Professional Documents
Culture Documents
Catchment Modelling
Catchment Modelling
2–3 9 2011
ABSTRACT
This paper presents a brief historical excursus on the development of hydrological catchment E. Todini (corresponding author)
Department of Earth and Geo-Environmental
models together with a number of possible future perspectives. Given the wide variety of Sciences,
Via Zamboni,
available hydrological models which, according to the embedded level of prior physical 67 – 40126 Bologna,
Italy
information, vary from the simple input–output lumped models to complex physically meaningful E-mail: [email protected]
ones, the paper suggests how to accommodate and to reconcile the different approaches. This
can be performed by better clarifying the roles and the limitations of the different models through
objective benchmarks or test-beds characterizing the diverse potential hydrological applications.
Furthermore, when dealing with hydrological forecasting, the reconciliation can be obtained
in terms of forecasting uncertainty, by developing Bayesian frameworks to combine together
models of different nature in order to assess and reduce predictive uncertainty.
Key words 9 conceptual, data-driven, hydrological models, physically based, predictive uncertainty,
validation uncertainty
INTRODUCTION
The history of hydrological modelling ranges from the gov/oh/hrl/dmip) in order to assess the performance of
Rational Method to recent distributed physically meaningful distributed hydrological models.
models. Over the same period of time, starting from the Today the plethora of available models has grown beyond
simple Unit Hydrograph, input–output models, now called any possible limit and the need for accommodating under a
data-driven models, have evolved into artificial neural net- unifying view and reconciling the different approaches has
work (ANN) models and data-based mechanistic (DBM) become of great priority. Unfortunately, hydrology is one of
models. From the wide range of models available, the choice the few scientific branches where standards on the use and
of the one most appropriate for any specific task is difficult, development of models are difficult to set and, although there
particularly as each modeller tends to promote the merits of is an increasing awareness at developing standards for
his own approach. Moreover, apart from the WMO (1975) calibration and verification (Refsgaard et al. 2005; Jakeman
inter-comparison of conceptual models, the WMO (1986) et al. 2006), there is still a long way to go.
inter-comparison of snow accumulation and melting models Moreover, hydrological models serve many purposes, one
and the WMO (1992) inter-comparison of real-time updating of the most important applications being flood forecasting
approaches applied to hydrological flood forecasting models, where uncertainty plays a major role. Unfortunately, the
no further objective comparisons using benchmarks or stan- implication of using uncertainty in the decision-making pro-
dard data sets have been proposed or effected in the last cess and even the concept of uncertainty seem to deter
decades. Only recently an inter-comparison of distributed hydrologists from addressing the problem. Indeed, many
model was started by the US-NWS (http://www.nws.noaa. hydrologists do not appear to be aware of the need to
doi: 10.2166/nh.2011.096
quantifying predictive uncertainty and tend to describe the UH, with the advent of system’s theory, was then extended to
model sensitivity rather than the decision makers’ uncertainty a variety of shapes in the form of impulse responses of
on the outcome of possible future occurrences given (or causative linear dynamic systems, known in the literature as
conditional upon) the model forecasts. the ‘‘linear models’’ (Dooge 1973; Nash 1958, 1960).
This paper will briefly describe the historical development Initially, the need to extend the results obtained with the
of the different hydrological models and will try to suggest linear models to larger and non-entirely impervious catch-
possible ways to reconcile the different approaches both on ments prompted the separation of the ‘‘effective rainfall’’,
the basis of their potential use as well as in terms of their namely the component of rain that would generate surface
Bayesian combination aimed at benefiting of all possible runoff (Chow et al. 1988), to be then modelled using the linear
information generated by the use of alternative models within models. Further on, with the advent and the wider availability
the frame of the decision making process. Finally, the paper of digital computers ‘‘conceptual models’’ started to appear
concludes with an overview of possible future perspectives in (Dawdy & O’Donnell 1965) with the aim of modelling the
hydrological research. complex interdependence of soil-surface runoff generating
mechanisms.
A BRIEF HISTORY OF QUANTITATIVE HYDROLOGICAL From conceptual to variable contributing area models
MODELS (1960–2000)
Although a more detailed overview of the historical develop- To achieve a better physical interpretation of catchment
ment of hydrological models was given in an earlier paper response, the 1960s saw the development of models in
(Todini 2007), a brief historical description was deemed which individual components in the hydrological cycle
necessary to better understand the development of thought were represented by interconnected conceptual elements;
and the future perspectives in hydrological modelling. each of these represented, in the hydrological model, the
response of a particular subsystem: for example Dawdy &
From the rational method to the conceptual models O’Donnell (1965), Crawford & Linsley (1966) – Stanford
(1850–1960) Watershed IV; Burnash et al. (1973) – Sacramento (Figure 1);
Rockwood (1964) – SSARR; Sugawara (1967, 1995) – Tank.
It must be acknowledged that from the onset, the develop-
ment of hydrological models has always been motivated by
practical engineering problems. For instance, the generally
recognised first hydrological model, the Rational Method,
proposed by Mulvany (1850) came as a response to the need
for designing sewers, to which it was extensively applied
(Kuichling 1889; Lloyd-Davies 1906), or dam spillways in
small, practically impervious catchments. The Rational
Method, uses the concept of time of concentration to estimate
the peak flow in small impervious catchments, such as the
urban or mountain catchments, where flow can be well
assimilated to a purely kinematic process.
Also, the development of the unit hydrograph (UH),
introduced by Sherman (1932), was motivated by the need
for providing not only the peak flow, but the shape and the
volume of the flood wave to improve the design of reservoirs
and flood defences. From its original triangular shape, the Figure 1 9 Schematic representation of a typical conceptual model: the Sacramento model.
