Instream physical habitat modelling types: an analysis as stream hydromorphological modelling tools for EU water resource managers

Abstract

The introduction of the EU Water Framework Directive (WFD) is providing member state water resource managers with significant challenges in relation to meeting the deadline for ‘Good Ecological Status’ by 2015. Overall, instream physical habitat modelling approaches have advantages and disadvantages as management tools for member states in relation to the requirements of the WFD, but due to their different model structures they are distinct in their data needs, transferability, user-friendliness and presentable outputs. Water resource managers need information on what approaches will best suit their situations. This paper analyses the potential of different methods available for water managers to assess hydrological and geomorphological impacts on the habitats of stream biota, as requested by the WFD. The review considers both conventional and new advanced research-based instream physical habitat models. In parametric and non-parametric regression models, model assumptions are often not satisfied and the models are difficult to transfer to other regions. Research-based methods such as the artificial neural networks and individual-based modelling have promising potential as water management tools, but require large amounts of data and the model structure is complex. It is concluded that the use of habitat suitability indices (HSIs) and fuzzy rules in hydraulic–habitat modelling are the most ready model types to satisfy WFD demands. These models are well documented, transferable, user-friendly and have flexible data needs. They can easily be implemented in new regions using expert information or different types of local data. Furthermore, they are easily presentable to stakeholders and have the potential to be applied over large spatial scales. Integral care must be taken in the use of appropriate HSIs as these are the most sensitive part of the modelling and inaccurate results will be gained if not correctly formulated. If representative HSIs are not available, fuzzy rule-based modelling is recommended, but care must also be taken in the designing of the rule sets. For larger-scale modelling or when only few field data are available, generalized habitat models hold promise for quantifying habitat suitability based on average stream characteristics.

Keywords:

• Water Framework Directive
• water managers
• habitat modelling
• physical habitat
• management models

1 Introduction

Until the introduction of the EU Water Framework Directive (WFD), EU member states have had varying levels of water management legislation and assessment approaches for assessing the health of the biota–physical relationship of their streams (Dunbar et al. 2001, Tharme 2003, Harby et al. 2004). The WFD requires an integrated approach to manage three of the key components of stream ecosystems: water quantity, water quality and physical habitat that support the biology of the system (Commission 2000), with the assumption that there are links between stream biota and the three components of stream habitat (Poff et al. 1997, Humphries et al. 1999, Maddock 1999, Sullivan et al. 2004, 2006). The WFD implicitly assumes an underlying link between ecological status and abiotic quality elements; thus, a key aspect is to identify and assess the links between the physical and biological components of streams if a stream is found to be non-compliant biologically (Commission 2000). For example, some major variables to affect fish and invertebrate species carrying capacities in streams include physical habitat variables such as flow and velocity (Beecher et al. 1995, 2002, Poff et al. 1997, Humphries et al. 1999, Holm et al. 2001, Jowett et al. 2005, Suren and Jowett 2006, Schweizer et al. 2007), water depth (Egglishaw and Shackley 1982, Kennedy and Strange 1982, Beecher et al. 1995, Bardonnet et al. 2006, Schweizer et al. 2007), substrate composition (Heggenes 1988, Cobb et al. 1992, Knapp and Preisler 1999, Kondolf 2000, Vadas and Orth 2001) and cover (Heggenes and Traaen 1988, Vehanen et al. 2000, Vadas and Orth 2001). It is the knowledge of the relationships between such physical habitat variables and biota that is needed for many member states for both research purposes and implementation of the WFD (Commission 2000, Logan and Furse 2002).

Instream physical habitat modelling approaches could help water managers in the decision-making process in relation to biota–physical habitat links. They are modelling tools for predicting the suitability of a habitat to support a target species by quantifying the relationship between stream biota and physical habitat (Hardy 1998). Instream physical habitat modelling can be used to (i) identify the individual biota–physical habitat relationships, (ii) assess the quality of the physical habitat variables through its impact on the biota and (iii) predict likely biological responses if hydromorphological changes to a system are to occur, and make decisions on this information. They are currently used in both stream management and research worldwide (Tharme 1996, 2003, Jowett 1997, Dunbar et al. 1998, Parasiewicz and Dunbar 2001, Leclerc et al. 2003). However, instream physical habitat modelling types within the broader habitat modelling approaches differ in their applicability to biota–physical habitat assessments for EU water managers (Rosenfeld 2003, Harby et al. 2004, Jorgensen 2008). For models to be considered by water managers to be used now, they will need to specifically:

1. supply a clear understanding of the links between the biological and physical habitat elements using available data (e.g. monitoring data);

2. be user-friendly (i.e. managers use them without intensive training support);

3. be transparent (i.e. easily presented to different stakeholders); and

4. be widely applicable (i.e. can be applied on a large spatial scale and for different biota elements) (Pollard and Huxham 1998, Logan and Furse 2002).

