years, RAND Corporation,
5–12. Santa Monica.
24. Georgakakos, K.P., Seo, D.-J., Gupta, H.V., Schaake, 39. Kanning, W. and van Gelder, P.H.A.M. (2006). Analysis
J. and Butts, M.B. (2004). “Towards the Character- and Influence of Uncertainties on the Reliability of Flood
ization of Streamflow Simulation Uncertainty Through Defence Systems, T07-06-07.
Multimodel Ensembles,” Journal of Hydrology, 298, 40. Kleindorfer, P.R., Kunreuther, H.G. and Schoemaker,
222–241. P.J.H. (1993). Decision Sciences: An Integrative Perspec-
25. Goldman, D. (1997). “Estimating Expected Annual Dam- tive, Cambridge University Press, Cambridge.
age for Levee Retrofits,” Journal of Water Resource. 41. Knight, F.H. (1921). Risk, Uncertainty, and Profit Hart,
Planning and Management, 123(2), 89–94. Schaffner & Marx, Boston, MA.
26. Hall, J.W. (2003). “Handling Uncertainty in the Hydroin- 42. Lempert, R.J., Popper, S.W. and Bankes, S.C. (2003).
formatic Process,” Hydroinformatics, 5(4), 215–232. Shaping the Next One Hundred Years: New Methods for
27. Hall, J.W. (2006). “Uncertainty-Based Sensitivity Indices Quantitative, Long-Term Policy Analysis, RAND, Santa
for Imprecise Probabilities,” Reliability Engineering and Monica, CA.
Systems Safety, 91(10–11), 1443–1451. 43. Liu, Y. and Gupta, H. (2007). “Uncertainty in Hydro-
28. Hall, J.W., Meadowcroft, I.C., Sayers, P.B. and Bram- logic Modeling: Toward an Integrated Data Assimilation
ley, M.E. (2003). “Integrated Flood Risk Management Framework,” Water Resources Research, 43(7).
in England and Wales,” Natural Hazards Review, 4(3), 44. Meadowcroft, I.C., Hall, J.W. and Ramsbottom,
126–135. D.M. (1997). Application of Risk Methods in Flood and