All these models, also known as Explicit Soil Moisture between the volume of water stored in the soil and the extent
Accounting (ESMA) models, represented in different ways of saturated areas (the topographic index function) on the
the responses of, and the interconnections between, the basis of physically meaningful parameters. Unfortunately, also
various subsystems from which the overall catchment due to a water balance error which was present in the original
response could originate (see Figure 2); at the time, they TOPMODEL, recently detected and corrected by Saulnier &
were regarded as the very best that could be achieved with the Datin (2004), the physical meaning of parameters proved to be
then current data and computational resources. true only for very small hill-slope catchments represented with
At the end of the 1970s, a new type of lumped model was extremely fine meshes as found by Franchini et al. (1996).
introduced, based on the idea that the rainfall runoff process
is mainly dominated by the dynamics of saturated areas, The distributed process models (1965–today)
which can be related to the soil moisture storage using a
simple monotone function, thus leading to the variable con- As an alternative to conceptual models several authors aimed
tributing area models. These models generally employed the at improving the physical representation of the rainfall-runoff
Dunne (1978) assumption that all precipitation enters the soil process. For instance, Wooding (1965a, b, 1966) and Wool-
and that surface runoff originates by saturation of the upper hiser & Liggett (1967) used kinematic models for the study of
soil layer. These variable contributing area models, the Xinan- small urban basins, while Freeze & Harlan (1969) proposed a
jiang due to Zhao (1977) and the Probability Distribution mathematical model based on distributed physical knowledge
(PDM) proposed by Moore & Clarke (1981), the ARNO of surface and subsurface phenomena. By numerical integra-
(Todini 1996) were characterized by few significant para- tion of the coupled sub-systems of partial differential equa-
meters, which unfortunately could not be directly derived, tions describing surface flow and flow in the unsaturated and
but needed to be estimated during model calibration. saturated zones, and by matching the solutions of each sub-
More recently, Beven & Kirkby (1979), with the increased system with the boundary conditions of another, catchment
availability of digital terrain models (DTM), originated the scale predictions could be produced. This concept was then
TOPMODEL, based on the distribution function of a topo- developed into SHE (Système Hydrologique Européen)
graphic index, that can be derived from the DTM. Based on (Abbott et al. 1986a, b), by the Danish Hydraulic Institute
the assumption that the accumulation of soil moisture can be (DK), the Institute of Hydrology at Wallingford (UK) and
approximated by successive steady states of the water table SOGREAH (France). Figure 3 shows a sketched representa-
originating in the upper soil layer, they derived a relation tion of the SHE model.
SAC-SMA Model
Precipitation
Impervious and Direct Runoff
Figure 2 9 The different components forming a flood wave as in the Sacramento model.
The limitation to its practical use is the large requirement fact the link between the Transfer Function or Impulse
for data and computational time which restrict its use to small Response models and the Auto-Regressive with Exogenous
extensively instrumented catchments. variables models (ARX). Later, system engineering
More recently, the wider availability of distributed approaches, including various types of input–output techni-
information, ranging from soil types and land use to radar ques, were applied to develop better performing and more
rainfall, have facilitated the production of simplified physi- parsimonious models to represent the hydrological behaviour
cally meaningful distributed hydrological models. These of a catchment, at the expense of a larger loss of physical
models, based on the upscaling of point process equations interpretation.
to finite-dimension pixels, with simpler and more parsimo- This loss of physicality increased further with ANN
nious parameterizations than those employed in MIKE approaches, which can be viewed as nonlinear analogues of
SHE (Refsgaard & Storm 1995) and SHETRAN (Ewen the original linear transfer function models. Although Daw-
et al. 2000), can also be applied to operational flood fore- son & Wilby (2001) and Shamseldin (1997) review applica-
casting. This is the case of TOPKAPI (Todini 1995; Liu & tions of ANN to rainfall-runoff modelling, few operational
Todini 2002; Todini & Ciarapica 2002), LISFLOOD (De forecasting systems are presently based on ANN (Garcı̀a-
Roo et al. 1998, 2000) and TETIS (see Figure 4) (Vélez 2001; Bartual 2002) since outside of the range of the training set, the
Francés et al. 2007). ANN may be less robust and may sometimes diverge (Gaume
& Gosset 2003).
The data-driven models (1970–today) More recently, there is an interesting return to the need
for justifying the identified models on physical grounds. The
The Sherman (1932) UH together with all the ‘‘linear models’’ DBM modelling approach, introduced by Young (2002), is a
(Nash 1958, 1960; Dooge 1973) can be viewed as the first data- tentative attempt to go beyond the black-box concept by
driven models in hydrology. Box & Jenkins (1970) showed in selecting, among the resulting model structures, those that
Evapo-transpiration
Snow model
Infiltration
Surface runoff
Lake / Reservoir
Drainage network flow
are considered physically meaningful (Young 2001, 2002). on the other hand at successfully modelling the relations
Unfortunately, apart from simple routing models, scant exam- between inflow and outflow regardless of the complexity of
ples of this modelling approach are presently available in phenomena, is still an unresolved problem. Needless to say,
operational hydrology. no model is either fully data driven or entirely physically
based; nonetheless, a discussion on these two lines of thought
is essential, given the practical implication that they have over
ACCOMMODATING AND RECONCILING the present and future development and availability of hydro-
HYDROLOGICAL MODELS logical models. In broad terms, the development of a model is
the synthesis of what the modeller assumes to know with
Data-driven or physics-based models what he regards as unknown, to be derived from data and, in
principle, the modeller should introduce as prior knowledge
The dichotomy between the two lines of thought on the one all the information he or she is sure of or comfortable with,
hand of hydrologists aiming at understanding and improving unless this results in excessive data demand.
the physical representation of processes in their models and From the lower end a modeller may assume to know
virtually nothing and statistical as well as system engineering
or ANN approaches will provide him the techniques to
identify at the same time the structure of the model and the
relevant parameters. The structure of the model will be then
based upon one or more simple model forms, a typical
representation of which is given in Figure 5, such as for
instance the negative exponential (linear reservoir) in impulse
response models, the polynomial ratio forms in the Laplace
or Z transformed spaces or the Logistic function in the case of
the ANN approaches.