It is not the intention of this paper to provide a detailed description of all the instream physical habitat modelling types (appropriate references are provided), but to provide an overview and analysis of the suitability and readiness of each habitat model type in relation to the above-mentioned criteria.

2 Instream physical habitat modelling types

2.1 Habitat suitability models

Habitat suitability index (HSI) modelling uses physical habitat variables to assess the physical quality of an area to support a target indicator species (Bovee 1986, Raleigh et al. 1986, Hardy 1998). HSIs are also referred to as ‘habitat suitability curves’. They are the biological basis of HSI modelling and represent the functional relationship between the selected physical habitat variable and the response of the selected species/lifestage (Bovee and Cochnauer 1977, Bovee et al. 1998). The suitability index varies between 0 (unsuitable) and 1 (most suitable) and provides a probability measure on how suitable a habitat is for a target species (Bovee and Cochnauer 1977, Bovee et al. 1998). They are commonly used for assessing site-specific suitability, but are also effective at a wider spatial range under a set of generalized criteria and can be used to predict outcomes if the input variables are manipulated (Raleigh et al. 1986).

HSIs can be derived from univariate suitability functions that consider individual habitat variables independently from one another (Figure 1(a)) (Bovee 1986) or from multivariate functions (Ahmadi-Nedushan et al. 2006) that consider the interaction of several habitat variables to determine the species response for the cumulative effects of several habitat variables (Figure 1(b)) (Vismara et al. 2001).

Figure 1 Examples showing (a) univariate HSI curve for the habitat variable ‘depth’ and (b) a multivariate HSI showing the species response to the cumulative effects of both habitat variables depth and velocity

Univariate HSIs are the simplest and most commonly used and are derived in three ways depending on the available information (Bovee 1986, Bovee et al. 1998, Parasiewicz and Dunbar 2001, Ahmadi-Nedushan et al. 2006): Category I – expert opinion indices are derived from both professional judgement (expert opinion) and life history literature; Category II – habitat use indices are based on the frequency of occurrence from in situ studies of habit use of the target species and Category III – preference indices are Category II adjusted for in situ habitat availability measured at the same time as habitat use sampling (Bovee 1986). Categories II and III are the most ecological defensible due to their use of local in situ environmental data, but Category I is considered useful in data-poor situations and is also useful to generate ‘regional’ curves, applicable to all rivers of a certain hydrological and morphological ‘type’ within the same region. Category I HSIs can also be upgraded as more information becomes available. Category III HSIs can be derived using a number of ‘index of selection’ techniques; Ivlev index (Ivlev 1961), Jacobs selectivity index (Jacobs 1974) and forage ratio (Edmondson and Winberg 1971), the simplest and most common being the forage ratio. The equation for the forage ratio is as follows:

where P _i is a measure of preference for each variable category (i), U _i a percentage of utilization of a specific interval (e.g. 0–0.1 m for depth) by the target species for the measured variable and A _i the percentage of available habitat within each variable category. The more complicated Jacobs selectivity index is used less frequently, but preferred by some authors (Dunbar et al. 2001). The equation for the Jacobs selectivity index is as follows:

where R _i is the proportion of habitat used in that variable category and P _i the proportion of habitat available in that category. Positive indices indicate selective use of that habitat class and negative values indicate avoidance of that habitat class.

HSIs can be defined to take into account diel, seasonal and size class/age variation in habitat use or preference, all of which must be considered if suitable HSIs are to be used (e.g. a set of HSIs developed for a certain size class of fish may not be applicable to another, or a set of HSIs developed in summer will most likely not be applicable in winter). HSIs of individual habitat variables such as velocity, depth and substrate can be combined to give composite suitability indices (CSIs) expressing an overall assessment of the physical habitat quality of a defined river reach (Bovee 1986, Vadas and Orth 2001, Beecher et al. 2002). CSIs can be calculated in different ways with the ‘product equation’ being the most common. The equation assumes that habitat variables are independent of each other and of equal importance. The equation is

If one of the HSIs is zero, then the CSI will yield zero no matter the values of the other HSIs. This equation can be modified to include weighting factors that consider the relative importance of each variable to the target biota (Ahmadi-Nedushan et al. 2006).

Alternative methods for calculating CSI include the ‘arithmetic mean’, where high-quality habitat variables can compensate for poor quality habitat variables, as it takes an average of the HSI values. The ‘geometric mean’ is similar to the arithmetic mean, but instead of taking a direct average of the sum of the HSI values, the values are multiplied together and the nth root of the product taken, so some compensation is present, but similar to the product equation yields zero if one of the variables gets zero habitat suitability. The method ‘lowest HSI’ assumes that the lowest quality habitat variable is the most limiting, so determines the upper limit of the CSI, and high-quality HSIs cannot compensate for poorer ones (Korman et al. 1994).

