Deliverable 3.5 final - CSIC · Figure 38. A) Data-driven fuzzy sets obtained from the Shannon-Waver entropy-based optimization for the standardized density in Ephemeroptera, Trichoptera,

Assessing and predicting effects on water quantity and quality in Iberian rivers caused by global change (2009-2014). Consolider-Ingenio 2010 CSD2009-00065

WP3: MORPH

DELIVERABLE 3.5 HABITAT SUITABILITY MODELS FOR KEY-SPECIES OF FISH AND INVERTEBRATES,

BASED ON UNIVARIATE AND MULTIVARIATE ANALYSES AT DIFFERENT SCALES (E.G. ANNA

MODELS). DETERMINATION OF HYDROLOGICAL GUILDS OF RIPARIAN WOODY

SPECIES


I

WP3: MORPH

TECHNICAL DETAILS

Description: Report corresponding to the deliverable 3.5 of the Work Package 3

MORPH: Habitat suitability models for key-species of fish and invertebrates, based on

univariate and multivariate analyses at different scales (e.g. ANNA models).

Determination of hydrological guilds of riparian woody species (Consolider-Ingenio

2010 CSD2009-00065)

Elaboration: WP3 Members (led by Francisco Martínez Capel)

[email protected]

Contact: Arturo Elosegi, [email protected]

Delivery date: October 2013


II

WP3: MORPH

CONTENTS 1 INTRODUCTION 1

1.1 HABITAT SUITABILITY MODELLING FOR FISH 1 1.2 HABITAT SUITABILITY MODELLING FOR RIPARIAN VEGETATION 7

2 METHODS 11

2.1 HABITAT SUITABILITY MODELLING FOR FISH AND MACROINVERTEBRATES ‐ FIELD DATA COLLECTION 11 2.1.1 Cabriel River 11 2.1.2 Siurana River and Mijares River 15 2.1.3 Ésera River 17 2.1.4 Macroinvertebrates 19

2.2 HABITAT SUITABILITY MODELLING FOR RIPARIAN VEGETATION ‐ RIVERINE CHARACTERIZATION & DATA COLLECTION

21 2.2.1 Hydrological characterization 22 2.2.2 Riverine characterization 23 Field data collection 24

2.3 HABITAT SUITABILITY MODELLING FOR FISH AND MACROINVERTEBRATES 27 2.3.1 Univariate Habitat Suitability Curves (HSCs) 27 2.3.2 Data‐driven fuzzy modelling 31 2.3.3 Probabilistic neural network modelling 34

2.4 HABITAT SUITABILITY MODELLING FOR RIPARIAN VEGETATION 38 2.4.1 Laboratory analysis 38 2.4.2 Data analysis 39 2.4.3 Definition of vegetation indicators 39

3 RESULTS 44

3.1 HABITAT SUITABILITY MODELLING FOR FISH AND MACROINVERTEBRATES 44 3.1.1 Univariate Habitat Suitability Curves (HSCs) 44 3.1.2 Multivariate Habitat Suitability Models 48

3.2 HABITAT SUITABILITY MODELLING FOR RIPARIAN VEGETATION 77 3.2.1 Vegetation description 77 3.2.2 Positional patterns and riparian guilds 79 3.2.3 Dominance curves 80 3.2.4 Relationship with soil characteristics 82 3.2.5 Multivariate interpretation 85

4 DISCUSION 89

4.1 HABITAT SUITABILITY MODELLING FOR FISH AND MACROINVERTEBRATES 89 4.1.1 Univariate Habitat Suitability Curves (HSCs) 89 4.1.2 Multivariate Habitat Suitability Models 93

4.2 HABITAT SUITABILITY MODELLING RIPARIAN VEGETATION 105 4.2.1 Composition 105 4.2.2 Transverse variation across the floodplain: definition of guilds 106


III

WP3: MORPH

4.2.3 Importance and uncertainty of the variables in the definition of guilds 106 4.2.4 Management implications and further research 109

5 CONCLUSIONS 110

6 REFERENCES 114


IV

WP3: MORPH

Figure captions

Figure 1. General example of a univariate Habitat Suitability Curve.

Figure 2. Importance of different environmental variables across the transverse

gradient from the channel to the upland forest. Different water levels are indicated (low:

LW; mean: MW; high: HW). Figure modified from Dr. Peter J. Horchler (pers. com.).

