462

Integrated Environmental Assessment and Management Volume 8, Number 3pp. 462472

2011 SETAC

A Bayesian Network Model for Integrative River

Rehabilitation Planning and Management
Mark E Borsuk,*y,z Steffen Schweizer,y, and Peter Reicherty,k
y Eawag, Swiss Federal Institute of Aquatic Science and Technology, Du
bendorf, Switzerland
z Thayer School of Engineering, Dartmouth College, Hanover, New Hampshire, USA
Kraftwerke Oberhasli AG, Meiringen, Switzerland
k ETH Zu
rich, Swiss Federal Institute of Technology, Zu
rich, Switzerland

(Submitted 15 July 2010; Returned for Revision 4 November 2010; Accepted 16 May 2011)

Special Series

EDITORS NOTE
This paper represents 1 of 7 review and case study articles generated as a result of a workshop entitled Scenario and decision
analysis in environmental management using Bayesian Belief Networks (12 October 2009, Oslo, Norway) hosted by the
Norwegian Institute for Nature Research (NINA) and the Strategic Institute Project Nature 2020R and funded by the
Research Council of Norway. The main aim of the workshop was to compare Bayesian network applications to different
environmental and resource management problems from around the world, identifying common modeling strategies and
questions for further research.

ABSTRACT
As rehabilitation of previously channelized rivers becomes more common worldwide, exible integrative modeling tools are
needed to help predict the morphological, hydraulic, economic, and ecological consequences of the rehabilitation activities.
Such predictions can provide the basis for planning and long-term management efforts that attempt to balance the diverse
interests of river system stakeholders. We have previously reported on a variety of modeling methods and decision support
concepts that can assist with various aspects of the river rehabilitation process. Here, we bring all of these tools together into a
probability network model that links management actions, through morphological and hydraulic changes, to the ultimate
ecological and economic consequences. Although our model uses a causal graph representation common to Bayesian
networks, we do not limit ourselves to discrete-valued nodes or conditional Gaussian distributions as required by most Bayesian
network implementations. This precludes us from carrying out easy probabilistic inference but gives us the advantages of
functional and distributional exibility and enhanced predictive accuracy, which we believe to be more important in most
environmental management applications. We exemplify model application to a large, recently completed rehabilitation project
in Switzerland. Integr Environ Assess Manag 2012;8:462472. 2011 SETAC
Keywords: Bayesian network

Decision support system

Integrative modeling

INTRODUCTION
Many rivers worldwide have been channelized to extend
agricultural, industrial, or residential land use, facilitate river
navigation, and reduce ood risk (Rosgen 1994; Buijse et al.
2002). These engineering projects have resulted in straight,
often erosive, rivers lined by articial ood levees. Thus,
riparian zones are homogenized and natural dynamics are lost
(Ward et al. 2001), decreasing the habitat quality for
organisms living in or near the river. This is often reected
in a reduction in the abundance and diversity of resident
terrestrial and aquatic organisms (Ward 1998). Human
recreational opportunities, such as shing, bird watching,
and swimming or wading are also usually diminished, leading
to an underappreciation of the ecosystem services that the
river has the potential to provide (Millennium Ecosystem
Assessment 2005).
In recent years, the rehabilitation of channelized systems
has become increasingly common, with some countries
spending billions of dollars to improve the ecological
* To whom correspondence may be addressed: mark.borsuk@dartmouth.edu
Published online 23 May 2011 in Wiley Online Library
(wileyonlinelibrary.com).
DOI: 10.1002/ieam.233

Uncertainty analysis

River rehabilitation

condition of rivers while maintaining ood protection for

adjacent land uses (Buijse et al. 2002; Shields et al. 2003;
Peter et al. 2005). This may involve the creation of localized
river widenings in which the levees are moved back to
allow natural channel movement within a limited area (Rohde
et al. 2005). Within the widened reach, the river might shift
and adjust, possibly reestablishing the range of riparian
habitats that could be found before channelization (Ernst
et al. 2010). This can certainly enhance local recreational
opportunities, and, if enough reaches are addressed, overall
river ecosystem health may improve (Naiman et al. 1993).
Despite the increasing number and expense of restoration
projects, there are few attempts to provide integrative
assessments of the benets that various rehabilitation options
are likely to incur. Most models designed to support the
restoration process do so in a limited manner, focusing
exclusively on hydraulic changes (Liu et al. 2004), gravel
transport dynamics (Singer and Dunne 2006), or single
targeted species (Tyler and Rutherford 2007; Noble et al.
2009). We have previously developed and reported on a
number of such models ourselves (Borsuk et al. 2006;
Schweizer et al. 2007a, Schweizer et al. 2007b) with intended
application primarily to large, midland rivers in Switzerland.
Here, we bring all of our previous efforts together into an
integrative model that links potential management actions,

A Bayesian Network Model for River RehabilitationIntegr Environ Assess Manag 8, 2012

through morphological and hydraulic changes, to ecological

and economic consequences. To do this, we use the structure
of a graphical probability network (often referred to as a
Bayesian network) because of the methods consistency with
the decision analytic paradigm (Reichert et al. 2007).
Specically, the ability to represent and propagate uncertainty
through multiple models that are based on varying types and
scales of knowledge is essential for evaluating candidate
decision alternatives according to a probabilistic assessment of
their anticipated outcomes.
As the specic model components have largely been
published elsewhere, the goal of the present article is to
describe how these components have been linked together
using the graphical probability network formalism. Specically, we use a directed acyclic graph to represent causal
relationships that are characterized by conditional probability
distributions. However, we do not require that the model be
used to carry out probabilistic inference (using Bayes
theorem), as this task leads to many practical limitations
and is not necessary for our purpose. In fact, true Bayesian
inference in fully integrated networks is required in only a
small minority of environmental management applications,
and the common software requirements of discrete-valued
nodes or conditional Gaussian distributions can be restrictive
and introduce unnecessary ambiguity. Bayesian inference is
still important to our effort but only in characterizing the
submodel relationships, and it is performed as an ofine step
in separate software. We exemplify application of our model
to a large, recently completed rehabilitation project in
Switzerland. This gives us the opportunity to test some
model predictions and the models potential use for providing
decision support.