2.2 Fuzzy rule-based modelling

Fuzzy rule-based modelling is being proposed as an alternative to the HSI modelling types for stream biota–habitat analyses (Silvert 2000, Ahmadi-Nedushan et al. 2008). Fuzzy rule-based modelling does not use conventional input data sets known as ‘crisp sets’ (e.g. 0.1–0.2 m for depth), as in conventional HSI modelling (Figure 1(a)), but uses IF–THEN rules that are more flexible with biota suitability being able to partially belong to different sets (Figure 2) (Adriaenssens et al. 2004, 2008). Fuzzy rule-based suitability data are presented in the form of verbal type commands similar to the way a human brain thinks (e.g. low, medium, high), and the CSI attained is the result of how the sets of fuzzy rules are defined (Ahmadi-Nedushan et al. 2006). All possible combinations have to be considered and the output from the model is also fuzzy until a transformation process called ‘defuzzification’ (Ahmadi-Nedushan et al. 2008) is used and the ‘crisp output’ data presented as CSIs (Adriaenssens et al. 2004, Harby et al. 2004, Silvert 1997).

Figure 2 An example of a fuzzy set describing the relative degree of habitat suitability membership (i.e. low or high) for the habitat variable depth

Fuzzy rule-based approaches have been used for fish (Jorde et al. 2001, Kerle et al. 2002, Mouton et al. 2008) and benthic invertebrate studies (Adriaenssens et al. 2006, Van Broekhoven et al. 2006, 2007). They are also being hybridized with other types of approaches such as artificial intelligence-based models to increase their coverage beyond HSI modelling types (Adriaenssens et al. 2004, Ahmadi-Nedushan et al. 2006).

2.3 Linear models

Linear regression can be used to investigate biota–physical habitat relationships (Gordon et al. 2004, Harby et al. 2004). Linear regression approaches follow four main assumptions: (i) the observations are independent, (ii) the residuals are assumed to be identically and individually distributed, (iii) model residuals are assumed to follow a normal distribution and (iv) the regression function is a linear function of its parameters.

2.3.1. Simple linear regression

Simple linear regression (SLR) assesses habitat variables independently using a univariate function (Quinn and Keough 2006). The equation is

where Y _i is the value of the dependent variable, A the slope of the regression line, X _i the independent variable, B the y-intercept of the regression line and E _i the random error.

2.3.2. Multiple linear regression

Multiple linear regression (MLR) approaches consider the joint effect of multiple explanatory variables and their interactions. The equation is

where Y _i is the value of the dependent variable, B ₀ a constant, X _1i , X _2i , … , X _mi the values of m independent variables and E _i the random error. The same assumptions hold as for SLR models, but because they treat more than one independent variable MLR analyses are one of the most common approaches used in physical habitat modelling approaches for explanatory analysis phase (Ahmadi-Nedushan et al. 2006). MLR techniques have been used primarily for fish (Annoni et al. 1997, Bult et al. 1999, Inoue and Nakano 2001) and benthic invertebrates (Brey et al. 1996, Gray 2005, Jacobsen 2005, Kennard et al. 2006).

2.3.3 Generalized linear models

Generalized linear models (GLMs) allow the dependent variable to have any distribution from the exponential family (e.g. binomial, Poisson, etc.); however, a transformation is used to achieve a linear dependency of the explanatory variables. This makes GLMs more flexible for analysing biota–physical habitat relationships based on input data that are not normally distributed (Agresti 1996). The equation is

where Y is a vector representing a set of outcome variables, X a set of pre-program variables or covariates, B ₀ the set of intercepts (value of each y when each x = 0) and B a set of coefficients, one for each X. Although GLMs have major advantages over the SLR or MLR, they have not been extensively used in instream physical habitat modelling approaches (Ahmadi-Nedushan et al. 2006). Labonne et al. (2003) used GLMs to determine the fish–habitat relationships and concluded that GLMs were an effective analysis tool for biota–hydrological studies, especially if the assumptions of simpler methods could not be met.

2.3.4. Logistic regression

Logistic regression (LR) is appropriate for binary dependent variables such as the presence/absence (Jongman et al. 1995). Even though this methodology requires more data, an advantage is that binary data are common in instream studies (Filipe et al. 2002). LR is used similar to MLR (Guay et al. 2000); however, LR is based on fitting data to the logistic curve:

where P is the probability of occurrence, Z = B ₁ X ₁ + B ₂ X ₂ + … + B _n X _n  + A, X _{1,  … , n} the significant independent variables, B _{1, …, n} the coefficients and A a constant (Harby et al. 2004).

LR has been widely applied in the field of habitat modelling for fish (Geist et al. 2000, Guay et al. 2000, Garland et al. 2002) and benthic invertebrates (Udevitz et al. 1987, Peeters and Gardeniers 1998, Castella et al. 2001), and also macrophytes to a lesser extent (Narumalani et al. 1997, Wagner and Falter 2002, Van den Berg et al. 2003). It was successfully used to predict habitat selection by Apache trout in six east-central Arizona streams (Cantrall et al. 2005), also being instigated successfully to study fish–habitat relationships in a Californian river (Harvey et al. 2002), and used it to identify the most important environmental variables affecting freshwater fish in a semi-arid stream in Portugal (Filipe et al. 2002).