Figure 3. Location of the Tagus Basin (TB) and Jucar River Basin District (JRBD) in the

Iberian Peninsula, and detail of sites where microhabitat surveys were carried out

(2005 – 2009).

Figure 4. Map of the sites where redfin barbel microhabitat surveys were carried out

during the summer 2012, in the Mijares River.

Figure 5. Frequency analysis of the number of individuals per size category at each fish

location.

Figure 6. Locations of the microhabitat survey in the tributaries of the Mijares, Palancia

and Turia Rivers.

Figure 7. Location of the study reach ‘Rabo del Batán’ within the Cabriel River Basin

and the Júcar River Basin District (Eastern Spain). The twenty cross-sections are

shown in red.

Figure 8. Daily river discharges (m3/s) in the Cabriel-Rabo del Batán reach. Period:

1949-2009.

Figure 9. Photographs illustrating the upstream (a) and downstream (b) sections of the

Cabriel-Rabo del Batán study reach.

Figure 10. Cross-sections and steel rods in the Cabriel-Rabo del Batán study reach.

Figure 11. Example of a steel rod marking the location of a cross-section (a) and geo-

referenced survey of vegetation units using a total station FOIF® (b).

Figure 13. A) Example of discretization for depth in 3 fuzzy sets for adult brown trout.

B) Expert-knowledge Output discretization. C) Data-driven Output discretization. Notice

no overlapping, but still providing a continuous output within the range from 0 to 1.

Figure 14. General architecture of a Probabilistic Neural Network (PNN) in a presence-

absence classification problem. x corresponds to the assessed pattern. XP correspond


V

WP3: MORPH

to the presence training patterns (n=98) whereas XA to the absence training patterns

(m=1457).

Figure 15. Effect of the selection of different values of a single smoothing parameter (σ)

in the habitat assessment of a pool at the Cabriel River (Z=elevation). The larger the σ,

the smoother the classification. Large values of the smoothing parameter sigma (0.25)

do not provide the extremes of the output range (0-1) whereas lower values (0.025)

provide sharper transitions achieving the extremes of the output range.

Figure 16. Cross-sectional view of a theoretical transect. The position of the points

recorded in the field is shown: beginning and end of the vertical projection (to estimate

cover) and predicted ground surface level of the main stem (to estimate distance and

elevation to/above thalweg).

Figure 17. Habitat Suitability Curves for adult brown trout. Upper sequence Category II

½ curves (Nuse = 98). Lower sequence Category III curves (Nuse = 98,Navailability = 2289).

Figure 18. Habitat Suitability Curves for juvenile brown trout. Upper sequence Category

II ½ curves (Nuse = 140). Lower sequence Category III curves (Nuse = 140,Navailability =

1978).

Figure 20. Habitat Suitability Curves for redfin barbel. Upper sequence Category II ½

curves (Nuse = 91). Lower sequence Category III curves (Nuse = 91, Navailability = 345).

Figure 21. Habitat Suitability Curves for adult, juvenile and fry brown trout. Adult and

juvenile based on Bovee (unpublished 1995) (Nadult = 74, Njuvenile = 153). Fry based on

data collected in the Jucar River Basin District (Nfry = 44).

Figure 22. Category II ½ and Category III Habitat Suitability Curves and their

corresponding fuzzy sets for adult brown trout. The last sequence corresponds to the

Data-driven fuzzy sets obtained from the Shannon-Waver entropy based optimization.


corresponding fuzzy sets for juvenile brown trout. The last sequence corresponds to

the Data-driven fuzzy sets obtained from the Shannon-Waver entropy based

optimization.


VI

WP3: MORPH


corresponding fuzzy sets for brown trout fry. The last sequence corresponds to the

Data-driven fuzzy sets obtained from the Shannon-Waver entropy based optimization.

Figure 25. Habitat assessment for adult brown trout carried out with the four generated

models. EK type A means Expert-knowledge based on Category II ½ curves, EK type

B means Expert-knowledge based on Category III curves, DD type A means

Unmodified Data-driven model and DD type B means Modified Data-driven model.

Figure 26. Habitat assessment for juvenile brown trout carried out with the four

generated models. EK type A means Expert-knowledge based on Category II ½

curves, EK type B means Expert-knowledge based on Category III curves, DD type A

means Unmodified Data-driven model and DD type B means Modified Data-driven

model.