But, regardless to the simplicity of the lumped input–
output model, when dealing with physical systems, a modeller
would like to build in some lumped form of the obvious
physical laws such as, for instance, water mass balance or
Figure 5 9 A typical representation of an ANN model. some simple form of energy or momentum balance. There-
fore, to introduce this prior information into the model one The next step in increasing prior knowledge into a
must either introduce it in terms of constraints over the space hydrological model not only involves model complexity,
of parameters (Natale & Todini 1976a, b) or to justify the which is now generally expressed in the form of distributed
shape of the selected impulse response on physical grounds. differential equations, but also a tremendous amount of
For instance, Kalinin & Milyukov (1957) demonstrated that, additional distributed information. When performing flood
by linearizing the unsteady flow equations, the integral routing using the full Saint Venant equations, the modeller
solution (the impulse response of the system) has the shape must be prepared to introduce a very detailed description of
of a Gamma density function, better known in hydrology as the channel morphology and of its hydraulic conveyance
the Nash (1958, 1960) cascade with parameters n and k, where characterization. Not to say the amount of information
the parameter n is now extended to the domain of real needed to develop a fully distributed catchment model such
numbers. as the SHE model (Abbott et al. 1986a, b). In this case a major
At this point the modeller has two choices: either to problem lies in the fact that the original process differential
estimate the model parameters by matching the results of equations were studied and are valid at the infinitesimal scale,
the model to a set of historical observations or, if he believes while in the model their validity is extended up to the pixel
in the physical approximation, he can relate the parameters to scale (from metres to hundreds of metres), without taking
measurable quantities. For instance, in the case of the above- into account the obvious spatial averaging effect that occurs
mentioned routing problem the parameters can be expressed in the upscaling process. This in turn impacts on the validity
in terms of the Froude number, the bed slope, the velocity, of the used ‘‘physically meaningful’’ parameter values: these
etc., as shown by Dooge (1973). values may be correct at a point, but, due to the not infrequent
Another step, towards increasing the complexity of prior high spatial variability, they are not necessarily true or valid at
information to be introduced in a model, arises when the the pixel scale over which the model is discretized. Therefore,
modeller wants to reproduce not only the outflow, but also effective parameters have again to be calibrated using series
the state of the internal state variables, such as for instance of available historical data. This is why a more realistic
the state of soil moisture in a hydrological model. The approach seems then, starting from the same point process
modeller will then have to specify the structure and the equations, to derive their average behaviour over the finite
interlinks of ‘‘sub-models’’ such as evapo-transpiration, soil pixel scale, as for instance in TOPKAPI (Todini 1995; Liu &
moisture infiltration and balance, overland and channel flow. Todini 2002; Todini & Ciarapica 2002). This allows one to use
This was in fact done in the conceptual models where each representative average parameter values (such as hydraulic
component of the hydrologic cycle was represented in the conductivity, porosity, slope) at the pixel scale up to scales of
form of buckets, thresholds, connections, etc., but the result- the order of a few hundred metres without losing model
ing models became dependent on far too many parameters to performance, as shown by Martina et al. (2009), with the
be estimated with input (rainfall) and overall output (dis- additional advantage that these parameters can be reasonably
charge in the river). Using this approach it was virtually derived from available maps (DTM, soil maps and land-use
impossible to relate the parameters to measurable quantities maps).
for many reasons: the ‘‘sub-models’’ were mostly coarse Unfortunately, the proponents of the data-driven models
bucket analogues of the processes involved; validation of and the fans of distributed differential models do not discuss
the assumed internal structure was not possible due to the in terms of the advantages and the fields of applications of the
fact that only precipitation and flow in the channel were alternative approaches. The first ones say that complex
available. Also, calibration could be extremely difficult due to models require too many data and are still undetermined in
the large number of parameters and their interdependence terms of representative parameter values, while the others say
(Gupta & Sorooshian 1983; Sorooshian & Gupta 1983), which that the data-driven models can hardly be extrapolated
resulted into a loss of confidence in ‘‘complex physically beyond the range of the historical data used for calibration.
based’’ models as the conceptual models were considered in It is therefore hoped that efforts will be made towards a
the 1970s and 1980s. reconciliation of the two lines of thought, since they are both
valid and fruitful for the improvement of the quality of by Chow et al. (1988) becomes more and more difficult to
representation of hydrological systems. actually represent the wide variety of available models.
As an alternative, Todini (1988), in order to assess the
Towards new possible classifications of models and state of the art of hydrological models, proposed a simple
the need for test beds classification based upon the level of prior knowledge intro-
duced in the model and assumed for the parameter values.