With all parametric-based analyses types, there are inherent problems when dealing with biota–physical habitat relationships as many of the assumptions of the modelling are not realized (i.e. data are not normally distributed data or there is no linear dependency), and transformations are also inherent with problems (i.e. more difficult to interpret results) (Townend 2002, Quinn and Keough 2006).

2.4 Non-parametric regression models

Non-parametric models are available and are starting to be used as an alternative to parametric-based types. If the assumptions of parametric regression approaches do not hold, then non-parametric techniques may be more appropriate (Quinn and Keough 2006, Jowett and Davey 2007).

Although these techniques have been in use for over 30 years, it is only in the last 10 years that they have been used in instream physical habitat modelling approaches. One technique that is gaining popularity is the general additive model (Jowett and Davey 2007) and to a lesser extent the multivariate adaptive regression splines (MARS) (Leathwick et al. 2005, 2006, Elith and Leathwick 2007).

The generalized additive models (GAMs) are a non-parametric extension of GLMs (Guisan et al. 2002). They use smooth functions of the explanatory variables, allowing the form of a relationship between the independent variable and the dependent variable to take on any shape as defined by the data (Jowett and Davey 2007). GAMs provide more flexibility in biota–physical habitat analyses because of their ability to deal with curvilinear relationships and non-monotonic relationships between independent and dependent variables. The only assumptions of GAMs are that functions are additive and the components are smooth (Guisan and Zimmermann 2000).

The MARS models are similar to GAMs as they make no assumption about the underlying functional relationship between the dependent and the independent variables, and they allow the independent variables to have any type of distribution (Friedman 1991, Hastie et al. 2001, Leathwick et al. 2005). MARS partitions the input space into regions, each with its own regression equation, which makes it suitable to solve problems involving more than two variables. MARS has been used to accurately describe the relationship between fish distribution and explanatory variables (Leathwick et al. 2006).

2.5 Artificial neural networks

Artificial neural networks (ANNs) evolved during the 1990s as a ‘black box’ nonlinear modelling type which is used to model complex nonlinear relationships between dependent and independent data, similar to the way a human brain works (Lek et al. 1996a, 1996b, Lek and Guegan 1999, May et al. 2008). Because biota–habitat relationships are often complex and non-linear, leading to problems with parametric regression-based approaches (Lek and Guegan 1999), ANNs have been proposed to overcome some of these problems (Lek et al. 1996b, Brosse et al. 1999a, 1999b, 2001). ANNs are able to learn and then identify relationships between input and output data, even if the data are imprecise and noisy. Owing to their ability to model (i) non-linear, (ii) complex multivariate data, (iii) noisy data and (iv) learning and adaptation ability, they are increasingly being used in aquatic habitat modelling (Dedecker et al. 2005, Dakou et al. 2006, Park et al. 2006). There are no specific raw physical habitat data requirements for ANNs, and similar to HSI and fuzzy rule-based modelling, any quantitative variable can be used as an input or output variable.

Two types of ANNs are being applied in habitat modelling approaches (Harby et al. 2004), that is, backpropagation networks (Lek and Guegan 1999) and Kohonen self-organizing maps (Lasne et al. 2007), with backpropagation networks being most widely used for fish (Lek et al. 1996a, Mastrorillo et al. 1997, Brosse et al. 1999b) and benthic invertebrates (Dedecker et al. 2001, 2005, Lencioni et al. 2006). Kohonen self-organizing maps have been used for benthic invertebrates (Lencioni et al. 2006).

2.6 Individual-based or agent-based modelling approaches

Individual-based models (IBMs) have been developed as an advanced alternative to conventional habitat suitability modelling (Bonabeau 2002, Grimm and Railsback 2005). They work at the individual level in the population and model biological responses (e.g. growth, fecundity, mortality, movement) of that individual to physical habitat variables (Stillman 2008). It is assumed that the target biota selects habitat to maximize their potential fitness (Railsback and Harvey 2002), and individuals (also called agents) will execute certain behaviour in relation to their interactions with their surroundings. The models are built by using IF–THEN rules to elucidate behavioural responses to changes in the habitat variables (Railsback and Harvey 2002). IBMs have been used in biota–physical habitat studies mostly for fish (Van Winkle et al. 1998, Railsback and Harvey 2002, Grand et al. 2006, Harvey and Railsback 2007).

Two emerging IBM sub-types are biological process-based models and bioenergetic models (Harby et al. 2004). Biological process-based models are an extension of conventional instream physical habitat modelling assessing the habitat suitability for a species by using factors such as food availability, movement ability, foraging behaviour and metabolic processes (Hayes et al. 2000). A simple biological process-based model could be fish growth under different physical habitat conditions, whereas a more complex model might include other variables such as feeding, reproduction or digestion rates (Hayes and Jowett 1994).

