Professional Documents
Culture Documents
Inverse Problem in Hydrogeology
Inverse Problem in Hydrogeology
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/225344454
CITATIONS READS
226 98
5 AUTHORS, INCLUDING:
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Andres Alcolea
letting you access and read them immediately. Retrieved on: 08 January 2016
Inverse problem in hydrogeology
Jesffls Carrera · Andrs Alcolea · Agustn Medina ·
Juan Hidalgo · Luit J. Slooten
Abstract The state of the groundwater inverse problem is d’tat (charges hydrauliques, concentrations) des pro-
synthesized. Emphasis is placed on aquifer characteriza- prits de l’aquifre. A cause de ces difficults, le cali-
tion, where modelers have to deal with conceptual model brage ne peut Þtre spar du processus de modlisation,
uncertainty (notably spatial and temporal variability), comme c’est le cas dans d’autres cas de figure. Par ail-
scale dependence, many types of unknown parameters leurs, il peut Þtre vu comme une des tapes dans le pro-
(transmissivity, recharge, boundary conditions, etc.), cessus de dtermination du comportement de l’aquifre. Il
nonlinearity, and often low sensitivity of state variables est montr que les mthodes d’valuation des paramtres
(typically heads and concentrations) to aquifer properties. actuels ne diffrent pas si ce n’est dans les dtails des
Because of these difficulties, calibration cannot be sepa- calculs informatiques. Il est montr qu’il existe une large
rated from the modeling process, as it is sometimes done panoplie de techniques d ‹inversion : codes de calcul
in other fields. Instead, it should be viewed as one step in utilisables par tout-un-chacun, accommodation de la va-
the process of understanding aquifer behavior. In fact, it is riabilit via la gostatistique, incorporation d’informa-
shown that actual parameter estimation methods do not tions gologiques et de diffrents types de donnes
differ from each other in the essence, though they may (temprature, occurrence, concentration en isotopes, ge,
differ in the computational details. It is argued that there etc.), dtermination de l’incertitude. Vu ces dveloppe-
is ample room for improvement in groundwater inversion: ments, la calibration automatique facilite normment la
development of user-friendly codes, accommodation of modlisation. Par ailleurs, il est souhaitable que son uti-
variability through geostatistics, incorporation of geo- lisation devienne une pratique standardise.
logical information and different types of data (tempera-
ture, occurrence and concentration of isotopes, age, etc.), Resumen Se sintetiza el estado del problema inverso en
proper accounting of uncertainty, etc. Despite this, even aguas subterrneas. El nfasis se ubica en la caracteriza-
with existing codes, automatic calibration facilitates cin de acuferos, donde los modeladores tienen que en-
enormously the task of modeling. Therefore, it is con- frentar la incertidumbre del modelo conceptual (princi-
tended that its use should become standard practice. palmente variabilidad temporal y espacial), dependencia
de escala, muchos tipos de parmetros desconocidos
Rsum L’tat du problme inverse des eaux souter- (transmisividad, recarga, condiciones limitantes, etc), no
raines est synthtis. L’accent est plac sur la caractri- linealidad, y frecuentemente baja sensibilidad de varia-
sation de l’aquifre, o les modlisateurs doivent jouer bles de estado (tpicamente presiones y concentraciones) a
avec l’incertitude des modles conceptuels (notamment la las propiedades del acufero. Debido a estas dificultades,
variabilit spatiale et temporelle), les facteurs d’chelle, no puede separarse la calibracin de los procesos de
plusieurs inconnues sur diffrents paramtres (transmis- modelado, como frecuentemente se hace en otros campos.
sivit, recharge, conditions aux limites, etc.), la non li- En su lugar, debe de visualizarse como un paso en el
narit, et souvent la sensibilit de plusieurs variables proceso de entendimiento del comportamiento del acu-
fero. En realidad, se muestra que los mtodos reales de
estimacin de parmetros no difieren uno del otro en lo
Received: 10 June 2004 / Accepted: 2 November 2004 esencial, aunque s pueden diferir en los detalles com-
Published online: 25 February 2005 putacionales. Se discute que existe amplio espacio para la
Springer-Verlag 2005 mejora del problema inverso en aguas subterrneas: des-
arrollo de cdigos amigables al usuario, acomodamiento
de variabilidad a travs de geoestadstica, incorporacin
J. Carrera ()) · A. Alcolea · A. Medina · J. Hidalgo · L. J. Slooten de informacin geolgica y diferentes tipos de datos
School of Civil Engineering, (temperatura, presencia y concentracin de istopos,
Technical University of Catalonia,
1-3, Jordi Girona Street, Building D2, 08034 Barcelona, Spain
edad, etc), explicacin apropiada de incertidumbre, etc. A
e-mail: jesus.carrera@upc.edu pesar de esto, affln con los cdigos existentes, la calibra-
Tel.: +34-93-4016890 cin automtica facilita enormemente la tarea de mode-
Fax: +34-93-4017251
lado. Por lo tanto, se sostiene que su uso debera de sivity and hydraulic conductivity will be used inter-
convertirse en prctica standard. changeably).
– Different types of parameters. While efforts are often
Keywords Inverse problem · Aquifer · Groundwater · concentrated on transmissivity, other parameters (re-
Modeling · Parameter estimation charge, boundary fluxes, etc.) may be equally uncer-
tain and relevant.
– Scale dependence. Parameters measured in the field
Introduction often represent a small portion of the aquifer. As a
result, they are qualitative and quantitatively different
In broad terms, inverse modeling refers to the process of from what is needed in the model.
gathering information about the model from measure- – Model uncertainty. Geometry of the aquifer and het-
ments of what is being modeled. This includes two related erogeneity patterns are controlled by the geology,
concepts: model identification and parameter estimation. which is never known accurately.
The latter will be used here as being synonymous with – Low sensitivity. Depending on the problem, state
calibration. Model identification applies to methods to variables may display low sensitivity to model pa-
find the nature (features) of the model, such as the gov- rameters (i.e. their information content is low). In
erning equations, boundary conditions, time regime, or particular, heads (the most frequent and sometimes
heterogeneity patterns. Parameter estimation, instead, is unique type of measurement) sometimes contain little
restricted to assigning values to the properties character- information about hydraulic conductivity.
izing those features.