DESCRIPTION OF METHOD
A (Bayesian) probability network is a directed acyclic graph
that leads to a compact representation of the joint probability
distribution of a set of variables in a system of interest (Pearl
2000). Graphically, system variables are represented by
nodes, and dependences between nodes are represented by
arrows. These arrows indicate patterns of probabilistic
dependence and not the ow of mass or process control.
Importantly, the absence of a directly connecting arrow
between any 2 nodes in a Bayesian network implies that the 2
variables are conditionally independent, given the values of
any intermediate nodes. This means that the probability
distribution of any variable Xi in any state of the network can
be determined by knowing only the values of its immediate
predecessors (called its parents, PAi), without regard to the
values of any other variables. In this way, the joint probability
distribution for the entire network can be written as the
product of a limited number of conditional distributions using
the chain rule of probability calculus
Px1 ; . . . ; xn

n
Y

Pxi jpai :

i1

Nodes without any parents are called roots and are

specied by marginal (i.e., unconditional) distributions.
The interpretation of a Bayesian network in terms of
causality is not necessary for extracting the conditional
dependence relations. However, it is usually the causal
interpretation that allows the structure of the network to be
determined. That is to say that appropriate experts can

463

construct a graphical network based on straightforward,

qualitative notions of cause and effect (that are the basic
building blocks of scientic knowledge) without necessarily
being uent in probabilistic reasoning. The network then
denes the appropriate factorization of all relevant variables
in the system into conditional distributions that can be used
to generate the necessary consequence probabilities (Eqn. 1).
In other words, the full diagram can be modularized to allow
the characterization of individual subnetworks to proceed
independently without regard to the broader context. This
means that each subnetwork can be specied using an
approach suitable for the type and scale of information
available (Borsuk et al. 2004). This is a property of which we
take particular advantage in the model described here.
Most published examples of Bayesian network models have
used either inherently discrete variables or continuous
variables that have been discretized into a nite number of
categories. When all variables in a Bayesian network are
discrete, then the relationships are specied by conditional
probability tables for each node that provide the probability
of it being in a particular state (or category), given any
combination of states of its parents. This simplies the
probability calculus involved in probabilistic (Bayesian)
inference of the states of ancestor nodes, given observed
states of descendants. This possibility for inference is the basis
for the term Bayesian network.
Inferring the states of some nodes, given the observed states
of some descendants (a process referred to as diagnosis), may
be useful in some river management contexts. For example, if
a network model is constructed to represent the relationship
between past disturbances (e.g., oods, drought, pollutant
spills) and particular ecosystem responses, then, when such a
response is observed in the future, the process of inference
can help determine which disturbance was likely to have been
the cause. In most cases of rehabilitation assessment however,
the primary task is to forecast the likely results of a potential
management action, rather than to diagnose the cause of a
symptom.
When the possibility for easy inference is not a requirement
of the model, then an alternative to constructing networks
entirely of discrete variables related by conditional probability
tables is to use continuous variables connected by functional
equations. Probabilities are introduced through the assumption that certain variables or parameters in the equations are
uncertain or unobserved. In many ways, this is more
consistent with the semideterministic way that causal models
are conceived and used in biology, physics, and engineering.
In its most general form, a probabilistic functional equation
for a variable Xi consists of an equation of the form
xi fi pai ; ui ;

where PAi are the parents of Xi, and Ui are the disturbances
caused by omitted variables or random effects (Pearl 2000).
This conceptualization can be considered a nonlinear version
of the more familiar linear structural equation models (Shipley
2002).
With all the conditional distributions of a Bayesian network
specied, generation of the distribution of nal outcome
nodes is straightforward and can generally proceed most
effectively using Monte Carlotype simulation (Borsuk et al.
2004). Each marginal node is set to a value specifying a
decision alternative or a distribution describing uncertain

464

background conditions. A large random sample is then

generated for the probabilistic marginal nodes, which is used
as input to the functions dening that nodes descendants,
along with samples from any other uncertain variables that
are required by the functions. This generates a sample of the
rst generation of descendants, which is propagated further
along the causal direction in an analogous manner until the
nal outcome nodes are reached. The marginal distributions
of each of the outcome nodes then represent the (probabilistic) predictions of the consequences of the specied
decision alternative.

MODEL DEVELOPMENT
Consistent with the decision analytic approach, our model
building process started with the construction of a comprehensive objectives hierarchy concerning the goals of river
rehabilitation (Reichert et al. 2007). This process was
undertaken in collaboration with a collection of organized
stakeholder groups for a river rehabilitation project in
Switzerland, as described by Hostmann et al. (2005a,
2005b). Measurable attributes describing the achievement
of the lower-level objectives of the hierarchy then provide the
set of variables that should ideally be predicted by the
probabilistic outcome model. The key variables resulting
from this process, in our case, in addition to the obvious ones
of ood protection level and rehabilitation cost, included:
river morphology, hydraulic habitat characteristics, erosion
potential, siltation of riverbed, benthic population diversity
and biomass, riparian community abundance, sh population
abundance and age structure, local employment potential,
and adjacent land use impacts (see Hostmann et al. 2005a for
details). These variables are related to each other and to input

Integr Environ Assess Manag 8, 2012ME Borsuk et al.

variables describing rehabilitation alternatives as shown in the

schematic network model of Figure 1. As most of the
submodels delineated by the dotted boxes in Figure 1 have
been described elsewhere, we will describe them only briey
in the following subsections.