Biological process-based models have been mostly developed for drift-feeding fish (Hughes and Dill 1990, Hughes 1992, 1998, Hayes et al. 2000, Hughes et al. 2003, Booker et al. 2004). Bioenergetic models are a special type of biological process-based model that defines the mass balance relationship between the amount of food consumed by a fish and its growth, and optimal fish location is based on energy budgets (Brant and Hartmann 1993, Hansen et al. 1993, Hill and Grossman 1993, Ney 1993). Bioenergetic models require inputs of bioenergetic parameters, water temperature data, growth data, energy density data and diet data (Brant and Hartmann 1993). Bioenergetics modelling has been used in fish habitat modelling to evaluate habitat quality (Booker et al. 2004), predator–prey dynamics (Stewart and Ibarra 1991), trophic interactions (Pothoven et al. 2001) and feeding patterns of fish on macroinvertebrates (Rand et al. 1995).

3 Hydraulic–habitat modelling

3.1 Conventional hydraulic–habitat models

Instream physical habitat models are often combined with a hydraulic model to form a ‘hydraulic–habitat’ model (Bovee and Milhous 1978, Stalnaker et al. 1995). The hydraulic–habitat model is very relevant as a water management tool under WFD conditions because it represents flow as the driver of the ecology within the river system (Poff et al. 1997, Bunn and Arthington 2002, Naiman et al. 2002), and this in turn affects the habitat quality available to the biota. Hydraulic–habitat models are mostly used when anthropogenic changes in flow regime (e.g. hydropower, water abstraction) are suspected to affect biota (Gibbins and Acornley 2000, Thorn and Conallin 2006, Olsen et al. 2009). Typically, hydraulic–habitat models are using HSI or fuzzy rule-based model types to represent the physical habitat requirements of species. Instead of giving a single HSI or CSI, the output of hydraulic–habitat models is a functional representation of the habitat availability quality for indicator biota under different flows (Bovee et al. 1998). For any surveyed river reach, the modelled habitat availability is expressed as a weighted usable habitat area (WUA) in units of square metre habitat per metre river length. Typically, as discharge increases from zero, WUA rises until an optimum flow, and then it decreases as discharge becomes too high to sustain optimal habitat conditions (Stalnaker et al. 1995). This functional representation can be converted into a hydraulic habitat suitability (HHS) index ranging between 0 and 1, similar to a HSI or CSI (Bovee et al. 1998).

Combined hydraulic–habitat models are probably the most widely applied instream physical habitat modelling type in the world in general and Europe in particular (Tharme 2003), and will be especially important for EU water managers where flow regime is being modified. There are already many well-established hydraulic–habitat models that have been tested on many stream types in Europe (Harby et al. 2004). The most famous and widely used approach is the instream flow incremental methodology and in particular one of its major components, the physical habitat simulation system (Bovee et al. 1998, Hardy 1998). This basic modelling system has been widely adapted and modified at the international level such as RHYHABSIM (Jowett 1992, 2006), CASIMIR (Schneider and Peter 1999, Jorde et al. 2001, Mouton et al. 2008), EVHA (Dunbar et al. 1998) and RHABSIM (Payne and Associates 2000). These models have been employed worldwide, and although they have been criticized for things such as transferability of HSIs and lack of ecological predictive capability (Orth and Maughen 1982, Mathur et al. 1985), they are also considered to be ‘the most scientifically and legally defensible methodology available’ for instream physical habitat assessments in many countries (Gore and Nestler 1988, Dunbar et al. 1998, Tharme 2003).

Two of the main advantages of hydraulic–habitat modelling are that multiple hydromorphological drivers (e.g. water use, morphological degradation) can be handled simultaneously, and physical habitat changes can be assessed at larger spatial scales (Cavendish and Duncan 1986).

3.2. Generalized habitat models

Generalized habitat models (also called statistical habitat models) are a modelling type being proposed as an alternative to conventional hydraulic–habitat modelling approaches (Lamouroux and Jowett 2005). Considerable field effort and experience is needed to run the conventional hydraulic–habitat models, whereas generalized habitat models obtain reach scale habitat values (HVs) based on a limited number of field measurements (Lamouroux et al. 1998), HVs range between 0 and 1 and are defined as the WUA divided by the wetted area of the reach. The generalized habitat models use reach hydraulic geometrical relationships (i.e. mean depth–discharge and mean width–discharge relationships) and simple average reach descriptions (i.e. mean particle size, Reynolds number) to estimate HV indices from discharge simulations (Lamouroux and Capra 2002, Stewardson and McMahon 2002, Stewardson 2005). It was shown that a limited reach scale data set collected at two different discharge levels was sufficient for a generalized habitat model to predict reach scale WUAs which were comparable with the ones predicted using the more data-demanding conventional hydraulic–habitat model types (Lamouroux and Capra 2002, Lamouroux and Jowett 2005). This makes the generalized habitat models particularly valuable for large-scale assessments or when only few reach data are available.