The above definitions apply to groundwater modelling Because of the above features, the aquifer model predic-
without much modification. In fact, it can be argued that tions are highly uncertain. Moreover, parameter estima-
some sort of inversion has always been standard practice tion cannot be formulated as clearly as in other fields.
in hydrogeology. For example, hydrogeologists perform That is possible for pumping test interpretation, where
pumping tests to characterize aquifers. These tests are one can indeed take heads straight from the test, enter
interpreted by, first, identifying the most appropriate them in a code and derive parameter values after a
model (on the basis of actual measurements and geology) moderately qualitative model analysis. In aquifers, one is
and, then, estimating model parameters by curve fitting. forced to cast inversion as one step of the modeling
This can be facilitated by automatic procedures, which process. Many of these features are shared by ground-
have become rather standard in fields where analytical water’s sister science, surface water hydrology, where
solutions are scarce, such as the unsaturated zone (Kool many of these issues also have been addressed (Beven
and Parker 1988; Hollenbeck and Jensen 1998). Auto- 1993; Gupta et al. 1999).
matic calibration has also become standard in fields Another consequence of the singularities of ground-
where appropriate tests can be performed to characterize water inversion is its relative isolation. Many inversion
spatial variability: medicine, where it is termed “scan- methods have been developed independently from those
ning” or “tomography” (Rudin et al. 1999; R
hli et al. in other fields. The earliest methods were based on simply
2002); geophysics (Sambridge and Mosegaard 2002); and substituting heads, assumed to be known, into the flow
many others. However, groundwater modeling displays equation, which leads to a first order partial differential
several peculiarities that collectively sets it apart from equation in transmissivity (Stallman 1956). This method,
modeling in other fields: termed “direct” by Neuman (1973), is relatively simple to
understand and has been widely used after Nelson (1960,
– Cost. Groundwater models are relatively expensive to 1961). In fact, it allowed deriving transmissivities from
run. They require building large systems of equations flow nets based on head measurements (Bennet and
(to get an accurate and realistic picture of the system) Meyer 1952). Unfortunately, this approach has several
that need to be solved for each model run. In addition, drawbacks. First, it requires knowing heads (and recharge,
advances in computer science are used to increase the storage coefficient and boundary conditions) over the
quality of the model performance rather than to reduce whole domain in space and time. This can only be
execution times. achieved through interpolation, which introduces
– Time dependence. State variables such as heads and smoothing and errors, so that the estimated transmissivity
concentrations are time dependent. Many flow prob- values become somewhat artificial. Second, it is unstable
lems are essentially under steady state conditions. Yet, (small errors in heads cause large errors in transmissivi-
even in this case, the information might be contained ty). To overcome the first problem, most recent inversion
in temporal fluctuations of heads, thus deserving methods use what Neuman (1973) termed the indirect
transient modeling. approach, which consists of acknowledging that mea-
– Heterogeneity. Values of hydraulic conductivity, K, surements contain errors and finding the hydraulic prop-
which is often the most dominant hydraulic property, erties that minimize these errors. That is, parameters are
may vary over several orders of magnitude. The same found by minimizing an objective function, which may
can be said about transmissivity, which is essentially become a huge computational task. To overcome insta-
equivalent to K in 2D models (hereinafter transmis- bilities, a number of approaches can be taken: adding a
regularization term to the objective function to dampen calibration should be viewed as a necessary step in
unwarranted oscillations, considering additional types of modeling and the paper is structured to mimic the mod-
data, reducing the number of parameters to be estimated eling process (Fig. 1). The following section of this paper
without losing the ability to reproduce spatial variability, is devoted to outlining the roles of knowledge and data in
etc. These issues (computation, stabilization, different the definition of the conceptual model, arguably the most
types of data, spatial variability) have remained the focus frequent source of errors in groundwater modeling (Bre-
of much research. Reviews of them are presented by Yeh dehoeft 2004). Methods for representing aquifer proper-
(1986), Carrera (1987), Kool et al. (1987), McLaughlin ties in terms of a hopefully small number of unknowns to
and Townley (1996) and de Marsily et al. (1999). be optimized (the so called model parameters) are pre-
Therefore, a thorough review of the state of the art in sented in the section entitled “What is to be estimated?”.
aquifer modeling inversion would be superfluous. The next natural step is to optimize (calibrate) the value
In this context, the objective of this paper is to describe of the model parameters. Actual calibration methods and
the current state of inverse modeling for aquifer charac- codes are presented in the sections, “How to estimate the
terization. Emphasis is placed on seeking similarities of model parameters, p”; and “How codes work”. Once the
existing methods and on discussing how they are used, parameters have been obtained, one needs to check the
rather than describing them exhaustively. It is argued that quality of the resulting model, which is addressed in the
ft may represent time varying recharge obtained by mass account for prior information on the parameters has been
balance in the soil and pi would be an unknown factor if q recognized. Traditionally, direct measurements were used
is areal recharge). for the first term in Eq. (7), but were disregarded during
The main advantage of zonation is its generality and the inversion process. Doherty (2003) includes a regu-
flexibility to accommodate the geological information larization criterion penalizing non-homogeneity of the
(e.g., zones may represent geological units or portions of model parameters and he does not use prior information.
them). It should be stressed, however, that zonation does Recently, Kowalsky et al. (2004) used a plausibility term
not preclude the use of geostatistics. In fact, Clifton and for the first time in the context of the pilot points method,
Neuman (1982) used zonation coupled to kriging. but the role of this term is not explored.