Today, users are frequently uncertain about the selection of This classification, which was just sketched in the refer-
the most appropriate hydrological model to suit their pur- enced paper, is not conclusive, but it is probably along these
poses given the wide variety of existing models (Singh & lines that the models should be assessed and classified with
Woolhiser 2002). A rather general classification of hydrologi- the aim of clarifying to possible users, in relation to the
cal models was provided in 1988 by Chow et al. (1988), as problem to be solved, the quantity and quality of assumptions
shown in Figure 6. Unfortunately, 30 years later this classifi- made; the need for geo-morphological information; the role
cation does not seem to be fully satisfactory. With the of uncertainty and the calibration requirements.
introduction of the concepts of ‘‘predictive uncertainty’’ (de This implies the definition of a number of standard test
Finetti 1975) and ‘‘equifinality’’ (Bertalanffy 1968; Beven & beds, covering a wide variety of engineering and water
Binley 1992; Beven & Freer 2001) many models, following the resources problems, in order to operationally compare the
basic Bayesian principle, are now more correctly viewed as a models also in relation to their declared objectives, their
combination of what is assumed to be known and what is performances and their ease of use.
derived from the observations. Under these new concepts, it
is difficult to classify even a routing component of a hydro- Predictive uncertainty and the use of multi-model
logical model. This could in fact be interpreted as physically approaches
based when using the Saint Venant equations with known
boundary conditions, but, at the same time, as stochastic Flood emergency management requires operational decisions
since all the uncertainty (model structure, parameters, initial that may lead to dramatic consequences (economic losses,
and boundary conditions, input and output measurement casualties, etc.) to be taken in real time. Knowing exactly
errors) would be taken to be concentrated in the roughness what would actually happen in the near future (next few
coefficient, which becomes now an uncertain (stochastic) hours or days), emergency managers could safely take, by the
parameter only characterized by its posterior probability book, the best possible decisions on the basis of pre-defined
density. Therefore it is evident that the classification proposed operational plans. Unfortunately, in real situations the
Model System
Input Output
accounts for f(randomness, space, time)
managers cannot choose the right decision due to their aim in this case is to reduce validation uncertainty in order to
uncertainty on the future evolution of events. As described improve model structure, model parameterization and
by Raiffa & Schlaifer (1961) and De Groot (1970), decision parameter estimation. In contrast, predictive uncertainty
theory studied this problem and provided solutions for deci- (Figure 7, right) expresses the uncertainty on what may
sions under uncertainty. These are generally obtained by happen at a future stage when the available information is
minimizing the expected value of an utility function, which provided by a model forecast, which at the time of the
represents either the actual losses (if they can be estimated) forecast is a known value. In this case, it is what will actually
or, more generally, the manager perception of losses, as a occur which is unknown. Predictive uncertainty is expressed
function of a quantity that may occur at a future time, such as in terms of the conditional density fŷjy ðŷjy ¼ y Þ of a future
the discharge or the water stage that will be reached at a given unknown value y given its prediction ŷ ¼ ŷ provided by a
cross-section. This quantity, which is called ‘‘predictand’’ in model, which is evidently known when issued.
the statistical literature, is not known when issuing the In the case of flood forecasting, predictive uncertainty can
forecast, but can be evaluated as an expected value condi- thus be defined as the uncertainty that a decision maker has
tional to one or more model forecasts, to which a predictive on the future evolution of a predictand that he uses to trigger
uncertainty is attached. a specific decision, such as issuing a flood warning or opening
Since the late 1990s, interest in assessing ‘‘uncertainty’’ in the gates of a water detention area or activating a bypass,
models has grown exponentially within the scientific com- conditional on all the model(s) forecasts he can be aware of.
munities of meteorologists and hydrologists. In particular, the Today, basically three approaches are available in the litera-
introduction, on the one hand, of meteorological ensembles, ture for the assessment of predictive uncertainty, the Hydro-
aimed at assessing meteorological meso-scale models fore- logical Uncertainty Processor (HUP) introduced by
casting uncertainty (Molteni et al. 1996; Buizza et al. 1999; Krzysztofowicz (1999), the Bayesian Model Averaging
Stephenson et al. 2005), and on the other hand, of the (BMA) promoted by Raftery (1993) and Raftery et al. (2003,
Hydrological Uncertainty Processor (Krzysztofowicz 1999), 2005), and the Model Conditional Processor (MCP), more
aimed at assessing predictive uncertainty in hydrological recently introduced by Todini (2008). All these approaches,
forecasts, has created the basis for the assessment of ‘‘flood and in particular the last two, aim at assessing and reducing
forecasting uncertainty’’. The interest in this subject is shown predicting uncertainty by combining together one or more
not only by the abundant literature, but also by the establish- than one predictive model.
ment of the International Research Programme HEPEX One of the major benefits arising from the use of the
(2004). Unfortunately, the statistical background of far too multi-model techniques is the possibility of reconciling alter-
many meteorologists and hydrologists was insufficient to native modeling approaches in terms of the estimation of a
really appreciate the definition of ‘‘predictive uncertainty’’ unique predictive uncertainty, conditional upon all the mod-
and its subtle difference from what could be defined as els, which will then be used in the decision making process.
‘‘validation uncertainty’’. This has generated, in the recent This is probably a satisfactory way of replying to Vit Klemes
literature, a large number of papers where the ‘‘validation (1983) wish of taking the maximum advantage from the
uncertainty’’ is estimated instead and is regarded as ‘‘predic- different characteristics of the physics based and the data
tive uncertainty’’, thus increasing the fogginess surrounding driven models.
the subject. Validation uncertainty, which expresses the abil-
ity at emulating (mimicking) reality with a model, is defined
as the uncertainty of the prediction given the observed value FUTURE PERSPECTIVES
to which the prediction refers and is expressed in terms of the
conditional density fŷjy ðŷjy ¼ y Þ of the model forecast ŷ Extending models to ungauged catchments
about a known observed value y ¼ y . As can be easily
understood validation uncertainty (Figure 7, left) is funda- As seen in the previous sections, the evolution of hydrological
mental to assess the quality and to improve a given model: the models proceeded from the simple conceptual models to the
Figure 7 9 Graphical representation of validation and predictive uncertainty. Validation uncertainty (left) is expressed in terms of the conditional density f ðŷjy ¼ y Þ of the model forecast ŷ
about a known observed value y ¼ y*, while predictive uncertainty (right) is expressed in terms of the conditional density f ðŷjy ¼ y Þ of a future unknown value y given its prediction
ŷ ¼ ŷ provided by a model, which is evidently known when issued.