Morphology and hydraulics submodel

Hydraulics and river morphology inuence all other
variables of interest (Figure 1). Therefore, model construction
began with this submodel (Schweizer et al. 2007b). Important attributes include river morphology, erosion potential,
hydraulic diversity, and riverbed siltation. Modules to predict
these attributes are briey described in the following
subsections. Schweizer et al. (2007b) provide details.
Channel morphology and erosion potential. A rivers morphology depends on the balance between stream power, external
width constraints, and gravel supply (van den Berg 1995). We
developed a stepwise procedure to consider these factors
(Figure 2). The procedure starts with the application of the
logistic regression model of Bledsoe and Watson (2001)
predicting the probability of a river being single- or multithreaded based on valley slope, mean annual ood discharge,
and median gravel diameter. Next, the impact of channel
width constraints is considered. This is done by predicting the
natural width for the indicated morphology using the results
of a regression of bankfull width on mean annual ood
discharge and median gravel diameter. This predicted natural
width is then compared against any constraints (e.g., levees)
expected to remain after rehabilitation. If the constrained
width is larger than the natural width, then single-threaded

Figure 1. Schematic probability network relating input variables describing catchment characteristics and rehabilitation alternatives (upper left) to submodels
predicting key decision attributes.

A Bayesian Network Model for River RehabilitationIntegr Environ Assess Manag 8, 2012

465

Figure 2. Flow chart of channel morphology determination.

rivers are predicted to be sinuous with alternating side bars,

otherwise they are predicted to be straight.
Rivers predicted to be multi-threaded according to the
logistic regression may yet be single-threaded if width
constraints are too severe. The pattern diagram of da Silva
(1991) indicates whether a river will be braided, meandering,
alternating, or straight for given gravel size, channel geometry,
and mean depth at bankfull discharge. For single-thread
rivers, the mean depth is estimated iteratively using the
equation of Strickler (1923), and for braided rivers the
iterative method of Zarn (1997) is applied. Finally, the gravel
transport capacity within the reach is calculated to check
whether the upstream gravel supply is sufcient to allow the
predicted morphological structures to develop. These calculations are based on the method of Meyer-Peter and Muller
(1948) with corrections based on Zarn (1997). When net
annual erosion is predicted, we assume that a straight, incising
channel will result. When there is net deposition, we assume
that the morphologies predicted by the previous calculations
will develop.
Hydraulics. Spatial variation in ow depth and velocity over a
river reach is a key determinant of ecosystem structure and
function (Allan 1995). We developed an empirical model for
the joint distribution of depth and velocity using survey data
from 92 stream reaches (Schweizer et al. 2007a). We found
that the bivariate distribution of relative velocity and relative
depth can be described by a mixture of 2 end-member
distributions, 1 normal and the other lognormal, each with
xed parameters. The relative contribution of each distribution can then be predicted from a combination of the
dimensionless characteristics: the reach mean Froude number,
the reach mean relative roughness, and the ratio of actual
discharge to mean discharge. We apply this method to
alternating gravel bar and braided morphologies. The joint
distribution of relative velocity and relative depth in a straight
river, on the other hand, is described by beta-distributed
marginals with xed parameters that are correlated with a

rank correlation coefcient of 0.94, as t to survey data from a

set of reaches in Switzerland (see Schweizer et al. 2007a for
details). The relative frequency of hydraulic units, such as
pools, runs, and rifes are calculated from the predicted
bivariate distributions of depth and velocity (Jowett 1993).
Riverbed siltation. Fish and benthic species use the gravel bed
matrix as cover and to provide egg incubation habitat.
Therefore, siltation and clearance of the interstices are
important ecological processes. Conceptually, we model
gravel bed siltation as a process that occurs at low to medium
discharges at a rate that depends on hydraulic and bed
characteristics. As a result of this process, the percent of nes
in the river bed increases and the hydraulic conductivity is
reduced. When the deposition and resuspension of particles
are in equilibrium, siltation of a river bed stops. The temporal
progression of the build up of nes is modeled according to
the methods of Schalchli (1995).
Occasionally, a high discharge event will ush the river
bed, restoring the original gravel structure and size distribution. The discharge sufcient to clear the river bed is
calculated according to comparisons of the dimensionless
shear stress with bed stability according to Gunther (1971).
The frequency of this discharge, together with the siltation
rate, determines the temporal extent and severity of siltation.
Because of spatial differences in the bottom shear stress, we
apply the siltation equations of Schalchli (1995) separately to
average conditions in pools and runs. We assume that
signicant siltation will not occur in rifes due to the very
high lter velocities.
Benthic population submodel
Periphyton and invertebrates are important components of
river ecosystems due to their ability to produce organic
material, decompose detritus, and serve as a food source for
organisms at higher trophic levels. We developed a set of
simple data-based models to describe the biomass of

466

Integr Environ Assess Manag 8, 2012ME Borsuk et al.