While the zones need not be large, the original spirit of
zonation is to reduce the dimension of p, while ensuring Conditional simulation
geological consistency (Stallman 1956). In fact, Carrera et Methods described so far are implicitly based on seeking
al. (1993b) argue that, when available, geological infor- some sort of optimal estimation. As will be seen in sub-
mation about parameter variability is so compelling (in sequent sections, it is sometimes preferable to seek
the sense that it can be included deterministically) that it equally likely simulations of q(x, t). Neuman and
overcomes the advantages of conventional geostatistics. Wierenga (2003) present a comprehensive strategy in the
Zonation is sometimes criticized as rigid. Hence, it is not context of simulation. Alternatives that have been used
surprising that efforts have been made to optimize the include the self calibration approach of Sahuquillo et al.
geometry of zones. A particularly appealing one is “ge- (1992), Gmez-Hernndez et al. (1997) and Capilla et al.
omorphing” where the geometry of zones is derived (1998), in which q0(x) represents a simulation of the log-T
during the calibration process (Roggero and Hu 1998). field conditioned by all (soft and hard) available infor-
mation and the terms ai (x, t)pi represent perturbations
Point estimation imposed by a set of master points. Another possibility is
It can be viewed as the limiting case of zonation, as the to express q(x) as a linear combination of random func-
size of zones tends to zero (actually, to the element or cell tions (Roggero and Hu 1998; Hu 2002), where ai0 s re-
size). The formalism of Eq. (4) can still be used (e.g. present conditional simulations and the model parameters
Meier et al. 2001). However, the dimension of the pa- are simply weights of those simulations.
rameter space becomes so large that it may be more ap- In summary, first, a representation of the variability of
propriate to seek alternative formulations (Kitanidis and the hydraulic properties is necessary and, second, the
Vomvoris 1983; Dagan 1985; McLaughlin and Townley most common parameterization schemes can be written
1996). These will be outlined in the next section. using Eq. (4) or, in spite of their quite different appear-
ance. The question now is how to find p.
Heuristic interpolation functions
The interpolation functions ai in Eq. (4) can be chosen
arbitrarily. Different types have been chosen, including How to estimate the model parameters, p
finite elements (Yeh and Yoon 1981), Ridge functions
(Mantoglou 2003), or others. These approaches offer The estimation problem deals with the concept of “best”
significant flexibility, but it is not clear how to define set of model parameters. A good question is what does
prior information on model parameters. “best” exactly mean? There is no perfect answer to this
question. In fact, there may not be a single set of model
Pilot points parameters leading to a “good” representation of reality,
In this case, p represents the unknown values of the which motivates conditional simulation methods Gmez-
property q at a set of (pilot) points. The method, originally Hernndez et al. (1997) that aim at finding a collection of
devised by de Marsily et al. (1984), has become very equally probable parameters sets. Following is a de-
popular (e.g. Ramarao et al. 1995; Vesselinov et al. 2001; scription of the most widely used methods.
Hernandez et al. 2003), to the point of becoming the
standard for non-linear geostatistical inversion. It can be Optimization methods
considered as a generalization of Eq. (4) for the case in In these methods, the parameter set is defined as the
which measurements at the pilot points are uncertain. In minimum of an objective function. Originally, the ob-
this case, Eq. (4) can be written as: jective function was motivated toward ensuring a good
dimðzp Þ zp
match between computed and measured data (model fit).
qðxÞ ¼ Si¼1 li ðxÞzpi þ SNj¼1 lPP
j ðxÞpj ð7Þ This agreement (performance criterion in this case) can be
zp measured as the squared difference between computed
where li and lPP i are the (co-) kriging weights for mea- and measured heads. This squared difference is, in
surements and pilot points, respectively. Comparing Eqs. mathematical terms, a norm (called L2 norm). However,
(7) and (4), it is clear that q0 is the first term on the right other norms are available and some have been used in
hand side and aqi ðx; tÞ are equal to lPP i ðxÞ. The main ad- groundwater. For example, errors can be quantified as the
vantage of the method is its relative ease and flexibility. absolute value of the difference between measured and
Surprisingly, it has not been until 2004 that the need to computed values (L1 norm). Each norm has its own ad-
vantages: the L2 norm is more sensitive to outliers than discussed later (sensitivity equations). Assuming this
L1, which is more difficult to work with than the L2 norm equation to hold, Eq. (2) can be generalized to obtain
(see Woodbury et al. (1987) for details and Xiang et al. parameters as the conditional expectation of p given h
(1992) for an application to modeling). Defining a per- (Dagan 1985; Rubin and Dagan 1987). For this purpose,
formance criterion leads naturally to a minimization one needs to define the cross-covariances between data
problem, i.e., to seek the model parameters that make the and parameters Qph and the covariance matrix of data
chosen norm minimum. Using the L2 norm, the objective Qhh, given by:
function to be minimized can be written as
Qph ¼ Cpp Jthp ð12Þ
t
Fh ¼ ðh h Þ C1
h ðh
h Þ ð8Þ
Qhh ¼ Jhp Cpp Jthp þ Ch ð13Þ
where C1 h is a matrix of weights. When Fh is used as
objective function, the problem often becomes ill-posed. Then
The solution becomes unstable, i.e. different parameter
sets lead to very similar values of Fh. Sometimes, this is ^p ¼ Eðpjp ; h Þ ¼ p þ Qph Q1
hh ðh hðp ÞÞ ð14Þ
shown as extreme sensitivity to initial parameters, which By extending matrices Qph and Q1 hh in a way similar to
prompts modelers to argue that the solution is non-unique. what was done in Eqs. (2), (14) can also be viewed as
To overcome these problems, some researchers tend to set cokriging (Kitanidis and Vomvoris 1983; Hoeksema and
upper and lower bounds on estimated parameters. In it- Kitanidis 1984), which results from minimizing the
self, this does not solve the problem. The solution simply variance of estimated parameters. Possibly, the most im-
fluctuates between those (arbitrary) bounds, but its reli- portant thing is that Eq. (14) and its kriging variants do
ability is not improved. This prompted Neuman (1973) to not rely explicitly on any geometrical parameterization
add a plausibility term, Fp, to the objective function: scheme. Once Qhh has been found, one can theoretically
F ¼ Fh þ lFp ð9Þ estimate p at any point. In fact, for Kitanidis and Vom-
voris (1983), the only parameters to be estimated are the
Fp ¼ ðp p Þt C1
p ðp p Þ ð10Þ ones characterizing the statistical properties of the log-T
field. In general, the covariance of measurement errors is
This equation was originally based on matching and also needed.