4 General discussion

Water managers are constantly in the process of adapting their existing national and regional assessment methods, modifying methods used in other countries or developing new methods to help them to meet the challenges of water resource management (Rosenfeld 2003), and for EU member states, it is essential for the implementation of the WFD (Wallin et al. 2003).

When deciding on modelling approaches for biota–physical habitat relationships, EU water managers must compromise between the accuracy of the approach and the policy, budget, timescale, expertise required, and ecological and hydromorphological data requirements of the approach (Jorgensen and Bendoricchio 2001). A trade-off between ‘optimizing accuracy’ versus ‘optimizing generality’ must be made in relation to management-based models (Guisan and Zimmermann 2000). Owing to a lack of environmental data, limited resources and the need to act immediately, certain modelling approaches are excluded or unsuitable as management-based models (Rosenfeld 2003).

Owing to the relatively short timeframe of the WFD, modelling types that require large amounts of data still to be collected will be substituted for model types that can use the already available data (e.g. monitoring data), even if this means decreasing the accuracy of the approach, and complex models substituted for simpler types (Rosenfeld 2003). Table 1 provides an overview of the various instream habitat modelling approaches in relation to the four criteria set out in Section 1. A scoring system has been used to assess the overall utility of the models for water managers (Table 1), and a more detailed discussion is given for each approach below.

Table 1 Summary of different instream physical habitat modelling types in relation to their applicability as modelling tools for EU water managers

Approach	Biological and physical links	User-friendly	Transparent	Widely applicable	Overall score
HSI models	***	***	***	***	***
Fuzzy rule-based models	***	**	**	***	**½
Linear models
SLR	***	***	**	*	**¼
MLR	***	***	**	*	**¼
GLMs	***	**	**	*	**
LR	***	**	**	*	**
Non-parametric regression models
GAMs	***	*	**	*	*¾
MARS	***	*	**	*	*¾
ANNs	***	*	*	***	**
IBMs	***	*	**	***	**¼
Applicability rating: *, low; **, medium; ***, high.

4.1 HSI models

HSI-based approaches have been in use for over 30 years and their popularity continues to grow probably due to their recognition worldwide as legitimate approaches when dealing with biota–hydromorphological analyses (Jowett 1997, Leclerc et al. 2003, Gordon et al. 2004, Harby et al. 2004). HSI approaches also deal with many types of data (e.g. continuous, discrete, binary, categorical) simultaneously and have been developed and tested for different biota groups (albeit mostly fish and benthic invertebrates). They can incorporate coarse monitoring data and expert opinion, and there are readily available user-friendly computer programs, many of which are used in combination with a hydraulic model (Bovee et al. 1998). They are also applicable over a wide spatial scale, but it is essential that care must be taken to make sure that the individual HSIs used in the modelling accurately represent the biota–physical habitat relationships of the area they are to be used in. Incorrect use of HSIs is a significant source of error in this modelling, mostly due to HSIs being transferred from one area to another area without adequate testing to determine transferability (Thomas and Bovee 1993, Moir et al. 2005).

Ideally, when using HSIs, they should be developed on a site-by-site basis, but this is impractical. Instead, Category III HSIs should be used as these habitat preference curves have been found to be more transferable than HSIs which are based on observed habitat use (Beecher et al. 1993, Dunbar et al. 2001). However, adequate evaluation of transferability should be conducted (Thomas and Bovee 1993, Groshens and Orth 1994), even if only in a literature review sense. If confidence is taken in the HSIs, then the conventional HSI methods are well suited for water managers, and the HSI approach is a clear candidate as a WFD management-based modelling approach.

4.2 Fuzzy rule-based modelling

Fuzzy rule-based approaches have advantages similar to HSI approaches, but can also better handle imprecise and qualitative data which are common with ecological data. The better use of imprecise and uncertain data including expert opinion makes the fuzzy rule-based approaches more appealing than HSI modelling, especially in data-poor situations (Adriaenssens et al. 2004, Mouton et al. 2008). Furthermore, they consider multivariate interactions of input variables, new parameters can easily be added and the system is relatively easy to implement (Ahmadi-Nedushan et al. 2008), all of which are attributes attractive to water managers.

However, fuzzy rule-based models are still in their exploration stage and it may be difficult to convince water resource managers of the benefits of such subjective methods when they are looking for defendable data-driven-type approaches (Adriaenssens et al. 2004). Another problem is the determination and tuning of the membership functions and fuzzy rules. This is not a straightforward procedure and relies entirely on the opinion of experts (Chen and Mynatt 2003). Expert opinion is obviously a major concern, both in the choice of input variables and in the formulation of the rules, but for some biota groups (e.g. macrophytes, diatoms) expert opinion will be all water managers have, no matter what method is used. To overcome these problems (Cornelissen et al. 2001), (i) stressed criteria are needed to determine necessary qualifications of experts, (ii) proper elicitation of expert knowledge is necessary to construct membership functions and (iii) methods are needed to test the validity of a membership function.