periphyton, total invertebrates, and 3 invertebrate functional

feeding groups (collector-gatherers, predators, and scrapers)
as a function of time since the last bed-moving ood, mean
water depth, grain size, mean ow velocity, and season (time
within the year) (Schweizer 2006). Parameters were estimated from a statistical t to survey data from as many as 8
sites in 3 rivers. The statistical approach made it easier to
derive reliable relationships across different rivers and sites
than would have been possible with a strictly mechanistic
modeling approach.
Considering the diversity of data sets and the simplicity of
the formulations, the models lead to a remarkably good
agreement with time series of measurements (Schweizer
2006). Because of the larger data set available, this was
particularly true for the periphyton model.
Riparian community submodel
Our riparian community model focuses on riparian
arthropods, a signicant contributor to overall riverine
biodiversity and a functionally important component of river
ecosystems. The model predicts the abundance of 3 major
groups (spiders, ground beetles, and rove beetles) as a
function of river morphology and riparian siltation. Both of
these variables are outputs of the Morphology and
Hydraulic submodel described above.
Using the data of Paetzold et al. (2008), we carried out
multiple regression analyses to relate the variation in species
abundance to the abiotic predictor variables (Figure 3). We
found that for all species there were signicant differences
between natural and channelized morphologies. In addition,
siltation reduced the abundance of all species similarly in both
types of morphologies, except for spiders at channelized sites
that were already so low that siltation had no further effect.
Uncertainty was accounted for through the inclusion of a
probabilistic error term, derived from the distribution of
residuals from the regression model.
Fish submodel
The core of our sh submodel is a dynamic, age-structured
population simulation (Borsuk et al. 2006). Because our
model was initially developed for application to midland
rivers in Switzerland, our focal species is brown trout, a
desirable sport sh and indicator of river ecosystem health
(Lasne et al. 2007).
We represent 5 major stages of the brown trout life cycle in
the model: eggs, newly emergent fry, late summer fry,
immature juveniles, and adult spawners. The distinction
between emergent and late summer fry was made to
differentiate the period of greatest density dependence. The
number of individuals in each life stage is determined by the
number in the previous life stage, as well as relevant
population parameters, such as survival and reproductive
rates. These parameters are inuenced, in turn, by intermediate variables, such as body size and growth rate, or by
external controls relevant to rehabilitation measures, such as
including substrate and habitat quality, stocking practices,
angling, prey resources, and competing species (Figure 4).
Consistent with the Bayesian viewpoint that probabilities
represent subjective degrees of belief, conditional probabilities in a Bayesian network may be based on any available
information, including experimental or eld results, processbased models, or the elicited judgment of experts. Many of

Figure 3. Predicted and observed abundance of (a) spiders, (b) rove beetles,
and (c) ground beetles as a function of riparian siltation and river morphology.
Squares represent data from channelized rivers and circles represent data from
rivers with natural morphology.

the relationships in the sh submodel were based on expert

elicitations (see Borsuk et al. 2006 for details).
Socioeconomic submodel
The socioeconomic submodel estimates the direct and
indirect effects of river rehabilitation on the local economy
(Sporri et al. 2007). The basic model is expressed in terms of
an inputoutput matrix (Miller and Blair 1985), tracking
goods and service ows between various sectors of the
economy. This matrix is then used to calculate the direct
and indirect changes in economic output and jobs per sector
resulting from the direct changes in demand associated with
the rehabilitation project (e.g., in the construction, real estate,
and other involved industries). We account for uncertainty in
the data and in some of the model assumptions using a
probabilistic formulation and propagating these probabilities
through the model equations (see Sporri et al. 2007 for
details).
Model users can input into the model a local inputoutput
matrix, if available, or modify a national matrix using, for
example, local employment statistics (the location quotient
method; Miller 1998). The inputoutput method can also be

A Bayesian Network Model for River RehabilitationIntegr Environ Assess Manag 8, 2012

467

Figure 4. Graphical probability network of brown trout population model. Dark shaded nodes indicate life cycle stages. Light shaded nodes indicate variables
likely to be inuenced by rehabilitation measures.

used to estimate the longer-term effects on the local economy

of changes in tourism and recreation expected to result from
rehabilitation.
Model implementation
The submodels described in the above sections were
implemented in Analytica, a commercially available software
program for propagating uncertainty through models formulated as graphical networks (Lumina 1997). In principle, the
model could be implemented using other software; we chose
Analytica because it allows for a wide variety of probability
distributions and functional model forms.
The inputs required to run the integrative river rehabilitation model consist of both site-specic characteristics (e.g.,
discharge statistics, reach slope, gravel size, water temperature patterns) and rehabilitation design criteria (e.g., project
expenditures, dike height and spacing, land use changes,
revegetation plans, sh stocking). The former can generally be
determined from historical or site survey data for the river
reach of interest, and the latter should be viewed as decision
variables to be set to values corresponding to current
conditions, specic design alternatives, or scenarios used for
sensitivity analysis.
In Analytica, each input variable is specied by either a
xed value, a probability density function, or, in the case of a
discrete variable, a probability table. Input nodes can also be
specied as multivariate joint distributions if the uncertainty
of 2 or more variables cannot be assumed to be independent.
Each child node is then dened by a conditional probability

table or, more generally, by a functional expression of the

form of Equation 1, derived from submodels such as those
described in the sections above. In the latter case, model
parameters and error terms are represented explicitly as
marginal nodes that serve as parents of a child node along
with that variables predictors. Covariance in model parameters is included by using appropriate multivariate distributions.
Once all model variables, parameters, and relationships are
specied, a large sample of realizations is drawn for each
marginal node using random Latin hypercube sampling
(McKay et al. 1979). These realizations are then propagated
through the functional expressions dening the conditional
distribution of each child node. These children then serve as
parents of the next set of nodes in the network, and their
uncertainty is propagated accordingly. The distributions of
nal model endpoints thus convey the combined uncertainty
and variability coming from their entire set of ancestors. This
process of Monte Carlotype simulation is implemented in
many other software packages, but to our knowledge,
Analytica is the only commercial software that provides a
graphical diagramming interface consistent with the Bayesian
network concept. Analytica also allows for the inclusion of
decision and objective nodes, making it easy to implement a
description of stakeholder preferences to yield decision
analytic results (e.g., Hostman et al. 2005a; Reichert et al.
2007).
The process of integrating our various submodels into a
cohesive network required some additional assumptions. For
example, whereas sufcient knowledge and data exist to build