Figure 27. Habitat assessment for brown trout fry carried out with the four generated

models. EK type A means Expert-knowledge based on Category II ½ curves, EK type

B means Expert-knowledge based on Category III curves, DD type A means

Unmodified Data-driven model and DD type B means Modified Data-driven model.

Figure 28 Frequency analysis of the habitat assessment carried out with the four

models and the three size classes (black bars), over the entire simulation reach

(Availability data). Frequency analysis of the habitat assessment carried out with the

four generated models and the three size classes (grey bars) over the corresponding

size class locations (Use data). EK mean Expert-knowledge and DD Data-driven.

Figure 29. Density map of brown trout in the Cabriel River evaluation site.

Figure 32. LEFT: comparison of the habitat assessment of the flow surveyed (Q=0.89

m³/s) during the biological validation using the developed PNN. Red areas mean

unsuitable locations and dark green mean areas with perfect suitability for adult trout.

Dots are adult trout positions at the moment of the validation survey. RIGHT: frequency

histogram of the assessment of PNN. Availability (assessment of the entire simulated

reach) is represented by black bars and Use (assessment of fish locations) by grey

bars. Notice that the assessment carried out through the PNN did not provide the

maximum suitability (range 0.8-1).


VII

WP3: MORPH

Figure 34. Frequency distribution of the original Availability database and of the Sub-

sample database. Notice that distributions are quite similar, but that frequencies are

lower in the sub-sample case.


corresponding fuzzy sets for Barbus haasi. Data-driven fuzzy sets obtained from the

Shannon-Waver entropy based optimization with similar amount than the HSC-based

and Data-driven fuzzy sets unconstrained.

Figure 36. Category II ½ Habitat Suitability Curves and their corresponding fuzzy sets

for adult, juvenile and fry brown trout based on Bovee (1995 unpublished).

Figure 37. Data-driven fuzzy sets obtained from the Shannon-Waver entropy-based

optimization. Substrate was discretized in two Fuzzy Sets, but an extra set was

included to cover the bedrock substrate.

Figure 38. A) Data-driven fuzzy sets obtained from the Shannon-Waver entropy-based

optimization for the standardized density in Ephemeroptera, Trichoptera, and EPT. The

Data-driven approach provided exactly the same discretization on the three cases, but

differed on the represented density values B) Presence/Absence output fuzzy sets for

Plecoptera.

Figure 39. Descriptors of the riparian vegetation species sampled by transects at the

Cabriel-Rabo del Batán study reach. LEFT: abundance (percentage of stems of each

species respect to the total number of plants sampled), cover (percentage of cover of

each species respect to the total cover value in the site). RIGHT: density (meter of

cover of each species per meter of transect, in bars; number of stems of each species

per meter of transect, in solid line). Species appear in ascendant ranking based on the

Huber estimator of elevation above thalweg. Species abbreviations: SP, Salix purpurea

L.; ST, Salix triandra L.; CS, Cornus sanguinea L.; SL, Salix alba L.; FA, Fraxinus

angustifolia Vahl; PN, Populus nigra L.; SE, Salix eleagnos Scop.; CR, Crataegus

monogyna Jacq.; PA, Populus alba L.; PC, Pinus spp.

Figure 40. Boxplots of the distribution (left plot, distance to thalweg; right plot, elevation

above thalweg) of the woody riparian species sampled at the Cabriel-Rabo del Batán

study reach. Species are in ascendant order according to their median values. Red

dots represent the Huber estimator.


VIII

WP3: MORPH

Figure 41. Dominance curves of the riparian species respect to the distance to thalweg

at the Cabriel-Rabo del Batán study reach. Left, absolute frequency. Right, normalized

giving a value of 1 to the peak frequency of each species.

Figure 42. Dominance curves of the riparian species respect to the elevation above

thalweg at the Cabriel-Rabo del Batán study reach. Left, absolute frequency. Right,

normalized giving a value of 1 to the peak frequency of each species.

Figure 43. Grain size distribution of the soil samples of the Cabriel-Rabo del Batán

study reach.

Figure 44. Percentages of organic matter and moisture content in the twelve soil

samples analyzed in the Cabriel-Rabo del Batán study reach.

Figure 45. Barplots with standard errors for the percentage of gravel, sand, silt and clay

for each species at the Cabriel-Rabo del Batán study reach. Species appear in

ascendant order according to the ranking defined by the Huber estimator in the

elevation gradient above thalweg.