more comprehensive and physically based ones, gradually lumped models can only be derived through an averaging
introducing more detailed equations in the effort of better (lumping) process conditional upon a correct representation
reproducing the complex reality (Singh 1988). At the same of these additional nonlinear phenomena at the catchment
time, several lumped models have been proposed, which tend scale, which are automatically resolved when using distrib-
to represent reality with widely different parameterisations of uted finer scales models but are not resolved by the present
the infiltration, soil saturation, drainage, runoff formation generation of lumped models. These phenomena are the
processes. But, given the present (as well as near-future hysteretic dependency of the saturated area on the mean
expected) availability of detailed spatial information such as soil water volume, also found by several other authors
for instance DTMs at tenths of metres, radar precipitation (Mishra & Seth 1996; Niedzialek & Ogden 2004; O’Kane &
estimates at few hundred metres, the basic question in order Flynn 2007; Norbiato & Borga 2008), and the exfiltration from
to be able to accurately extrapolate models to ungauged the soil which continues after the end of a rainfall event
catchments, is whether or not it is possible or worthwhile (Liu & Todini 2002). Owing to these nonlinear effects, one
to directly set up lumped hydrological models encapsulating has to realize that, currently, only the distributed models can
the physical properties and processes that can be described at be used if one wants to use the available distributed maps to
the different scales without the need of setting up distributed extrapolate on physical grounds the hydrological parameters
models. The reason for this question lies in the fact that the to ungauged catchments, while lumped version retaining
statistical approaches to extrapolate information from one the physical information can be successively derived via
catchment to another have mostly failed at providing efficient distributed modelling simulation.
models that can benefit from the large amount of available Thus, interesting research perspectives lie in the study
distributed and detailed geomorphological information, today (from their experimental analysis to their conceptualisation in
available through remote sensing. hydrological models) of the macro nonlinear phenomena that
In a recent paper, Martina et al. (2009) showed that, can be observed when aggregating from the pixel to the
unfortunately, the physical properties of the basic soil and catchment scale, but not directly at this latter scale, since
surface processes (such as hydraulic conductivity, soil moist- they cannot be not resolved in lumped form. Furthermore,
ure content, slope, overland friction) can only be retained at additional research should be concentrated in the derivation
finer spatial scales (up to few hundred metres), due to the of theoretical results that could overcome this need for
inherent topological nonlinearities of the hydrological pro- distributed modelling simulations, by more directly finding
cesses at the catchment scale that hide the small scale ones, the nonlinear relations occurring through different scales as a
which directly depend on the parameters. Physics-based function of the different available maps.
Linking hydrological models to LAMs for real-time Bayesian inferential schemes based on the uncertainty multi-
and flash flood forecasting model processors described in a previous section.
Another important area of development is the use of hydro- Linking hydrological models to GCMs for climate
logical models as part of a chain aimed at transforming studies
meteorological quantitative precipitation forecasts (QPFs) into
flood forecasts at given river cross sections. The use of QPFs is The pressure due to climate changes is also motivating a wide
common when one wants to extend the forecast beyond the variety of research activities and in particular the incorpora-
characteristic concentration time of a catchment. Several ten- tion of hydrologic models into the General Circulation
tative case studies have been implemented in the recent past, Models (GCMs). The importance of a more realistic repre-
particularly within EU funded projects such pas EFFS (2003), sentation of the water balance at the catchment scale was
which have not lead to satisfactory results. In addition, the use recognized by Dümenil & Todini (1992) who incorporated the
of meteorological ensemble predictions, namely predictions ARNO model (Todini 1996, 2002) in the ECHAM GCM in
based not on a single future precipitation scenario, but on a place of the Manabe (1969) on–off bucket, followed by Liang
set of 20–50 scenarios (members of the ensemble), has addi- et al. (1996a, b) who used the VIC model (Wood et al. 1992) in
tionally complicated the problem. the GFDL GCM for the same purpose.
As one can see from Figure 8, for an example of real-time One of the reasons that motivated the interest of climato-
flood forecast at Ponte Spessa on the Po river in Italy, logists in using more realistic surface schemes, rather than the
ensemble QPF forecasts tend to generate ensembles of simple on–off bucket, to represent the formation of runoff is
predicted discharges and water levels that hardly embed tied to the possibility of using river discharges, now available
the observed ones. This is due to the fact that meteoro- for most of the largest rivers of the world, to assess the
logical ensembles represent an envelope of model, parameter response of the GCMs not only in terms of average climato-
and boundary conditions uncertainty, the validation uncer- logy, but also in terms of actual monthly water volumes
tainty, instead of the predictive uncertainty, namely the uncer- delivered to the oceans.
tainty on future values given the model forecasts (Todini 2008). What appeared immediately evident was the need for a
Therefore, current research, particularly within the lumped hydrological model that could be applied to all the
frame of HEPEX, aims at finding the most appropriate ways GCM pixels which were, at that time, of the order of magni-
of making use of QPF ensembles by incorporating them into tude of 100 100 km2. Neither the ARNO nor the VIC
schemes could be extended on physical grounds to the differ-
ent pixels, taken as ungauged catchments, due to the lack of
physical meaning of their parameters. This motivated the
interest in possible hydrological model parameterizations
which parameters could be derived from digital elevation
maps, land use maps and soil type maps that are now available
for the entire globe at pixels of the order of 1 1 km2.
As described in a previous section, results in this area are
promising but not yet conclusive and additional research is
still needed.