468

the benthic population submodel at a owdynamic scale

(days to weeks), the sh population model focuses on life
cycle transitions at a longer time scale (months to years).
Thus, for estimating brown trout prey availability, the
seasonal average benthic biomass was used, neglecting shorter
term uctuations. In other network models, we have faced
the opposite situation, needing to express changes in highly
dynamic variables as functions of more slowly evolving
variables (Borsuk et al. 2004), a technique referred to as
variable speed splitting (Walters and Korman 1999). Either
of these methods may be necessary for cross-scale modeling
with Bayesian networks.
Another challenge comes with reconciling the spatial scales
of the various submodels. For example, although the physical
processes of siltation can be described mathematically at a ne
spatial and temporal scale, the brown trout spawning model
required the reach-wide seasonal average as input. Therefore,
we decided that a reasonable simplifying procedure would be
to: 1) assume spatially homogeneous conditions within each
hydraulic unit (i.e., runs, pools, rifes), 2) simulate the
siltation dynamics under each of these sets of conditions, then
3) calculate the seasonal average siltation level for each
hydraulic unit, and nally 4) calculate the weighted average
siltation level for the reach using the modeled proportions of
each type of hydraulic unit. Similar assumptions concerning
homogeneity within hydraulic units were also necessary for
other variables, including water depth and velocity, gravel
size, and habitat suitability.

MODEL APPLICATION
Case study description
To demonstrate application of the model, we use a case
study at the Thur River, Switzerland. The catchment of the
Thur River originates 2500 m above sea level and the river
ows 130 km before entering the Rhine River at 350 m above
sea level, giving a total catchment area of 1750 km2. Regular
ooding of the Thur prevented settlement of its oodplains
until 1890, when the rst river correction project occurred.
The meandering river was straightened and levees were built,
anking the river at widths of 30 to 50 m. However, the
corrections could not contain the largest oods and also gave
rise to new problems. In the straightened channel, water
owed more quickly and with a greater depth than before,
increasing its erosive power and undercutting the levees. The
monotonous stream with uniform bed morphology also
offered few breeding or spawning areas for birds, sh, or
aquatic organisms.
After a series of large oods between 1960 and 1980,
authorities realized that the condition of the Thur was not
sustainable and decided to dedicate funds and effort to
improving ood protection while simultaneously enhancing
ecosystem health (Amt fur Umwelt 1999). As a result, local
rehabilitation measures have been taking place at some
reaches along the length of the river for the past 15 y.
Our case study focuses on a rehabilitation project
conducted on the Thur River in 2004 near the cantonal
border of Thurgau and Zurich (Figure 5). The median
discharge at this location is 49 m3/s with an annual ood
value of approximately 570 to 725 m3/s. Median water
temperature is 10.5 8C with a maximum of approximately
22.4 8C. Channel constraints were widened from an average
of 40 m to 120 m over a 1.5-km stretch by reducing oodplain

Integr Environ Assess Manag 8, 2012ME Borsuk et al.

height and returning this area to the river. To protect nearby

residences from large oods, the historical dikes were
maintained at a width of 200 m. No retention basins or side
channels were constructed. The widening project cost 9.9
million Swiss Francs (Sfr) and removed approximately 3 ha
from agricultural production. Gravel bed characteristics and
transport estimates for this section of the Thur are available
from Schalchli et al. (2005), and time series data of discharge,
suspended particle concentration, and water temperature
come from unpublished sources. We generate model results
for pre- and postrehabilitation conditions to estimate the
morphological, hydraulic, ecological, and economic effects of
the project. For some attributes, limited measurements are
available for comparison.
Case study results
If the study section of the Thur River were to be of
sufcient length and free of any width constraints, the model
predicts that the natural river form is much more likely to be
single-threaded (i.e., straight, meandering, or alternating
gravel bars with 89% probability) than multi-threaded (i.e.,
braided with 11% probability). However, the actual constrained channel morphology will depend on the gravel
supply, as well as the severity of width constraints, as
predicted by the pattern diagram of da Silva (1991)
(Figure 2).
With the actual rehabilitated width of 150 m and length of
1.5 km, there is a negligible chance of braiding, a 56% chance
of alternating gravel bars, and a 44% chance that the river
might still be straight. The model estimates that gravel
transport out of the study section is sufciently reduced to
yield gravel deposition rather than incising. Thus, gravel
supply should not be a limiting factor in the formation of
bank structures at this width. Indeed, since the 2004
rehabilitation, the river at this section has gone from highly
channelized to forming signicant gravel bar structures.
The model can be used to predict the mean ow depth and
velocity, as well as the distribution of habitat units (i.e., pools,
rifes, runs), for any discharge below bankfull. To illustrate
the difference in hydraulic conditions between the historically
straight and a rehabilitated alternating gravel bar morphology,
we used the modal discharge (Q 28 m3/s). We nd that the
2 morphologies do not differ considerably with respect to
mean depth and velocity, but a substantial increase in the
variability of velocity and depth is predicted for the
alternating form, as reected in the anticipated spatial
distribution of habitat units (Figure 6). For a straight
morphology, the hydraulic conditions are extremely homogeneous (mostly runs, with pools and rifes almost absent),
whereas for an alternating form, approximately 40% of the
river stretch will consist of either pool or rife habitat.
Model results indicate that ushing events will occur less
frequently after widening, due to the lower water depths and
thus lower dimensionless shear forces for a given discharge
compared to a constrained river. This leads to an average ne
particle content (assuming an alternating gravel bar form) of
7% (90% interval of 3% to 11%). Although siltation
conditions appear to worsen as a result of rehabilitation, it
is important to note that a widened river reach will contain
20% to 30% more rifes than a constrained river, and these
can be expected to remain clear of nes due to their hydraulic
conditions.