Figure 46. Barplots with standard error for the percentage of organic matter and

moisture content for each species at the Cabriel-Rabo del Batán study reach. Species

appear in ascendant order according to the ranking defined by the Huber estimator in

the elevation gradient above thalweg.

Figure 47. PCA diagram for the Cabriel-Rabo del Batán study reach. Riparian woody

species (dots in black colour) are abbreviated as: ST, Salix triandra; CS, Cornus

sanguinea; SL, Salix alba; SP, Salix purpurea; FA, Fraxinus angustifolia; PN, Populus

nigra; PA, Populus alba; CR, Crataegus monogyna; SE, Salix eleagnos; PC, Pinus

spp.

Table captions Table 1. Summary of the sample sizes for the three life stages of microhabitat data.

JRBD means Jucar River Basin District and TB means Tagus Basin.

Table 2. Summary of the sample sizes for spawning brow trout. JRBD means Jucar

River Basin District. In total, 356 data of habitat use, 712 for habitat availability.

Table 5. Main characteristics of the Pajaroncillo flow gauging station at the Cabriel

River.


IX

WP3: MORPH

Table 6. Characteristics of the flow time series (m3/s) of the gauging station at the

Cabriel River.

Table 9. Summary of the Fuzzy Rules for the generated models. Asterisk (*) means

uncovered rule and the proper output was based on the corresponding rule in the

Expert-knowledge approach, if it exists of in the Data-driven model with different

substrate. Suitability in 5 categories corresponds to Very Low (VL), Low (L), Medium

(M), High (H) and Very High (VH).

Table 10. Assessment of the training database through the models. The sensitivity (sn),

the specificity (sp) and the true skill statistic (TSS) were calculated.

Table 11. Summary of the Fuzzy Rules for the generated fuzzy models including a

model for adult, juvenile and fry brown trout based on Bovee (1995 unpublished).

Suitability in 5 categories corresponds to Very Low (VL), Low (L), Medium (M), High (H)

and Very High (VH).

Table 12 Summary of the Fuzzy Rules for the macroinvertebrates models. Asterisk (*)

means uncovered rule (no data for training); in such cases, the outputs were

determined through authors’ consensus, that is, expert-knowledge approach. Suitability

in 5 categories corresponds to Very Low (VL), Low (L), Medium (M), High (H) and Very

High (VH).

Table 13. Summary of the plant species surveyed at the Cabriel-Rabo del Batán study

reach. The scientific and common names of each species are indicated, as well as the

code assigned to each species. In the last column appears the number of specimens

surveyed.


X

WP3: MORPH

EXECUTIVE SUMMARY The present document describes fish and macroinvertebrate habitat suitability models applicable in the Cabriel, Siurana and Ésera River Basins. Additionally, the concept of habitat suitability curves was applied further, describing the distribution pattern of the most important riparian woody species was studied. The final goal of these curves and models is the implementation of habitat suitability maps and physical habitat simulation considering more components of the river ecosystem; namely, macroinvertebrates, fish and riparian vegetation.

HABITAT SUITABILITY FOR FISH

Freshwater fish and macroinvertebrates have been used as indicators of the ecological status of river systems. Habitat suitability models allow predicting changes on fish and macroinvertebrate populations. The univariate approach, widely used to model habitat requirement, presents several deficiencies. The present document summarizes univariate habitat suitability curves developed for brown trout (Salmo trutta L.) and redfin barbel (Barbus haasi Mertens). These species were selected because they are key species of the fish community in the rivers under study, and they allow the application of habitat suitability models to assess the effects of flow regulation and water scarcity on the aquatic ecosystem. Additionally, the fuzzy logic approach was used to develop multivariate habitat suitability models, for brown trout spawning and for macroinvertebrates. The habitat suitability models were validated in independent rivers demonstrating the capability to predict fish distribution. The macroinvertebrate models did not shown robust results thus recommending site-specific studies. Finally, the capability of the probabilistic neural networks in predicting adult brown trout distribution was tested with satisfactory results.