stability arguments. The term Fp can be viewed as a
regularization term in the sense of Tihonov (1963). In Non linear methods
hydrology, Emselem and de Marsily (1971) used it to One question that arises from Eq. (11) is why should the
dampen oscillations. Also, Eq. (9) can be derived by linearization be performed at the prior estimates of the
statistical means in which case Ch is viewed as the co- parameters (linearization is made around the constant
variance matrix of measurement errors. Gavalas et al. mean in the original papers). Actually, Carrera and Glo-
(1976) derived it by maximizing the posterior pdf of the rioso (1991) prove that it is much better to linearize
model parameters (maximum a posteriori, MAP). The around the estimated parameters themselves, which
obtained objective function is equal to Eq. (9) with l should be close to the true ones and may be far from the
equal to 1. Carrera and Neuman (1986a) derived Eq. (9) prior estimates. In practice, this is achieved by iteratively
by maximizing the likelihood of the parameters given the linearizing at the parameters obtained in the previous it-
data (maximum likelihood estimation, MLE). The ad- eration. Carrera and Glorioso (1991) also showed that the
vantages of formulating Eq. (9) in a statistical framework result is identical to the one obtained by minimizing Eq.
lie in the fact that it yields ways to estimate not only the (9), and that Eq. (14) yields the same results as the first
parameters controlling aquifer properties but also those iteration of a Gauss-Newton method to minimize Eq. (9).
controlling their uncertainty (variances, variogram and This closes the loop. As it turns out, all the methods
the like). As it turns out, the latter are no less important discussed previously and summarized in Table 2, are
than the former (Zimmerman et al. 1998). identical from the estimation viewpoint.
The minimization of the objective function is an ar- Differences among methods are restricted to the
duous task because the relation between state variables characterization of statistical parameters and to the eval-
and parameters is usually non-linear. This is why for- uation of uncertainty. Without discussing how different
mulations of the inverse problem can be classified as methods deal with these issues, the next section is devoted
linear and non-linear. to arguing that indeed they are important.
Linear methods
Regardless of the formulation, the problem can be lin- How codes work
earized as:
In essence, most codes work by following an iterative
hðpÞ ¼ h0 þ Jhp ðp p Þ ð11Þ
process such as:
where Jhp is the Jacobian (sensitivity matrix), containing
the derivatives of h with respect to p. Its computation is
1. Initialization: Read input data, set iteration counter Computing the updating step, d
i=0, initialise parameters, p0 Computation of d falls within the realm of optimization.
2. Solve the simulation problem, h(pi), compute the ob- Literature on numerical optimization methods is exten-
jective function Fi, and possibly its gradient (assuming sive, but only some of these have been applied to the
that it is continuously differentiable), gi and the Jaco- solution of the groundwater inverse problem. They are
bian matrix, Jhp. summarized in Table 3. Cooley (1977) proposed using
3. Compute an updating vector, d, possibly using infor- Marquardt’s method (Marquardt 1963), which is based on
mation on previous iterations, as well as gi and Jhp. minimizing a quadratic approximation of the objective
4. Update parameters, piþ1 ¼ pi þ d function while limiting the size of the step to be taken at
5. If convergence has been reached, then stop. Otherwise, each iteration. This method is quite powerful, in the sense
set i=i+1 and return to 2. that it often converges in a small number of iterations, but
expensive because it requires computing the Jacobian
These steps are rather straightforward, except for the matrix. Carrera and Neuman (1986b) propose using a
definition of the updating direction, d, and the computa- combination of quasi-Newton and conjugate gradient
tion of g and Jhp. The most frequent methods for these are methods. These are not as robust as Marquardt’s method
outlined below. but their cost per iteration is much smaller, as they only
need to compute the gradient of the objective function.
The choice of one or the other is problem dependent.
an increase in the use of codes based on it, such as Uncertainty about the model structure
UCODE and PEST (Poeter and Hill 1998; Doherty et al. Conceptual models have many uncertain features because
2002). data are never exhaustive and contain inconsistencies.
Modelers are forced to make simplifying assumptions.
Errors are introduced in the parameterization, in the dis-
How good is the model? cretization, in the selection of boundary conditions, etc.
Addressing this uncertainty often requires posing several
Following the modeling procedure outlined in Fig. 1, once conceptual models. Different models can be compared in
the parameters have been estimated, it is necessary to terms such as model fit Eq. (8), residual distribution,
assess how good are both the model concept and esti- parameter correlation, and confidence intervals for pa-
mated parameters. Model quality is affected by three rameters and predictions. Ideally, a good model should
factors (Beck 1987): (1) uncertainty about the model lead to a good match with observations, uncorrelated
structure, (2) uncertainty about the values of the param- residuals, and reasonable parameter values. Still, several
eters appearing in the model structure and (3) uncertainty models may fit all these criteria and one may need to
associated with predictions of the future behavior of the select one among them. Several model selection criteria
system. The term uncertainty is used here to mean not have been defined from the field of time series analysis
only random fluctuations in errors, but also biases (see, and applied to groundwater (Akaike 1974, 1977; Rissanen
e.g., Barth et al. 2001).
1978; Schwarz 1978; Hannan 1980; Kashyap 1982). comment on which prior information will most effec-
Carrera and Neuman (1986c) applied four of these criteria tively reduce uncertainty.
to a synthetic test case and concluded that the Kashyap – Reducing the number of parameters: This is what
criterion was the best. Similar results were obtained by motivates the parameterization schemes discussed in
Carrera et al. (1990b) and by Medina and Carrera (1996). the section entitled, “What is to be estimated?”.
Still, one might question the wisdom of selecting only – Increasing the number and types of data, which was
one model. This implies rejecting the others, which may discussed in the section, “The conceptual model:
not be logical if they are consistent with available Knowledge and data”.