Another solution that is being investigated to improve the flexibility of the system is the hybrid fuzzy approach, whereby fuzzy models are combined with techniques such as principal component analysis, genetic algorithms or ANNs (Meesters et al. 1998, Loia et al. 2000, Mackinson 2000), but at the same time it increases the complexity of the modelling process, something water managers want to avoid (Adriaenssens et al. 2004).

4.3. Parametric regression techniques

Linear regression models are robust if their assumptions are justified. Unfortunately, in stream ecology, these assumptions are rare, and even transformations of the data to help meet these assumptions are inherent with certain problems (Quinn and Keough 2006). However, linear regression techniques are well established, popular, available in most commercial computer packages and can deal with most types of data (Townend 2002).

SLR has been criticized because biota–hydromorphological relationships in streams are known to be a multivariate process (Jowett 1997, Geist et al. 2000, Guay et al. 2000, Vadas and Orth 2001, Vismara et al. 2001, Neumann and Wildman 2002, Leclerc et al. 2003) and treating variables as independent does not always give an adequate representation of the true situation (Nykanen et al. 2001). However, SLR is often used as a ‘first indication’ of which variables are important and then more appropriate methods are applied (Quinn and Keough 2006). MLR deals with more than one independent variable but a potential problem with MLR models is multicollinearity (Affifi and Clark 1996). Many variables are known to be highly correlated with each other (depth, velocity, substrate, etc.), and the final set of variables in the model may be a result of strong associations with other variables and can lead to redundant variables being included or important variables being excluded from the modelling (Ahmadi-Nedushan et al. 2006). Another drawback is that most parametric models are built up using local habitat data so they are site-specific, thereby leading to problems in transferring models between sites or comparing sites from different models, or extrapolating between sites, which are common requirements for water managers. Consequently, these types of techniques are not widely applicable and best suited in the explanatory phase for determining the importance of habitat variables to use in a later modelling phase.

4.4 Non-parametric techniques

GAMs have only been used within the last 5–10 years in instream physical habitat modelling. Jowett and Davey (2007) compared GAMs against traditional habitat suitability criteria for both fish and invertebrates. They found that the GAM performed slightly better, but it was more complicated and the user had to be cautious not to over-fit the data in which case the model may cause minor fluctuations in the data to be exaggerated. The main issue with GAMs is that they need a lot of data to make their use worthwhile. If these data are available, then there may be situations where they give better results in identifying sharp thresholds in habitat selection (Jowett and Davey 2007). MARS works with data sets similar to GAMs but according to Leathwick et al. (2006) they hold some advantages. They found that MARS results were easier to interpret ecologically. Furthermore, it is computationally much faster and more easily incorporated into other analyses using, for example, GIS modelling (Munoz and Fellicísmo 2004) than approaches such as GAMs and ANNs. However, both techniques require large data sets that may not be readily available.

As analysis types such as GAMs and MARS are used increasingly in research to supplement parametric-based models, they will become better documented and probably more acceptable as management-based analysis tools. But for the present, water managers should probably focus on more established parametric model types and only use more advanced techniques such as GAMs and MARS if transformations of the input data are not possible to fulfil the requirements of parametric models. However, this does not mean that they should be excluded if water managers have knowledge of them, as they hold advantages over those of parametric regression models.

4.5 Artificial neural networks

For water resource managers, ANNs have advantages in being able to process complex, nonlinear and noisy data, which is often the case with ecological data. Any type of raw data (i.e. discrete, continuous and dichotomous) can be used in ANNs (Lek and Guegan 1999).

Unfortunately, there are also some major limitations for water resource managers when it comes to using ANNs. First, ANNs often require large sets of training data. This is a problem as background ecological data are often limited, leading to too few data being used to train the model, and considerable time and resources can be spent gaining this information (Lek and Guegan 1999). Second, there are very few ANN ‘package’ models available. Because each study usually has varying amounts of input data, different amounts of input neurons, hidden layers, etc., are needed. This means that ANNs are usually designed specifically for each study being programmed using network toolboxes or by building new computer codes. For this purpose, the user needs to be experienced in computer programming and model building, and the model is therefore not easily transferable (Harby et al. 2004). Furthermore, there is also no standard method for choosing the optimal ANN characteristics, such as the number of hidden neurons and transfer functions. This leads to difficulties in following the pathways and defining the most important variables (Olden and Jackson 2002, Harby et al. 2004).

ANNs have also been criticized for their ‘black box’ approach, in the sense that ANNs do not provide a clear insight or understanding of the pathways from the input variables to the predicted output variables. This makes it very difficult to present the model to stakeholders. Therefore, ANNs are probably still better suited to the applied research field, but they may also be useful in site-specific situations where enough data are available. However, at this stage, they are not well suited for many member states as water management models, where the need to apply output at large spatial scales is required.