HABITAT SUITABILITY FOR RIPARIAN VEGETATION

Riparian vegetation is an essential component of fluvial ecosystems, as it is recognised in the legislation at national and European levels. However, basic information is still needed to integrate this ecosystem component into the water management. To conduct a pilot study in this direction, one free-flowing Mediterranean reach with high riparian quality was selected in the Cabriel River (Júcar River Basin District). The positional patterns of key riparian woody species were analyzed to define guilds of species that respond similarly to physical habitat conditions, in particular to river morphology (distance and elevation to/above thalweg) and soil characteristics (texture, organic matter and moisture). A combination of robust non-parametric statistics and multivariate statistics were performed to analyze the joint variation of all variables and thus obtain groups of species with a similar response (guilds). The results are relevant for the potential definition of riparian guilds of hydrological response, as well as to integrate riparian vegetation into water management decisions in Mediterranean rivers.


1

WP3: MORPH

HABITAT SUITABILITY MODELS FOR KEY-SPECIES OF FISH AND INVERTEBRATES, BASED ON UNIVARIATE AND MULTIVARIATE ANALYSES AT DIFFERENT SCALES

1 INTRODUCTION

1.1 Habitat suitability modelling for fish

For several decades freshwater fish and macroinvertebrates have been considered as

good indicators of water quality in river systems (Karr, 1981; Bischoff and Freyhof,

1999; Angermeier and Davideanu, 2004). The selection of some fish species as

indicators was based on their sensitivity to environmental alterations and their position

on the trophic chain; some macroinvertebrates are very sensitive to pollution and play a

critical role in the maintenance of the stream functional integrity (Wallace and Webster,

1996). Fish requirements have been determined in many studies, using a range of

methods from univariate to multivariate techniques. Regarding fish, Waters (1976)

proposed a habitat suitability index ranging from 0 to 1, under given hydraulic

conditions. Then the Habitat Suitability Criteria, in the form of Univariate Habitat

Suitability Curves (HSCs), have been developed and applied most commonly for

instream flow assessment (Vismara et al., 2001). The criteria are usually based on

unimodal curves that provide the suitability of variables such as velocity, depth,

substrate, cover and temperature for the target organism, fish or macroinvertebrate

genera, family or order (Figure 1).

Figure 1. General example of a univariate Habitat Suitability Curve.


2

WP3: MORPH

The HSCs were categorized according to the methodology used to develop the criteria

(Bovee, 1986): Category I includes curves generated from literature and expert

consensus. Category II are curves based on frequency analysis of hydraulics over the

fish or macroinvertebrate locations, but without reference to their invertebrate electivity.

Finally, Category III consists on preference curves derived also from observational data

on habitat use (i.e. hydraulics in the locations where fish or macroinvertebrates were

observed), but weighted by habitat availability (i.e. hydraulics over the surrounding

unoccupied locations). This weighting process has been usually carried out through the

forage ratio (Voos, 1981). However, several authors have considered an extra category

if the sampling approach followed the equal effort protocol (Johnson, 1980), the so-

called Category II ½ (Bovee, 1986).

The Habitat Suitability Index (HSI), ranging from zero to one, is the most widely applied

method to combine the results from several HSC in a single composite suitability index

(Vadas and Orth, 2001). Various approaches have been used to combine suitability

indices in a HSI, including the lowest (Korman, 1994), the product (Bovee, 1986), the

arithmetic mean (Terrell, 1984) and the geometric mean (Terrell, 1984). The lowest is a

'controlling method' that assumes the most limiting factor to determine the upper limit of

habitat suitability; therefore, the different suitability of the variables cannot compensate

each other. The product method is also a 'controlling method', whereas arithmetic and

geometric means are partially 'compensatory methods' (U.S. Fish and Wildlife Service

1981). The product approach assumes that unsuitable values of a variable cannot be

compensated by good conditions of another (Bovee, 1986), but does not determine the

upper limit. In contrast, the arithmetic mean assumes that good habitat conditions

based on one variable can compensate for poor conditions of others (Terrell, 1984).

Finally, the geometric mean assumes that each environmental variable is equally

important (Benaka, 1999; Rubec, 1999).