A Bayesian Network Model for River RehabilitationIntegr Environ Assess Manag 8, 2012

469

Figure 5. The Thur River at the location of rehabilitation. (Top) Map of Thur watershed. (Bottom left) June 2001, before the river was widened. (Bottom right)
May 2004, after the river was widened. Data source: swisstopo: Vector252006, DHM2520036, GG252006 (reproduced with permission of swisstopo/JA
100119). Photographs: C. Herrmann, BHAteam, Frauenfeld.

The total invertebrate biomass is expected to vary seasonally, with median predicted values after widening ranging
from 1 to 23 g dry mass m
2 with an uncertainty factor of
approximately 2 (Figure 7). Measurements taken in June
at this location led to total invertebrate biomass estimates of
2.8 to 5.3 g dry mass/m
2 (P. Baumann, personal communication). Periphyton biomass is expected to stabilize at
approximately 16 g dry mass/m
2 (results not shown).
Bed-moving oods are expected to be rare in the rehabilitated
condition, leading to little disruption of the benthic
population.

Total riparian arthropod abundance is predicted to be

26.5  6.5 individuals/m2 with the majority consisting of
spiders and ground beetles. This is a signicant improvement
relative to the 16.8  5.8 individuals/m2 predicted for the
channelized state of the river.
The study region of the Thur River lies in the transition
zone between ecoregions supporting grayling and barbell sh
species (Huet 1959). Thus, resident brown trout are not
expected in this reach. Indeed, electroshing over 40 sections
in this area found only 19 brown trout (Schager and Peter
2005). Fortunately, the brown trout portion of the model has

470

Figure 6. Predicted spatial frequency of habitat units by morphological type.

been tested successfully in a variety of other river reaches,

both channelized and natural (Borsuk et al. 2006).
Finally, for the 9.9 million Sfr spent on the project, one can
expect a total of 13.8  0.4 million Sfr in increased output
and 78  4 new full-time equivalent (fte) employment
positions. However, these short-term job gains will be
partially offset by the long-term loss of approximately 3.0
fte agricultural positions, and 0.4  0.6 million SwF in lost
economic output per year, as a result of taking 3 ha of
agricultural land out of production to provide for the wider
river corridor.

DISCUSSION
A variety of models and analyses have been developed to
anticipate the outcomes of river rehabilitation measures such
as corridor widening. However, to our knowledge, the model
described here is the rst attempt to combine models of
multiple features into an integrative forecasting tool. Having a
model that links actions to outcomes using the language of
probabilities is essential for making rational choices according

Figure 7. Predicted total invertebrate mass by day of year. Lines represent

bounds on indicated predictive intervals. Jagged lower edges are the result of a
rarely occurring bed-moving ood disturbing the benthic population.

Integr Environ Assess Manag 8, 2012ME Borsuk et al.

to a decision theoretic framework (Reichert et al. 2007). We

believe that our model is applicable to rivers similar to those
from which the data used to estimate our submodels
originated. Specically, we believe it is appropriate to apply
the model to gravel bed rivers in mid latitudes, with a
relatively natural ow regime, mean discharge between 1 and
60 m3/s, and slope less than 2%.
The case study application of our model to a recently
rehabilitated section of the Thur River at the cantonal border
of Thurgau and Zurich, Switzerland, shows that the results
can yield useful information for rehabilitation planning.
Specically, the model predicts (correctly) that, after widening of the channel at this location from 40 to 120 m, gravel
transport out of the reach would be reduced and alternating
gravel side bars would develop. This morphology is judged to
be more attractive by stakeholders and likely to increase
recreational use (Hostmann et al. 2005b).
As a result of this more natural morphology, the variability
of ow velocity and depth is expected to increase and a
signicant number of pools and rifes are expected to
develop. This is likely to lead to greater siltation in some
areas, but overall a more diverse mosaic of silted and clean
gravel conditions will exist. Although the total invertebrate
population is not expected to change dramatically, the model
predicts a near doubling in the riparian arthropod community, thus providing increased food for birds, sh, and other
riverine fauna. Although brown trout, a desirable species, is
not expected to reside in this section of the river, grayling and
barbel populations are likely to improve as a result of the
added prey resources and more diverse habitat (Weber et al.
2009).
Economically, rehabilitation of the Thur appears to be a
good investment. A 40% return, in the form of indirect
stimulation of other sectors of the local economy is expected.
In addition, new employment of approximately 80 fte is
anticipated during the construction phase, which is only
partially offset by the loss of 3 fte/y in agricultural positions.
Although in our case study our model was used retrospectively, we anticipate the model being used to complement stakeholder assessments in providing decision support
for managers (Reichert et al. 2007). Such a procedure can
help guide selection of reach-specic rehabilitation measures.
In addition, we are extending our reach-oriented model to
one that can support prioritization decisions at the regional or
catchment scale over an entire river network.
Methodologically, our model is distinct in the Bayesian
network literature in using functional equations to link
continuous variables, rather than employing discrete variables
and conditional probability tables. Representing all variables
in an inuence diagram as being discrete does have the
advantage that software is readily available to handle all
possible calculations one would want to make with such a
model, including inference against the direction of arrows.
However, discretizing variables that are inherently continuous
is often a subjective process that introduces a degree of
imprecision into the model that would otherwise not exist.
This is because of the vagueness that arises from assigning all
values within a specied range of a continuous variable to the
same discrete state.
Another problem associated with discretization is that it
encourages vagueness in variable denitions. For example,
many studies have been published that dene states of
variables to be low, medium, and high, without giving precise