CONCLUSIONS
Figure 8 9 Observed discharges (solid line) together with the ones simulated using
A significant advance has been made in terms of quantitative
observed rainfall (dashed line) and the ones resulting from ensemble
forecasts of QPF (thin solid lines) for the Po river at Ponte Spessa. representation of hydrological phenomena from the Rational
Method to the nowadays available distributed physics-based Dawdy, D. R. & O’Donnell, T. 1965 Mathematical models of
catchment behavior. J. Hydraul. Div.- ASCE, HY4 91, 123–137.
models. Nonetheless, there remains much scope in pursuing
Dawson, C. W. & Wilby, R. L. 2001 Hydrological modelling using
research into a number of interesting questions and problems artificial neural networks. Prog. Phys. Geog. 25, 80–108.
under the pressure of climate change and the need to cor- de Finetti, B. 1975 Theory of Probability, Vol. 2. Wiley, Chichester, UK.
rectly assess predictive uncertainty and the possibility of De Groot, M. H. 1970 Optimal Statistical Decisions, McGraw-Hill,
New York.
reconciling alternative modelling approaches.
De Roo, A. P. J., Wesseling, C. G. & Van Deursen, W. P. A. 1998
This paper, which aimed to present a historical overview Physically based river modelling within a GIS. The LISFLOOD
and future perspective of hydrological catchment modelling, model. Proc. 3rd Intl Conf. on Geo-Computation. In: Geo-Com-
concludes with the hope that future generations of hydrolo- putation CD-ROM produced by R. J. Abrahart. http://www.geo-
computation.org/1998/06/gc_06.htm.
gists will enthusiastically approach the new emerging
De Roo, A. P. J., Wesseling, C. G. & Van Deursen, W. P. A. 2000
research needs. Physically-based river basin modelling within a GIS: The LIS-
FLOOD model. Hydrol. Process. 14, 1981–1992.
Dooge, J. C. I. 1973 Linear Theory of Hydrologic Systems. Technical Bull.
No. 1468, United States Department of Agriculture, Washington,
REFERENCES DC.
Dümenil, L. & Todini E. 1992 A rainfall-runoff scheme for use in the
Abbott, M. B., Bathurst, J. C., Cunge, J. A., O’Connell, P. E. & Hamburgh Climate Model. In Advances in Theoretical Hydrology,
Rasmussen, J. 1986a An introduction to the European Hydrologi- a Tribute to James Dooge O’Kane, J. P.(Ed.), European Geophy-
cal System – Système Hydrologique Européen, "SHE", 1: History sical Society Series of Hydrological Sciences, 1. Elsevier, Amster-
and philosophy of a physically-based, distributed modelling dam, pp. 129–157.
system. J. Hydrol. 87, 45–59. Dunne, T. 1978 Field studies of hillslope flow processes, Chapter 7. In:
Abbott, M. B., Bathurst, J. C., Cunge, J. A., O’Connell, P. E. & Kirkby, M. J. (Ed.), Hillslope Hydrology. Wiley, New York, pp.
Rasmussen, J. 1986b An introduction to the European Hydrologi- 227–293.
cal System – Système Hydrologique Européen, "SHE", 2: Struc- EFFS 2003 An European Flood Forecasting System, http://effs.
ture of physically-based, distributed modelling system. J. Hydrol. wldelft.nl/.
87, 61–77. Ewen, J., Parkin, G. & O’Connell, P. E. 2000 . J. Hydrol. Engng. 5, 250–258.
Bertalanffy, L. v. 1968 General Systems Theory: Foundations, Develop- Francés F., Vélez, J. I. & Vélez, J. J. 2007 Split-parameter structure for
ment, Applications. George Braziller, New York. Revised edition the automatic calibration of distributed hydrological models.
1976: ISBN 0-8076-0453-4. J. Hydrol. 332(1), 226–240.
Beven, K. J. & Binley, A. M. 1992 The future of distributed models: Franchini, M., Wendling, J., Obled, Ch. & Todini, E. 1996 Physical
model calibration and uncertainty prediction. Hydrol. Process. 6, interpretation and sensitivity analysis of the TOPMODEL.
279–298. J. Hydrol. 175, 293–338.
Beven, K. J. & Freer, J. 2001 Equifinalilty, data assimilation, and Freeze, R. A. & Harlan R. L. 1969 Blueprint for a physically-based
uncertainty estimation in mechanistic modeling of complex envir- digitally-simulated hydrologic response model. J. Hydrol. 9,
onmental systems using the GLUE methodology. J. Hydrol. 249, 237–258.
11–29. Garcı̀a-Bartual, R. 2002 Short-term river forecasting with neural net-
Beven, K. J. & Kirkby, M. J. 1979 A physically based, variable contribut- works. Integrated assessment and decision support. Proc. 1st
ing area model of basin hydrology. Hydrol. Sci. Bull. 24, 43–69. Biennial Meeting of the International Environmental Modelling
Box, G. E. P. & Jenkins, G. M. 1970 Time Series Analysis Forecasting and Software Society, Vol. 2, pp. 160–165.
and Control, Holden Day, San Francisco. Gaume, E. & Gosset, R. 2003 Over-parameterisation, a major obstacle
Buizza, R., Miller, M. & Palmer, T. N. 1999 Stochastic representation of to the use of artificial neural networks in hydrology? Hydrol.
model uncertainties in the ECMWF Ensemble Prediction System. Earth Syst. Sci. 7, 693–706.
Q. J. R. Meteorol. Soc. 125, 2887–2908. Gupta, V. K. & Sorooshian, S. 1983 Uniqueness and observability of
Burnash, R. J. C., Ferral, R. L. & McGuire, R. A. 1973 A General conceptual rainfall-runoff model parameters: the percolation pro-
Streamflow Simulation System - Conceptual Modelling for Digital cess examined. Water Resour. Res. 19, 269–276.