Despite the existence of these possibilities, considering each hydraulic variable

independently may be questionable, as it could induce a bias as a result of overlooking

possible interactions between variables (Orth and Maughan, 1982; Lambert and

Hanson, 1989). The multivariate approach addresses this limitation successfully, thus it

has increased in popularity among researchers (De Pauw et al., 2006). For example,

several multivariate approaches have been successfully applied: fuzzy logic (Van


3

WP3: MORPH

Broekhoven et al., 2006; Mouton et al., 2008), Logistic regression (Guay et al., 2000;

Turgeon and Rodríguez, 2005), Generalized Additive Models (GAMs) (Jowett and

Davey, 2007) or artificial neural networks, specifically the multilayer perceptron (MLP),

(Lek et al., 1996; Reyjol et al., 2001). Among these multivariate approaches, the fuzzy

logic, firstly introduced by Zadeh (1965), has demonstrated a great versatility to deal

with several modelling situations (e.g. at the microscale in fish and macroinvertebrate

habitat suitability modelling: Van Broekhoven et al., 2006; Mouton et al., 2008; García

et al., 2011). Important advantages of the fuzzy logic are its transparency, which may

stimulate communication of model results to stakeholders (Van Broekhoven et al.,

2006) and its ability to incorporate the ecological gradient theory (Mouton, 2008).

There are several types of fuzzy inference systems, although the most widely used in

ecological modelling are the Mamdani fuzzy systems (Mamdani, 1974). The main

advantages of this type of systems are its widespread acceptance, intuitive character

and capability to formalise inputs in the form of expert knowledge (Gegov, 2007).

These fuzzy inference systems consist of three parts: (i) fuzzy input and output

variables in the form of Fuzzy Sets (FS) whose shapes are defined by their fuzzy

Membership Functions (MF); (ii) fuzzy rules and (iii) fuzzy inference methods

(Kasabov, 1998). To implement the first part, fuzzy systems categorize the input and

the output variable in linguistic terms such as: Low, Medium, High etc. defined by fuzzy

sets (Zadeh, 1965). The fuzzy sets are described by their membership function, which

indicates the membership degree, ranging from zero to one, to each fuzzy set of a

given variable value. Since membership functions have overlapping boundaries, called

fuzzy borders, instead of crispy borders, a given value may belong, with different

proportions, to two adjacent fuzzy sets.

For instance, an area could present low velocity but partially medium, whereas

methods of discretization with crispy borders would classify it as low or medium

velocity, but not both of them. The second part of the fuzzy inference system is

implemented defining the relationship of the categories by defining rules of association,

the Fuzzy Rules (FR). These rules are a collection of linguistic statements, where the

'if' part is the antecedent (input) and the 'then' part is the consequent (output). For

instance: If velocity is low and depth is medium and substrate is high then the suitability

is low. The fuzzy inference systems allow more than one rule to be partially applied to a


4

WP3: MORPH

given area under assessment. Therefore, the third part consists in the defuzzification of

output distribution, giving a single suitability index ranging from zero to one in

accordance with Waters' proposal (1976). There are several methods, the most

commonly applied is the Centre of Gravity (Ahmadi-Nedushan et al., 2008),

successfully applied in previous studies (Jorde et al., 2001; Muñoz-Mas et al., 2012).

To date, the development of fuzzy models applied to the ecological requirements of

freshwater organisms present two main approaches: the Expert-knowledge and the

Data-driven. Both approaches have demonstrated useful in fish habitat modelling. The

Expert-knowledge approach is based on the literature and the consensus of scientists

about the discretization of input variables (the fuzzy sets) and the fuzzy rules (i.e.

relations of the input fuzzy sets with the output). Data-driven approach optimizes the

discretization of the input variables and the consequent for each fuzzy rule. Some

authors have demonstrated that expert judgment is convergent and that the Expert-

knowledge fuzzy inference systems do not differ substantially depending on the

consulted expert (Ahmadi-Nedushan et al., 2008). However, in some cases the

consultation of an expert panel is unfeasible, forcing fuzzy inference systems to base

on literature.

Another multivariate approach recently tested are the Probabilistic Neural Networks

(PNNs) firstly introduced by Specht (1990). PNNs have been successfully applied in

pattern classification in some areas related to fish (i.e classification of sonar signals)

(Moore et al., 2003) and in the assessment of the suitability for bacteria growth in

culture (Hajmeer and Basheer, 2002). However, to our knowledge Probabilistic neural

networks had never been applied before to model fish microhabitat. The classification

strategy of the PNN is called a 'Bayes strategy' and is based on the minimization of the

risk of misclassification (Mood et al., 1974). For instance, in a given problem where two

categories were present (i.e. 'presence' or 'absence') the Bayes theorem considers a

single sample X=[x1,x2,x3 ...xp] and d(X)= hi·li·fi, the theoretical function that defines its

membership to a given category. This sample will be classified, for instance, in the

category 'presence' if the following inequality is fulfilled: hP·lP·fP(X) > hA·lA·fA(X) where hi

is the a priori probability of occurrence, Ii is the cost associated with misclassification

and fi is the probability density function of the corresponding category, and their

aggregation define the membership function.