A Bayesian Network Model for River RehabilitationIntegr Environ Assess Manag 8, 2012

quantitative denitions. This opens up the possibility for

model developers or users to have very different ideas of what
the variable and its different states represent. This can lead to
errors in assessing the probabilities required of the model or in
applying the results for decision making. Therefore, although
we lose the ability to infer the values of ancestors from
observed values of descendants, we generally prefer to build
networks of continuous variables, described by exible
distributional forms, and linked by functional causal expressions.
Although we do not benet from the inferential strength of
Bayesian networks, the effort of model building was
signicantly facilitated by the use of the graphical network
structure. By highlighting subnetworks (referred to as cliques)
that are relatively independent of other parts of the full
network, the causal graph encourages decomposition of the
larger modeling problem into manageable submodels. As
discussed in the Model Implementation section, however,
special attention needs to be given to address any potential
discrepancies in spatial or temporal scale.

CONCLUSIONS
We have brought a variety of submodels relevant to river
rehabilitation planning together into a graphical probability
network that links management alternatives to ecological and
economic consequences. Although we have adopted the
causal graph formalism common to Bayesian networks, we
have not made probabilistic inference a priority, providing us
with added exibility in functional and conditional distributional representation. This improves forecasting precision,
which we believe to be a more important model attribute in
environmental management applications than the possibility
of performing causal diagnosis. Our case study exemplies the
potential use of the model in addressing socioeconomic and
ecological consequences of rehabilitation, in addition to
strictly morphological and hydraulic impacts. The former
are usually the concerns of greatest interest to stakeholders
and therefore model predictions of these consequences can be
an inuential factor in supporting local rehabilitation decisions.
AcknowledgmentThis study was supported by the RhoneThur project, which was initiated and funded by the Swiss
Federal Ofce for the Environment (BAFU), the Swiss Federal
Institute for Environmental Science and Technology (Eawag),
and the Swiss Federal Institute for Forest, Snow and Landscape
Research (WSL). We thank Rosi Siber for producing the map
in Figure 5.

REFERENCES
Allan JD. 1995. Stream ecology: Structure and function of running waters. London
(UK): Chapman and Hall. 338 p.
Amt fu
r Umwelt KantonThurgau. 1999. Die Thureine unberechenbare Gro
sse
(The Thur Riveran Unpredictable River). Frauenfeld, Switzerland. 12 p.
Bledsoe BP, Watson CC. 2001. Logistic analysis of channel pattern thresholds:
Meandering, braiding, and incising. Geomorphology 38:281300.
Borsuk ME, Stow CA, Reckhow KH. 2004. A Bayesian network of eutrophication
models for synthesis, prediction, and uncertainty analysis. Ecol Model
173:219239.
Borsuk ME, Reichert P, Peter A, Schager E, Burkhardt-Holm P. 2006. Assessing the
decline of brown trout (Salmo trutta) in Swiss rivers using a Bayesian probability
network. Ecol Model 192:224244.

471

Buijse AD, Coops H, Staras M, Jans LH, Van Geest GJ, Grifts RE, Ibelings BW,
Oosterberg W, Roozen FCJM. 2002. Restoration strategies for river oodplains
along large lowland rivers in Europe. Freshw Biol 47:88907.
da Silva AMAF. 1991. Alternate bars and related alluvial processes [MS Thesis].
Kingston (ON): Queens Univ.
Ernst AG, Baldigo BP, Mulvihill CI, Vian M. 2010. Effects of natural-channel-design
restoration on habitat quality in Catskill mountain streams, New York. Trans Am
Fish Soc 139:468482.
Gu
nther A. 1971. Die kritische mittlere Sohlenschubspannung bei
Geschiebemischung unter Beru
cksichtigung der Deckschichtbildung und der
turbulenzbedingten Sohlenschubspannungsschwankung (The critical mean
bottom shear stress for varying bed materials regarding the development of
armoured layers and uctuations in the bottom shear stress. Mitteilung Nr.3
der Versuchsanstalt fu
r Wasserbau, Hydrologie und Glaziologie. Zu
rich (CH):
ETH.
Hostmann M, Borsuk M, Reichert P, Truffer B. 2005a. Stakeholder values in decision
support for river rehabilitation. Arch Hydrobiol Suppl 155:491505.
Hostmann M, Bernauer T, Mosler H-J, Reichert P, Truffer B. 2005b. Multi-attribute
value theory as a framework for conict resolution in river rehabilitation.
J Multi-Crit Decis Anal 13:91102.
Huet M. 1959. Proles and biology of Western European streams as related to sh
management. Trans Am Fish Soc 88:15511563.
Jowett IG. 1993. A method for objectively identifying pool, run, and rife habitats
from physical measurements. N Z J Mar Freshw Res 27:241248.
Lasne E, Bergerot B, Lek S, Laffaille P. 2007. Fish zonation and indicator species for
the evaluation of the ecological status of rivers: Example of the Loire Basin
(France). River Res Appl 23:877890.
Liu YB, Gebremeskel S, De Smedt F. 2004. Simulation of ood reduction by natural
river rehabilitation using a distributed hydrological model. Hydrol Earth Syst Sci
8:11291140.
Lumina. 1997. Analytica for Windows, Users Guide, Version 1.1. Denver (CO):
Lumina Decision Systems.
McKay MD, Conover WJ, Beckman RJ. 1979. A comparison of three methods for
selecting values of input variables in the analysis of output from a computer
code. Technometrics 21:239245.
Meyer-Peter E, Mu
ller R. 1948. Formulas for bedload transport. In: Proceedings of
the International Association of Hydro-Environmental Engineering and
Research; 3rd Congress; Stockholm, Sweden. p 3964.
Millennium Ecosystem Assessment. 2005. Ecosystems and human well-being:
Wetlands and water synthesis. Washington (DC): World Resource Institute.
68 p.
Miller RE, Blair PD. 1985. Inputoutput analysis: Foundations and extensions.
Englewood Cliffs (NJ): Prentice-Hall. 464 p.
Miller RE. 1998. Regional and interregional inputoutput analysis. In: Isard W,
editor. Methods of interregional and regional analysis. Aldershot (UK):
Ashgate. p 41133.
Naiman RJ, Decamps H, Pollock M. 1993. The role of riparian corridors in
maintaining regional biodiversity. Ecol Appl 3:209212.
Noble RAA, Bredeweg B, Linnebank F, Salles P, Cowx IG. 2009. A qualitative model
of limiting factors for a salmon life cycle in the context of river rehabilitation.
Ecol Inform 4:299319.
Paetzold A, Yoshimura C, Tockner K. 2008. Riparian arthropod responses to ow
regulation and river channelization. J Appl Ecol 45:894903.
Pearl J. 2000. Causality: Models, reasoning, and inference. Cambridge (UK):
Cambridge Univ Press. 400 p.
Peter A, Kienast F, Woolsey S. 2005. River rehabilitation in Switzerland: Scope,
challenges and research. Arch Hydrobiol Suppl 155:643656.
Reichert P, Borsuk ME, Hostmann M, Schweizer S, Spo
rri C, Tockner K, Truffer B.
2007. Concepts of decision support for river rehabilitation. Environ Model
Software 22:188201.
Rohde S, Schuetz M, Kienast F, Englmaier P. 2005. River widening: an approach to
restoring riparian habitats and plant species. River Res Appl 21:10751094.
Rosgen DL. 1994. River restoration utilizing natural stability concepts. Land Water
38:3641.
Schalchli U. 1995. Basic equations for siltation of riverbeds. J Hydraulic Eng
121:274287.