i i

i i
i i

“paper01” — 2008/5/26 — 20:07 — page 98 — #14


i i

98 Jim Hall and Dimitri Solomatine

Coastal Engineering: Scoping Study, Report SR483, HR 62. Silverman, B.W. (1986). Density Estimation, Chapman
Wallingford, Oxfordshire. and Hall, London.
45. Montanari, A. and Brath, A. (2004). “A Stochastic 63. Simon, H. (1956). “Rational Choice and the Structure of the
Approach for Assessing the Uncertainty of Rainfall-Runoff Environment,” Psychol. Rev., 63, 81–97.
Simulations,” Water Resources Research, 40(1), W01106. 64. Sluijs, J.P. van de et al. (2005). “Combining Quantitative
46. Munich Re Group (2007). Natural Catastrophes 2006: and Qualitative Measures of Uncertainty in Model-Based
Analyses, Assessments, Positions, 302-05217. Environmental Assessment: The NUSAP System,” Risk
47. National Academy of Engineering (2000). Risk Analysis and Analysis, 25(2), 481–492.
Uncertainty in Flood Damage Reduction Studies, National 65. Sluijs, J.P. van de et al. (2003). RIVM/MNP Guidance for
Academy Press, Washington DC. Uncertainty Assessment and Communication, Netherlands
48. National Research Council (1999). Improving American Environmental Assessment Agency (RIVM/MNP).
River Flood Frequency, National Academy Press, Wash- 66. Stedinger, J.R. (1997). “Expected Probability and Annual
ington DC. Damage Estimators,” Journal of Water Resources Planning
49. National Research Council (2000). Risk Analysis and and Management, 123(2), 125–135.
Uncertainty in Flood Damage Reduction Studies, National 67. Todini, E. (2008). “Predictive Uncertainty in Flood Fore-
Academy Press, Washington DC. casting Models,” Journal of River Basin Management (this
50. Newcastle University and Halcrow (2006). Early Concep- issue).
tual Options, A Framework for Uncertainty Analysis in 68. USACE (1996). Risk-Based Analysis of Flood Damage
TE2100, ECO Report R10X, Environment Agency, London. Reduction Studies, U.S. Army Corps of Engineers, Wash-
51. Office of Science and Technology (2002). Foresight Futures ington, DC.
2020: Revised Scenarios and Guidance, Department of 69. USACE (2006). Risk Analysis for Flood Damage Reduction
Trade and Industry, London. Studies, U.S. Army Corps of Engineers, Washington, DC.
52. Pappenberger, F. and Beven, K. (2006). “Ignorance is 70. USWRC (1983). Economic and Environmental Princi-
Bliss: 7 Reasons not to use Uncertainty Analysis,” Water ples and Guidelines for Water Related Land Resources
Resources Research, 42(5). Implementation Studies, U.S. Water Resources Council,
53. Pappenberger, F., Harvey, D.P., Beven, K., Hall, Washington DC.
J.W. and Meadowcroft, I.C. (2006). “Decision Tree for 71. van Gelder, P.H.A.J.M., Devictor, N. and Marquès,
Choosing an Uncertainty Analysis Methodology: A Wiki M. (2008). “Representation of Statistical Uncertainties
Experiment,” Hydrological Processes, 20(17), 3793–3798. in the Estimation of Extremes,” Journal of River Basin
54. Pasman, H.J. and Vrijling, J.K. (2003). “Social Risk Management (This issue).
Assessment of Large Technical Systems,” Human Factors 72. Vick, S.G. (2002). Degrees of Belief: Subjective Probability
and Ergonomics in Manufacturing, 13(4), 305–316. and Engineering Judgement. ASCE Press, Reston VA.
55. Pate-Cornell, M.E. (1996). “Uncertainties in Risk Anal- 73. Vrijling, J.K. (1993). “Development in Probabilistic
ysis, Six Levels of Treatment,” Reliability Engineering and Design of Flood Defences in the Netherlands,” in Yen, B.C.
System Safety, 54, 95–111. and Tung, Y.-K. (eds.), Reliability and Uncertainty Analysis
56. Rosenhead, J. and Mingers, J. (2001). Rational Analysis in Hydraulic Design, ASCE, New York, pp. 133–178.
for a Problematic World Revisited, Wiley, Chichester. 74. Vrijling, J.K. and Van Gelder, P.H.A.J.M. (2006). “Impli-
57. Saltelli, A., Chan, K. and Scott, M. (2000). Sensitivity cations of Uncertainties on Flood Defense Policy,” in
Analysis, Wiley, New York. Vrijling, J.K. et al. (eds.), Stochastic Hydraulics ’05, Proc.
58. Savage, L.J. (1954). The Foundations of Statistics, Wiley, 9th Int. Symp. on Stochastic Hydraulics, IAHR, Nijmegen,
New York. the Netherlands, 8 pp.
59. Sayers, P.B., Hall, J.W. and Meadowcroft, I.C. (2002). 75. Walker, W.E. et al. (2003). “Defining Uncertainty: A
“Towards Risk-Based Flood Hazard Management in the Conceptual Basis for Uncertainty Management in Model-
UK,” Civil Engineering, 150(1), 36–42. Based Decision Support,” Integrated Assessment, 4(1),
60. Shafer, G. (1976). A Mathematical Theory of Evidence, 5–17.
Princeton University Press, Princeton, NJ. 76. Walley, P. (1991). Statistical Reasoning with Imprecise
61. Shrestha, D.L. and Solomatine, D.P. (2008). “A Method Probabilities, Chapman and Hall, London.
of Localised Model Uncertainty Prediction and its Applica- 77. Zadeh, L.A. (1965). “Fuzzy Sets,” Information and Con-
tions,” Journal of River Basin Management (this issue). trol, 8, 338–353.

i i

i i

You might also like