4.6 Individual-based approaches

IBMs provide advantages in which they require little in situ historical background data (e.g. field data) as they work off biological behavioural response–physical habitat relationships (Railsback et al. 2003, Harvey and Railsback 2007). Once these are established, they should be universal (Grimm and Railsback 2005). IBMs are also advantageous in that they are able to model complex nonlinear relationships where space is crucial (bottlenecks), and the positions of the individuals are not fixed (Railsback and Harvey 2002). The individuals exhibit complex behaviour which includes learning and adaption behaviour (Bonabeau 2002). This makes IBMs ideal for biota such as fish and to a lesser degree macroinvertebrates (Grimm and Railsback 2005). The major limitation is the determination of the critical biological responses to physical habitat (especially within bottlenecks) which are poorly known for most biota. For this purpose, expert opinion must be relied on to make the rules, thereby increasing subjectivity. Another problem is that IBM computer packages are not widely available and therefore not easily accessible.

From an EU water manager's view, IBMs are still not developed enough to be seriously considered as a WFD management model, as there are insufficient data to be applied at a large scale, and the hydraulic modelling is complex (usually 3D). Currently, the modelling is limited to drift feeding fish such as salmonids. Therefore, at this point in time, it is better suited for academic or research modelling approaches. However, it has the potential to be coupled with conventional physical habitat suitability modelling approaches to better understand and predict biota–physical habitat relationships or to study specific problems identified using coarser habitat model types (Grimm and Railsback 2005).

4.7 Hydraulic–habitat modelling

In order for the instream physical habitat models to be applicable at the larger scales often considered in water management, a combined hydraulic–habitat model should be used. This requires that the habitat model which is built using microscale data is applicable at river or reach scale. For this purpose, the fuzzy logic and HSI-based habitat models are most useful because of their flexible data use and their functional representation which means that the transferability can be accomplished using expert opinion.

Because of model assumptions that cannot be satisfied and a close empirical data dependency of parametric and non-parametric regression methods, it is very uncertain that these models can be transferred to other regions. Owing to their flexible data use, ANNs also have a high potential for larger-scale applications; however, this model type requires a great deal of data for training. If the model has been trained in a region with different data availabilities, experience in computer programming and model building is needed to adopt the model to a new region. IBMs also hold promising perspectives as they are based on biological behaviour responses and therefore less dependent on in situ data. However, detailed information on the spatial variation in physical habitat variables is required, and the hydraulic model needs a 3D structure which complicates its operation.

Among the two most suitable instream physical habitat model candidates, the HSI and fuzzy logic-based methods, the use of representative HSIs is very important for successful HSI modelling. However, the fuzzy rule-based model can be based solely on expert opinions, but this method can be very subjective and perhaps less convincing in the eyes of managers and other stakeholders. Consequently, if HSIs are validated, the HSI model is probably preferable, but if data availability is sparse, the fuzzy rule-based method should be used. Fuzzy-based rules can also be based on available field data. If field observations are imprecise, fuzzy rule-based curves are the preferred option.

4.8 Generalized habitat models

Generalized habitat models essentially perform the same function as hydraulic–habitat models, but hold advantages over them as they do not inherit their complexity and their high field and biological data requirements (Lamouroux et al. 1998, Lamouroux and Capra 2002). However, due to their relative simplicity, information is sacrificed. This loss of information must be weighed up against the intensive fieldwork requirements of conventional hydraulic–habitat models, especially if HSIs have to be devised for the conventional modelling approach. Therefore, this model type is very attractive to water managers as they are widely applicable with a limited amount of field work (Lamouroux and Jowett 2005, Lamouroux and Cattaneo 2006).

5 Conclusion

As a management-based modelling approach, there are substantial differences between the instream physical habitat modelling types available to water managers. HSI and fuzzy logic-based methods appear to be most ready for operational application by water managers. These models are well documented, user-friendly, easily presentable to stakeholders and transferable to be applied over large spatial scales. Furthermore, they can deal with different types of data and they are applicable to a range of biota. However, for HSI modelling, it is essential that care is taken in the use of appropriate HSIs as these are the most sensitive part of the modelling. Inaccurate results will be gained if HSIs are not correctly formulated.

Other methods are also suitable, especially parametric and non-parametric approaches but restrictive model assumptions and their applicability over a limited spatial scale remains a substantial disadvantage. Fuzzy-based models hold promise in either replacing HSIs or supplementing them and may be a better alternative than HSIs for biota where little is known of their biota–physical habitat relationships (e.g. macrophytes and diatoms). The remaining model types will certainly increase in use as they become more developed, refined and independently tested in the research field, but at present they remain research or academic-based. However, modelling types such as ANNs and IBMs that take away the problems of in situ restrictions that many other model types have hold promise and should be researched further as they are universal and therefore widespread applicable, a major attribute for all water managers. In relation to hydraulic–habitat modelling, generalized habitat modelling approaches hold promise for reducing the fieldwork and experience needed in conventional hydraulic–habitat approaches.

Acknowledgements

The authors of this paper would like to thank Ian Jowett, Thom Hardy and Mike Dunbar for their comments and help with the paper. This study was partly funded by the project ‘Climate change impact on ecological conditions in streams’, contract no. 274-06-0474.