472

Integr Environ Assess Manag 8, 2012ME Borsuk et al.

Schalchli U, Abbegg J, Hunzinger L. 2005. Geschiebestudie Thur und Einzugsgebiet

Strickler A. 1923. Beitrage zur Frage der Geschwindigkeitsformel und der

(Gravel report for the Thur River and its catchment area). Amt fu
r Umwelt,
Kantons Zu
rich, Thurgau, Appenzell and St. Gallen, Switzerland.

Rauhigkeitszahlen fu
r Stro
me, Kanale und geschlossene Leitungen
(Contributions to the formula of ow velocity and to the estimation of ow

Schager E, Peter A. 2005. Endangered Flow-Loving Cyprinids in the River Thur:

resistance for rivers, channels and closed conduits). Mitteilung Nr. 16 des Amtes

Status and Future. Study commissioned by the AWEL Zurich. Eawag Aquatic
Research, Du
bendorf, Switzerland. Available from: http://www.rhone-

fu
r Wasserwirtschaft; Eidg. Departement des Inneren, Bern, Switzerland.
Tyler JA, Rutherford ES. 2007. River restoration effects on steelhead populations in

thur.eawag.ch/THUR_bericht_05.pdf
Schweizer S, Borsuk ME, Jowett I, Reichert P. 2007a. Predicting joint frequency

the Manistee River, Michigan: Analysis using an individual-based model. Trans

Am Fish Soc 136:16541673.

distributions of depth and velocity for instream habitat assessment. River Res
Appl 23:287302.

Van den Berg JH. 1995. Prediction of alluvial channel pattern of perennial rivers.
Geomorphology 12:259279.

Schweizer S, Borsuk ME, Reichert P. 2007b. Predicting the morphological

and hydraulic consequences of river rehabilitation. Riv Res Appl 23:303

Walters CJ, Korman J. 1999. Cross-scale modeling of riparian ecosystem responses

to hydrologic management. Ecosystems 2:411421.

322.
Schweizer S. 2006. Predicting the consequences of river rehabilitation measures on

Ward JV. 1998. Riverine landscapes: Biodiversity patterns, disturbance regimes, and
aquatic conservation. Biol Conserv 83:269278.

morphology, hydraulics, periphyton and on invertebrates [PhD thesis]. Zurich

(CH): ETH-Zurich.

Ward JV, Tockner K, Uehlinger U, Malard F. 2001. Understanding natural patterns

and processes in river corridors as the basis for effective river restoration. Reg

Shields FD, Copeland RR, Klingeman PC, Doyle MW, Simon A. 2003. Design for
stream restoration. J Hydraulic Eng 129:575584.

Rivers Res Manag 17:311323.

Weber C, Schager E, Peter A. 2009. Habitat diversity and sh assemblage structure

Shipley B. 2002. Cause and correlation in biology: A users guide to path analysis,
structural equations and causal inference. Cambridge (UK): Cambridge Univ

in local river widenings: A case study on a Swiss river. River Res Appl 25:687
701.

Press. 332 p.
Singer MB, Dunne T. 2006. Modeling the inuence of river rehabilitation scenarios

Zarn B. 1997. Einuss der Flussbettbreite auf die Wechselwirkung zwischen

Abuss, Morphologie und Geschiebetransportkapazitat (Inuence of river

on bed material sediment ux in a large river over decadal timescales. Water

Resour Res 42:W12415.

width on interactions between discharge, morphology and transport

capacity). Mitteilungen der Versuchsanstalt fu
r Wasserbau, Hydrologie und

Spo
rri CC, Borsuk M, Peters I, Reichert P. 2007. The economic impacts of river
rehabilitation: a regional input-output analysis. Ecol Econ 62:341351.

Glaziologie, 154, Eidgeno

rich,
ssische Technische Hochschule, ETH), Zu
Switzerland. p 209.