knowledge and data. This line of argument leads to ac- – Optimizing the observation scheme: Observation net-
cepting a large set of models and using them all to works and experiments can be designed to minimize
characterize uncertainty in model predictions (Beven and model uncertainty and/or to increase the ability of data
Binley 1992; Beven and Freer 2001) to discriminate among alternative models (Knopman
and Voss 1989; Usunoff et al. 1992)
Difficulties associated with the calibration process
The most important difficulties associated with the opti- Despite this, one may have to acknowledge that it is not
mization are the problems of non-uniqueness, non-iden- possible to find a unique solution to the problem. This
tifiability and instability. Non identifiability occurs when motivates some researchers to use stochastic simulation of
more than one set of parameters leads to a given solution parameter fields conditional to data, rather than estima-
of the forward problem. Non-uniqueness is present when tion (e.g. Gmez-Hernndez et al. 1997). These tech-
more than one set of parameters leads to minima of the niques generate large numbers of e.g. transmissivity fields
objective function. Instability is present when small that satisfy the available head and transmissivity data. In
changes in the observations lead to large changes in the this way, one ends up with a number of models, rather
estimated parameters, but is usually identified by a de- than one. Uncertainty is associated with the ensemble set
pendence of the solution on the initial parameters. In of all simulations, rather than to statistical measures of
addition, optimization algorithms such as Marquardt and uncertainty. As another alternative, Yapo et al. (1998)
conjugate gradients may get stuck into a local minimum, study instability using a concept known as the pareto
a set of model parameters whose performance cannot be optimum, which denotes the set of parameter vectors for
improved by small changes in data, but which does not which improving one component of the objective function
represent the overall optimum. Carrera and Neuman causes a deterioration in another component.
(1986b) discuss extensively these concepts and show that
they are closely related. In fact, they can all be charac- Difficulties associated with predictions
terized by the region of uncertainty, even if it is linearized of the future behavior of the system
(Fig. 2). All the above difficulties cause the model parameters to
Instability and large uncertainty are different concepts, have some uncertainty, which is inherited by the model
but both are often associated to elongated confidence predictions. The possibility of multiple conceptual models
regions, which can be characterized by the eigenvalues causes further uncertainty in the model predictions.
and eigenvectors of the posterior covariance matrix. The Therefore, evaluating uncertainty in prediction requires
eigenvectors of this matrix form a set of orthogonal analyzing the effect of both parameter and model uncer-
vectors, each of which is associated to an eigenvalue. The tainty. The latter is usually analyzed by simulating with
vector associated to the largest eigenvalue represents the the models available and evaluating the range of predic-
linear combination of parameters that has the largest un- tions (e.g., Medina and Carrera 1996). While systemati-
certainty, whereas the vector with the smallest eigenvalue zation is needed, the fact that conceptual model building
defines the direction with least uncertainty. If eigenvalues is not systematic makes this objective hard to meet.
are dramatically different, then one should expect insta- Hence, what follows concentrates on evaluating the effect
bilities to occur. One of the effects is parameter correla- of parameter uncertainty.
tion. The parameters in Fig. 2 suffer from correlation: a There exist various methods for quantifying the pre-
shift in the estimation of one parameter (say, P1) causes a diction uncertainty: Linear approximation, non-linear
shift in the optimal value of the other (P2). In other words, approximation and Monte Carlo methods. They are out-
the estimated parameters are not independent. This de- lined in Fig. 3. The linear method is relatively simple. It is
pendence causes the confidence intervals of the parame- based on the a posteriori covariance matrix:
ters to be larger than they would have been if they were
independent. Sp ¼ Jthp C1 1
h Jhp þ Cp ð19Þ
To summarize the above, instabilities and large un- where Jhp is the Jacobian matrix, Ch the measurements
certainties, though frequent, can be easily identified and covariance matrix and Cp the parameters covariance
characterized. Tactics to combat these problems include: matrix.
If f is a prediction to be made with the model (f is a
– Regularization: This is what motivated adding Fp to function of parameters, p), a lower bound of its variance
the objective function in (9). Weiss and Smith (1998) is given by:
of science” (http://go5.isiknowledge.com/portal.cgi),
@f t @f which keeps track of all papers published in major jour-
Var ðf Þ ¼ Sp þ s2p0 ð20Þ
@p @p nals, was performed. Results suggest that the number of
papers about inverse modeling remains more or less
where s2p0 represents model errors independent of pa- steady (around 5% of those about groundwater modeling),
rameter uncertainties. This equation represents a linear while the number of papers using inverse modeling has
approximation, so that actual uncertainty may be larger slowly but steadily increased in the last 13 years. These
than that represented. This is why more sophisticated papers cover a broad range of topics. Many have been
approaches may be needed. Still, this equation allows one cited in previous sections and some will be cited in the
to spell out quantitatively how the different factors affect remaining sections. It is difficult, however, to identify
prediction uncertainty. In essence, prediction uncertainty trends. The only ones that emerge is that geostatistical
grows with: inversion tends to be used in relatively small scale
problems, such as the interpretation of hydraulic tests,
1) Sensitivity of predictions to model parameters and while large scale models tend to be based on zonation.
initial conditions. Obviously, a parameter is a source of Since most groundwater applications can be classified in
concern only if predictions are sensitive to it, as one of these two categories, they deserve further attention
measured by @f/@p. This is why sensitivity analyses and are discussed below.
are sometimes performed instead of a formal error
analysis. Geostatistical inversion
2) Uncertainty in model parameters (and/or initial state). The use of geostatistics is motivated by the need to ad-
This is measured by the covariance matrix Sp . dress the variability of hydraulic properties (specifically
transmissivity) when modeling aquifers. In practice,
Non linear methods are more difficult to apply. Vecchia however, this need can be understood in two different
and Cooley (1987) present a way to compute nonlinear ways: (1) to constrain model parameters and (2) because
confidence intervals. They conclude that (i) correspond- it is deemed necessary for model predictions. While these
ing linear and nonlinear confidence intervals are often two views are not exclusive, they rarely go together.
offset or shifted towards each other, and the nonlinear Methodologically, geostatistical inversion follows the
ones are often larger, (ii) the variability in sizes of non- steps originally proposed by Clifton and Neuman (1982).
linear confidence intervals is usually larger than the cor- That is, one starts by proposing a stochastic model (i.e.,
responding variability in linear confidence interval sizes, whether the log-T is stationary, what is its variogram and
(iii) the difference between the sizes of linear and non- mean, etc). Second, available data is used to produce a
linear confidence intervals increases as the sizes of the prior, best estimate of log-T and its covariance, using Eqs.
intervals increase, and (iv) prior information can alter the (2) and (3). Finally, these are used to obtain an estimate of
size of the confidence intervals. Christensen and Cooley model parameters either by minimizing Eq. (9) or using
(1999) present a measure to quantify model non-linearity. Eq. (14). Kitanidis and Vomvoris (1983) and Carrera and
The prediction analyzer of PEST currently includes Neuman (1986a) modified this concept by emphasizing
nonlinear confidence intervals. the need for optimal estimation of statistical parameters
The Monte Carlo method is the most computationally (variances of errors, correlation distance and the like).
intensive method. It is based on many forward problem The importance of these parameters has been recognized
evaluations with different sets of parameters. Its main by many. In fact, this is one of the conclusions of Zim-
advantages are that it is easy to understand, it yields a merman et al. (1998), after comparing different geosta-
probability density function and does not require difficult tistical inversion techniques.
assumptions. Its main problem is that it is hard to ascer- Successful applications to real field data are restricted
tain the required number of simulations. An example of to relatively small-scale problems. These include hy-
this is the GLUE methodology (Generalized Likelihood draulic test interpretation (Yeh and Liu 2000; Meier et al.