Computers. Report by the Joint Federal State River Forecasts HEPEX, Hydrological Ensemble Prediction Experiment 2004 http://
Center, Sacramento, CA. www.ecmwf.int/newsevents/meetings/workshops/2004/HEPEX/
Chow, V. T., Maidment, D. R. & Mays, L. V. 1988 Applied Hydrology. index.html.
McGraw-Hill. Kalinin, G. P. & Milyukov, P. I. 1957 O raskete neustanovivshegosya
Crawford, N. H. & Linsley, R. K. 1966 Digital Simulation in Hydrology, dvizhenia vody v otkrytykh ruslakh (On the computation of
Stanford Watershed Model IV, Tech. Rep. 39. Department of Civil unsteady flow in open channels). Meteorologiya i gidologiya
Engineering, Stanford University. zhuzurnal 10, 10–18 Leningrad (in Russian).
Klemes, V. K. 1983 Conceptualization and scale in hydrology. J. Hydrol. Raiffa, H. & Schlaifer, R. 1961 Applied Statistical Decision Theory. MIT
65, 1–23. Press, Cambridge, MA.
Krzysztofowicz, R. 1999 Bayesian theory of probabilistic forecasting Raftery, A. E. 1993 Bayesian model selection in structural equation
via deterministic hydrologic model. Water Resour. Res. 35, models. In: Testing Structural Equation Models Bollen, K. A. &
2739–2750. Long, J. S. (Eds.), Sage, Newbury Park, CA, pp. 163–180.
Kuichling, E. 1889 The relation between the rainfall and the discharge of Raftery, A. E., Balabdaoui, F., Gneiting, T. & Polakowski, M. 2003
sewers in populous districts. Am. Soc. Civil Eng. Trans. 20, 1–56. Using Bayesian model averaging to calibrate forecast ensembles,
Jakeman A. J., Letcher, R. A. & Norton, J. P. 2006 Ten iterative steps in Tech. Rep. 440, Department of Statistics, University of
development and evaluation of environmental models. Environ. Washington, Seattle.
Modell. Softw. 21, 602–614. Raftery, A. E., Gneiting, T., Balabdaoui, F. & Polakowski, M. 2005 Using
Liang, X., Lettenmaier, D. P. & Wood, E. F. 1996a One-dimensional Bayesian model averaging to calibrate forecast ensembles. Mon.
statistical dynamic representation of subgrid spatial variability of Weather Rev. 133, 1155–1174.
precipitation in the two-layer variable infiltration capacity model. Refsgaard J. C., Henriksen H. J., Harrar W. G., Scholten H. & Kassahun
J. Geophys. Res. 101(D16), 21, 403–21, 422. A. 2005 Quality assurance in model based water management –
Liang, X., Wood, E. F. & Lettenmaier, D. P. 1996b Surface soil moisture review of existing practice and outline of new approaches.
parameterization of the VIC-2L model: Evaluation and modifica- Environ. Modell. Softw. 20, 1201–1215.
tions. Global Planet Change 13, 195–206. Refsgaard, J. C. & Storm, B. 1995 Chapter 23: MIKE SHE. In: Computer
Liu, Z. & Todini, E. 2002 Towards a comprehensive physically-based Models of Watershed Hydrology, Singh, V. P. (Ed.), Water
rainfall-runoff model. Hydrol. Earth Syst. Sci. 6, 859–881. Resources Publications, Littleton, CO.
Lloyd-Davies, D. E. 1906 The elimination of stormwater from sewerage Rockwood, D. M. 1964 Streamflow synthesis and reservoir regulation.
systems. Inst. Civil Eng. Proc., 164, 41–67.
US Army Engineer Division, North Pacific, Portland, OR, Engi-
Manabe, S. 1969 Climate and Ocean circulation: 1. The atmospheric
neering Studies Project 171, Technical Bulletin No. 22.
circulation and the hydrology of the earth’s surface. Mon. Weather
Saulnier G. M. & Datin R. 2004 Analyical solution to a bias in
Rev. 97, 739–774.
the TOPMODEL framework balance. Hydrol. Process. 18,
Martina, M. L. V., Todini, E. & Liu, Z. 2009 Can physically meaningful
1195–1218.
properties and parameters be directly retained in lumped hydro-
Shamseldin, A. Y. 1997 Application of neural network technique to
logical models? Paper submitted to J. Hydrol., under review.
rainfall-runoff modelling. J. Hydrol. 199, 272–294.
Mishra, S. K. & Seth, S. M. 1996 Use of hysteresis for defining the
Sherman, L. K. 1932 Streamflow from rainfall by the unit graph method.
nature of flood wave propagation in natural channels. Hydrologi-
Engng News Record 108, 501–505.
cal Sci. J. 41(2), 153–170.
Singh, V. P. 1988 Hydrologic Systems: Rainfall-Runoff Modelling. Vols.
Molteni, F., Buizza, R., Palmer, T. N. & Petroliagis, T. 1996 The
1–2. Prentice Hall, Englewood Cliffs, NJ.
ECMWF Ensemble Prediction System: methodology and valida-
Singh, V. P. & Woolhiser, D. A. 2002 Mathematical modeling of
tion. Q. J. R. Meteorol. Soc. 122, 73–119.
watershed hydrology. J. Hydrol. Engng. 7, 270–292.
Moore, R. J. & Clarke, R. T. 1981 A distribution function approach to
Sorooshian, S. & Gupta, V. K. 1983 Automatic calibration of conceptual
rainfall-runoff modelling. Water Resour. Res. 17, 1367–1382.
rainfall-runoff models: The question of parameter observability
Mulvany, T. J. 1850 On the use of self registering rain and flood gauges.
and uniqueness. Water Resour. Res. 19, 260–268.