Uncertainty Estimation; Beven and Freer 2001). 2001; Vesselinov et al. 2001), well capture zone delin-
eation (Vassolo et al. 1998; Kunstmann et al. 2002; Harrar
et al. 2003), and others (Barlebo et al. 2004). The fact that
What is actually done? applications to large-scale aquifers are scarce reflects the
Application trends difficulty in modeling them, but also points out the two
The literature in scientific journals tends to concentrate on limitations of geostatistics as it is most frequently used: it
the development of new methods and new interpretations. fails to include geological information and it fails to re-
Therefore it is not appropriate for identifying application produce actual variability.
trends. Applications are most often found in internal re- Geological information is normally expressed in terms
ports or special sessions of congresses. None of these are that are difficult to account for during inversion (depo-
easy to track exhaustively. Therefore, this section is bi- sitional patterns, orientation of conductive features, and
ased by what can effectively be found and by the authors’ the like). This type of information can become precise at
personal views. Despite the above, a search on the “web the large scale but rarely at the test scale. Here, the only
thing that can be said is that permeability is variable,
which is properly acknowledged by geostatistics. Still, 1999, was that official pumping rates were badly under-
when discrete features are identified, inversion gains estimated) and the nature of river-aquifer interaction.
significantly by explicitly including them as deterministic The somewhat chaotic nature of the process makes it
features. This was acknowledged by Meier et al. (2001), hard to describe in detail. Attempts have been made by
who used information about the stress state of a shear Cooley et al. (1986) and Hill (1998). Some trends can be
zone to define the anisotropy direction of fracture trans- identified:
missivity.
Ironically, the main problem with geostatistical esti- – Point values of hydraulic conductivity and transmis-
mation is the fact that it fails to reproduce actual vari- sivity are prone to error. Moreover they may be of little
ability. As discussed in the section, “How to estimate the use when modeling at scales much larger than the
model parameters, p”, Eqs. (9) and (14) yield the condi- pumping test in which they are based. These mea-
tional expectation of log-T given the available data. As surements need to be put into the related geological
such, the estimation says little about actual deviations context.
because the expected value filters them out. When data – Dominant features (i.e. conductive fractured zones,
are abundant, which is frequent in detailed hydraulic tests, paleochannels, or the like) must be included in the
variability patterns can be delineated with some certainty model even if they are not known accurately.
(e.g. Meier et al. 2001). Otherwise, only the estimation – Much information about aquifer behavior is contained
covariance provides some information about variability in discrete events (floods, big rainfalls). Taking full
and about the continuity of high conductivity zones (paths advantage of these requires transient simulations.
for solutes migration) or low conductivity barriers. An- – Model calibration is rarely unique (i.e., different model
other source of hope for large-scale geostatistical inver- structures may fit hard data satisfactorily). This un-
sion is the use of geophysical data (Hubbard and Rubin certainty ought to be acknowledged when performing
2000). model predictions. Reducing it often requires the use
Gmez-Hernndez and coworkers (Gmez-Hernndez of complementary data as discussed in the section,
et al. 1997; Capilla et al. 1998) address the above issue by “The conceptual model: Knowledge and data”.
rejecting optimal estimation altogether and, instead, per-
forming conditional simulations. Ideally, the average of This kind of approach displays several drawbacks. On one
all these simulations should be equal to the conditional the hand, it does not account properly for uncertainty. The
estimation, but each of them reproduces the assumed resulting covariance matrix of model parameters is con-
variability. As such, when used for predicting processes ditioned not only on hard data, but also on the many
that are sensitive to variability (e.g., contaminant trans- subjective decisions the modeler has made. While these
port), simulations are much more appropriate. can be taken into account by making predictions with
several conceptual models, it is rarely done. On the other
Geologically based inversion hand, one is never sure about the validity of the model
The main difference between aquifer scale models and beyond calibration conditions. In summary, the procedure
test scale models is the degree of reliance on geology needs to be systematized. This is best done in a geosta-
when defining variability and the multiplicity of param- tistical framework, hence the need to seek geostatistical
eter types. Regarding variability, geological data are descriptions that take advantage of qualitative geology
rarely precise but cannot be ignored. Different formations, data.
or different units within a formation, may have different
properties. Thus, when these units can be outlined, the
worth of this information cannot be ignored. Unfortu- What comes next?
nately, boundaries between units are rarely known accu-
rately. Hence, a lot of work may be needed to test dif- The discussion in the previous section makes it clear that
ferent geometries whenever the effect of geometry is it is believed that the time is ripe for standard use of
found to be important. The process involves defining inverse modeling in groundwater studies aimed at aquifer
geometry and testing it against available data, which is characterization and management, a view which we share
repeated until a satisfactory fit is found. The procedure is with Poeter and Hill (1997).
tedious, it involves interactions between modelers and In the 1970s, the U.S. Geological Survey promoted the
geologists, and it is not systematic. In fact, it does not get use of numerical modeling for aquifer studies. This led to
properly documented, so that if the model is revised years a significant rise in the quality of understanding of
later, one is not sure about the reason behind the selected groundwater flow by many hydrologists. Numerical
model structure. Things are made worse by the fact that modeling forced them to be quantitatively consistent
not only transmissivity but also other types of parameters when integrating different data types. Since this is never
need to be specified. In the experience of the authors of straightforward, hydrologists were forced to test different
this paper, ambiguities frequently exist about recharge parameter values, to perform sensitivity analyses, and to
(both average amount and time variability), boundary guess at what could be the cause behind observed data. In
fluxes, pumping rates (the main finding of Castro et al. the end, they might not be fully successful, but they
gained understanding, which is what counts for proper soft information), existing codes lack sufficient ro-
decision making. bustness and flexibility to be of general use.