Inst. Civ. Eng. Proc. 4, 1–8. Dublin, Ireland.
Stephenson, D. B., Coelho C. A. S., Balmaseda M. & Doblas-Reyes F. J.
Nash, J. E. 1958 The form of the instantaneous unit hydrograph. IUGG
2005 Forecast Assimilation: A unified framework for the combina-
General Assembly of Toronto, Vol. III – IAHS Publ. 45, 114–121.
Nash, J. E. 1960 A unit hydrograph study with particular reference to tion of multi-model weather and climate predictions. Tellus A 57,
British catchments. Proc. Inst. Civil. Eng. 17, 249–282. 253–264.
Natale, L. & Todini, E. 1976a A stable estimation for linear models – 1. Sugawara, M. 1967 The flood forecasting by a series storage type model.
Theoretical development and Monte Carlo experiments. Water Int. Symp. Floods and their Computation, IAHS Publ. 85, IAHS
Resour. Res. 12, 667–671. Press, Wallingford, UK. pp.1–6.
Natale, L. & Todini, E. 1976b A stable estimator for linear models – 2. Sugawara, M. 1995 Tank model. Chapter 6. In Computer Models of
Real world hydrologic applications. Water Resour. Res. 12, Watershed Hydrology, Singh, V. P. (Ed.), Water Resources
672–675. Publications, Littleton, CO.
Niedzialek, J. M. & Ogden, F. L. 2004 Numerical investigation of Todini, E. 1988 Rainfall-runoff modeling – Past, present and future.
saturated source area behavior at the small catchment scale. J. Hydrol. 100(1), 341–352.
Adv. Water Resour. 27(9), 925–936. Todini, E. 1995 New trends in modeling soil processes from hillslope to
Norbiato, D. & Borga, M. 2008 Analysis of hysteretic behaviour of a GCM scales. In The Role of Water and the Hydrological Cycle in
hillslope-storage kinematic wave model for subsurface flow, Adv. Global Change, Oliver, H. R. & Oliver, S. A.(Eds.), NATO ASI
Water Resour. 31(1), 118–131. Series. Vol. I 31. Springer-Verlag, Berlin-Heidelberg-New York,
O’Kane, J. P. & Flynn, D. 2007 Thresholds, switches and hysteresis from pp. 317–348.
the pedon to the catchment scale: a non-linear system theory. Todini E. 1996 The ARNO Rainfall-Runoff model. J. Hydrol. 175, 339–
Hydrol. Earth Syst. Sci. 11(1), 443–459. 382.
Todini, E. 2002 The ARNO model. Chapter 16. In Mathematical Models WMO 1992 Simulated real-time intercomparison of hydrological mod-
of Large Watershed Hydrology, Singh, V. P., Frevert, D. K. & els. Operational Hydrology Report No. 38. WMO Publ. No. 779,
Meyer, S. P. (Eds.),. Water Resources Publications, Littleton, CO, 241 pp. Geneva, Switzerland.
pp. 687–716. Wood, E. P., Lettenmaier, D. P. & Zartarian, V. G. 1992 A Land-surface
Todini, E. 2007. Hydrological catchment modelling: past, present and hydrology parameterization with subgrid variability for general
future. Hydro. Earth Syst. Sci. 11(1), 468–482. circulation models. J. Geophys. Res. 97(D3), 2717–2728.
Todini, E. 2008 A model conditional processor to assess predictive un- Wooding, R. A. 1965a A hydraulic model for the catchment-stream
certainty in flood forecasting. Intl J. River Basin Mgmnt 6(2), problem, I. Kinematic wave theory. J. Hydrol. 3(3, 4), 254–267.
1–15. Wooding, R. A. 1965b A hydraulic model for the catchment-stream
Todini, E. & Ciarapica, L. 2002 The TOPKAPI model. Chapter 12. In problem, II. Numerical solutions. J. Hydrol. 3(3, 4), 268–282.
Mathematical Models of Large Watershed Hydrology, Singh, V. P., Wooding, R. A. 1966 A hydraulic model for the catchment-stream pro-
Frevert, D. K. & Meyer, S. P. (Eds.), Water Resources Publica- blem, III. Comparison with runoff observations. J. Hydrol. 4,
tions, Littleton, CO, pp. 471–506. 21–37.
Vélez, J. I. 2001 Desarrollo de un modelo hidrológico conceptual y Woolhiser, D. A. & Liggett, J. A. 1967 Unsteady, one-dimensional flow
distribuido orientado a la simulación de las crecidas. Universidad over a plane – the rising hydrograph. Water Resour. Res. 3, 753–771.
Politécnica de Valencia, Departamento de Ingenierı́a Hidráulica y Young, P. C. 2001 Data-based mechanistic modelling and validation of
Medio Ambiente. Ph.D. Thesis (in Spanish). rainfall-flow processes. In Model Validation: Perspectives in
WMO 1975 Intercomparison of conceptual models used in operational Hydrological Science, Anderson, M. G. & Bates, P. D. (Eds.),
hydrological forecasting. Operational Hydrology Report No. 7, Wiley, Chichester, UK, pp. 117–161.
WMO Publ. No. 429, 172 pp. Geneva, Switzerland. Young, P. C. 2002 Advances in real-time flood forecasting. Phil. Trans.
WMO 1986 Intercomparison of models of snowmelt runoff. Operational R. Soc. London, A. 360, 1433–1450.
Hydrology Report No. 23. WMO Publ. No. 646, Geneva, Zhao, R. J. 1977 Flood Forecasting Method for Humid Regions of China.
Switzerland. East China College of Hydraulic Engineering, Nanjing, China.
First received 14 August 2009; accepted in revised form 21 December 2009. Available online February 2011