The situation is now changing on several accounts. For – Incorporation of age, environmental isotopic data,
one, hydrologists are increasingly expected to make hard temperature and other sources of information. As
decisions for which qualitative understanding is not suf- discussed, using different types of data improves dra-
ficient. Instead, accurate quantitative models are needed. matically the stability of inversion and the robustness
Second, the volume of data is also increasing. Long data of the model. In fact, a number of authors have shown
sets of heads, pumping history, and hydrogeochemical that incorporating these data does improve the reli-
measurements are becoming available. While they con- ability of the model. The problem, again, lies in the
tain valuable qualitative information, it is clear that much availability of easy-to-run, flexible codes.
more can be gained by using them quantitatively, i.e., by – Representation of uncertainties. It has also been dis-
building models that can match those data. Third, mod- cussed that dealing with uncertainty is an integral part
eling exercises such as INTRAVAL (Larsson 1992) have of modeling. This is true both at the characterization
made it clear that the most important issue when model- stage, where data contain errors and model structure is
ing an aquifer is the conceptual model. Since manual never accurately known, and at the prediction stage.
calibration is very tedious, modellers have not been able Well informed decisionmaking can not be based on
to concentrate sufficiently on conceptual issues. Auto- single model predictions. Acknowledging uncertainties
matic calibration should change that. in both model concept and parameters is required. As
In short, it is contended that conceptually sound and discussed, linear estimates of uncertainty are very poor
quantitatively consistent modeling requires automatic in- (they may underestimate actual errors in orders of
version. This allows the modeler to concentrate on the magnitude; Carrera and Glorioso 1991). Alternative,
qualitatively important issues, such as the definition of the non-linear estimates of error are difficult to use for
conceptual model, representation of spatial variability, hydrologists who are not familiar with statistics, even
discussion of alternative management strategies, etc. though tools are becoming increasingly user-friendly.
Also, this allows improved model building and use. Therefore, Monte Carlo simulation remains the most
Therefore, it should be one of the challenges in the future appealing. An advantage is the relative ease to ac-
of hydrogeology. As it is, however, inverse modeling has commodate conceptual model uncertainty. Still, at
not reached a desired level of maturity. A number of present, Monte Carlo methods are extremely expensive
difficulties must be overcome. Some are listed below: and again, not easy to use.
– Coupling to GIS. One of the drawbacks of inverse
– Difficulty in running inversion codes. Since one has to modeling is the need to incorporate all causes of
introduce data on parameterization, observations, their variability; geology, soil properties and use, pumping,
reliability, etc., general purpose inversion codes are etc. If any of them is missed (e.g., if the effects of a
cumbersome and hard to run. The emergence of codes pumping well are not incorporated), the algorithm will
such as PEST or UCODE has changed that trend sig- react by modifying the parameters to circumvent the
nificantly, but places an additional burden on the effect of the error so as to best fit the measurements
modeler by requiring him or her to know programming (e.g., reduce recharge, increase transmissivity, etc.).
and the inside of codes. This could be alleviated by The resulting model is thus erroneous not only because
developing ‘utility programs’ to enable the interfacing of the missing factor, but also because of those mod-
of the inversion engine with the forward model. ifications. Conventional modeling is less sensitive to
– Incorporation of geological information. As discussed this kind of error (one may be aware that model cal-
earlier, zonation is the most convenient and widely culations may be erroneous in areas affected by un-
used approach for realistic incorporation of geological known factors, but that does not affect the rest of the
data available in the form of maps. This is too rigid model). Therefore, inverse modeling requires careful
and incomplete. Geological information is usually accounting. As the history of aquifers becomes in-
“soft”, in the sense that boundaries between formations creasingly complex (growing number of pumping
underground are rarely known accurately. Moreover, wells, evolving soil uses, improved knowledge of
important features (paleochannels, water conducting aquifer geometry), so does the difficulty to incorporate
faults, etc.) may have gone unmapped. Furthermore, all factors. The tendency to incorporate all territorial
even within a formation, there may be a lot of quali- data in GIS sheds some hope on the possibility of in-
tative information (depositional patterns, continuity of corporating all these data in a model. In fact, there
water conducting features, gradation of materials and have been a number of efforts to link GIS to traditional
the like) that is difficult to incorporate in a zonation models (Gogu et al. 2001; Chen et al. 2002). It is clear,
framework. For solute transport problems, these sour- however, that this need is more pressing for inverse
ces of small scale variability may be just as important models. Efforts along the line of Graphical User In-
as large scale trends. Presumably, they can be best terfaces are starting.
handled in a geostatistical framework. While some
approaches are available for incorporating them (i.e., All the above issues might suggest that inverse modeling
treat each zone geostatistically, use geological maps as is not yet mature. It should be stressed, however, that all
the above-mentioned difficulties are just that, difficulties. Carrera J, Neuman SP (1986c) Estimation of aquifer parameters
They can be, and indeed are, overcome in practice under transient and steady-state conditions, 3. Application to
synthetic and field data. Water Resour Res 22(2):228–242
through laborious work. Moreover, these difficulties are Carrera J, Navarrina F, Vives L, Heredia J, Medina A (1990a)
not specific of inverse modeling; they are mostly shared Computational aspects of the inverse problem. In Proc. of VIII
by conventional modeling (although they are often ig- international conference on computational methods in water
nored). Finally, while it is true that further improvement resources. CMP, pp 513–523
Carrera J, Heredia J, Vomvoris S, Hufschmied P (1990b) Fracture
of programs is needed, it is not clear that their applica- Flow Modelling: Application of automatic calibration tech-
bility is more difficult than it was for the use of early niques to a small fractured Monzonitic Gneiss Block. In:
codes when the U.S. Geological Survey made modeling a Neuman N (ed) Proc hydrogeology of low permeability envi-
routine for aquifer studies. Routine application of inverse ronments, IAHPV, Hydrogeology, selected papers, vol 2, pp
modeling is the future. The sooner it starts, the better 115–167
Carrera J, Glorioso L (1991) On Geostatistical Formulations of the
prepared will hydrologists be to face the challenges of the Groundwater Flow Inverse Problem. Adv Water Resour
soon-to-come future. 14(5):273–283
Carrera J, Medina A, Galarza G (1993a) Groundwater inverse
Acknowledgments The final manuscript benefitted from comments problem. Discussion on geostatistical formulations and valida-
by Mary Hill, Matt Tonkin and Johan Valstar tion. Hydrogologie (4):313–324
Carrera J, Mousavi SF, Usunoff E, Sanchez-Vila X, Galarza G
(1993b) A discussion on validation of hydrogeological models.
Reliability Eng Syst Saf 42:201–216
References Carrera J, Medina A (1994) An improved form of adjoint-state
equations for transient problems. In: Peters, Wittum, Herrling,
Akaike H (1974) A new look at statistical model identification. Meissner, Brebbia, Grau, Pinder (eds) Proc X international
IEEE Trans Automat Contr AC-19:716–722 conference on methods in water resources, pp 199–206
Akaike H (1977) On entropy maximization principle. In: Krish- Castro A, Vazquez-Su
e E, Carrera J, Jaen M, Salvany JM (1999)
naiah PR (ed) Applications of statistics. North Holland, ms- Calibracin del modelo regional de flujo subterrneo en la zona
terdam, pp 27–41 de Aznalcllar, Espa
a: ajuste de las extracciones [Calibration
Anderman ER, Hill MC (1999) A new multi-stage ground-water of the groundwater flow regional model in the Aznalcollar site,
transport inverse method, Presentation, evaluation, and impli- Spain: extractions fit]. In Tineo A (ed) Hidrologa Subterrnea.
cations. Water Resour Res 35(4):1053–1063 II,13. Congreso Argentino de Hidrogeologa y IV Seminario
Barlebo HC, Hill MC, Rosbjerg D (2004) Identification of Hispano Argentino sobre temas actuales de la hidrogeologia
groundwater parameters at Columbus, Mississippi, using three- Chen CX, Pei SP, Jiao JJ (2003) Land subsidence caused by
dimensional inverse flow and transport model. Water Resour groundwater exploitation in Suzhou City, China. Hydrogeol J
Res 40(4):W0421110 11(2):275–287
Barth GR, Hill MC, Illangasekare TH, Rajaram H (2001) Predictive Chen Z, Huang GH, Chakma A, Li J (2002) Application of a GIS-
modeling of flow and transport in a two-dimensional interme- based modeling system for effective management of petroleum-
diate-scale, heterogeneous porous media. Water Resour Res contaminated sites. Env Eng Sci 9(5):291–303
37(10):2503–2512 Christensen S, Cooley RL (1999) Evaluation of confidence inter-
Beck MB (1987) Water quality modelling: a review of the analysis vals for a steady state leaky aquifer model. Adv Water Resour
of uncertainty. Water Resour Res 23(8):1393–1442 22(8):807–817
Bennet RR, Meyer RR (1952) Geology and groundwater resources Clifton PM, Neuman SP (1982) Effects of kriging and inverse
of the Baltimore area. Mines and Water Resour Bull 4, Mary- modeling on conditional simulation of the Avra valley aquifer
land Dept of Geology in southern Arizona. Wat Resour Res 18(4):1215–1234
Beven K (1993) Prophecy, Reality and uncertainty in distributed Cooley RL (1977) A method of estimating parameters and as-
hydrological modeling. Adv Water Resour 16(1):41–51 sessing reliability for models of steady state groundwater flow,
Beven KJ, Binley AM (1992) The future of distributed models: 1, Theory and numerical properties. Water Resour Res
model calibration and uncertainity prediction. Hydrol Process 13(2):318–324
6(3):279–298 Cooley RL (1985) A comparison of several methods of solving
Beven KJ, Freer J (2001) Equifinality, data assimilation, and un- nonlinear-regression groundwater-flow problems. Water Re-
certainty estimation in mechanistic modelling of complex en- sour Res 21(10):1525–1538
vironmental systems using the GLUE methodology. J Hydrol Cooley RL, Konikow LF, Naff RL (1986) Nonlinear regression
249:11–29 groundwater-flow modeling of a deep regional aquifer system.
Bredehoeft J (2004) Modeling: the conceptualization problem- Water Resour Res 22(13):1759–1778
surprise. Hydrogeol J (this issue) Dagan G (1985) Stochastic modeling of groundwater flow by un-
Capilla JE, Gmez-Hernndez JJ, Sahuquillo A (1998) Stochastic conditional and conditional probabilities: the inverse problem.
simulation of transmissivity fields conditional to both trans- Water Resour Res 21(1):65–72
missivity and piezometric head data—3. Application to the de Marsily GH, Lavedan G, Boucher M, Fasanino G (1984) In-
Culebra formation at the waste isolation pilot plan (WIPP), terpretation of interference tests in a well field using geosta-
New Mexico, USA. J Hydrol 207(3–4):254–269 tistical techniques to fit the permeability distribution in a res-
Carrera J (1987) State of the art of the inverse problem applied to ervoir model. In: Verly et al (ed) Proc Geostatistics for natural
the flow and solute transport problems. In: Groundwater flow resources characterization. Part 2. D. Reidel Pub. Co. : pp 831–
and quality modeling, NATO ASI Ser: 549–585 849
Carrera J, Neuman SP (1986a) Estimation of aquifer parameters de Marsily G, Delhomme JP, Delay F, Buoro A (1999) 40 years of
under transient and steady-state conditions, 1. Maximum like- inverse problems in hydrogeology. Comptes Rendus de
lihood method incorporating prior information. Water Resour l’Academie des Sciences Series IIA. Earth and Planet Sci
Res 22(2):199–210 329(2):73–87. Elsevier Science
Carrera J, Neuman SP (1986b) Estimation of aquifer parameters Doherty J, Brebber L, Whyte P (2002) PEST-Modelling dependent
under transient and steady-state conditions, 2. Uniqueness, parameter estimation. Water Mark Computing. Corinda (Aus-
stability and solution algorithms. Water Resour Res 22(2):211– tralia)
227