5

WP3: MORPH

Bayes’ theorem favours a class that has high density in the vicinity of the unknown

sample (fi(X)) or if the cost of misclassification (Ii) or prior probability (hi) are high

(Hajmeer and Basheer, 2002). The main problem at this point is that the corresponding

probability density function is ignored and the samples in the training database provide

the only clue to infer the unknown underlying probability density function (Specht,

1990). The development of the methodology to calculate the probability density

function was proposed by Parzen (1962) for the univariate case and Cacoullos (1966)

generalized the formulation for the multivariate case. The resulting multivariate

probability density estimator combines all the explanatory variables and all the samples

in the database. Therefore the combination or samples' connection becomes a network

and this methodology is considered as an artificial neural network.

The decreasing trends in many native fish populations, specifically the brown trout

(Almodóvar et al., 2012), or the temporary absence of the species (Lütolf et al., 2006)

may lead to low prevalence databases. Prevalence can have a strong effect on model

performance (Manel, 2001). Among other methodologies (Mouton et al., 2009b) PNN

are able to cope with low prevalence databases, thus making them potentially ideal to

model fish microhabitat suitability. PNN basically compares the assessed conditions

with the measured conditions included in the training database, and determines the

probability of membership of the assessed conditions to each of the categories present

in that training database.

Most of the aforementioned studies and methods have been focused on fish, but

macroinvertebrates have been also the target of several studies (Mérigoux and

Dolédec, 2004; Dolédec et al., 2007; Pastuchová et al., 2008; Mérigoux et al., 2009). At

the univariate level there are several examples in the development of studies for

macroinvertebrate species, thus relating variations on the hydraulic conditions and

macroinvertebrate density (Gore and Judy Jr, 1981; Orth and Maughan, 1983; Morin et

al., 1986; Jowett et al., 1991; Collier, 1993; Mérigoux and Dolédec, 2004; Dolédec et

al., 2007; Mérigoux et al., 2009), and even developing the old fashion HSCs (Gore et

al., 2001), although the macroinvertebrates have been the targets for more complex

techniques such as the fuzzy logic approach (Van Broekhoven et al., 2007; Mouton et

al., 2009a; Mouton et al., 2009c). The interest for invertebrates derives from their role

as processors of organic matter, and because they are a critical food for fish (Gore et


6

WP3: MORPH

al., 2001). Besides, macroinvertebrates play a crucial role in the presence of keystone

organisms such as large brown trout (Jowett, 1992; Jowett et al., 1996; Morante et al.,

2012), which are correlated with drift abundance (Shannon et al., 1996), (Mas-Martí et

al., 2010). However, natural systems are complex, what results in frequent

discrepancies (Gore et al., 2001), which should encourage the fulfilment of new

studies. Stream hydraulics is also a major determinant of macroinvertebrate

communities (Statzner et al., 1988; Davis and Barmuta, 1989; Carling, 1993).

Hydraulic characteristics are thought to structure assemblages by influencing the

metabolism, feeding and behaviour of macroinvertebrates (Statzner et al., 1988).

Therefore, small differences in combinations of velocity, depth and substrate have an

important role in the spatial distribution of macroinvertebrates (Brooks et al., 2005).

Macroinvertebrates are more sensitive to habitat changes than fish because of their

lower swimming capability, which prevents them from returning to their original

locations (Gore et al., 2001). Therefore, instream flow assessment should also

consider habitat for benthic invertebrates, particularly where food availability may limit

fish numbers and/or growth (Jowett, 2003). Most recent habitat suitability models for

macroinvertebrates used multivariate approaches (Jowett and Davey, 2007). Some

authors defend using variables separately (Jowett et al., 1991; Jowett, 2003) because,

apparently, macroinvertebrates and fish (Scruton et al., 2003) vary their preferences

depending on the available hydraulic conditions (Mérigoux and Dolédec, 2004). Hence,

in several cases site-specific and season-specific models outperformed those more

generalist (Bovee, 1986; Gore, 1987) because ecological interactions (i.e. changes on

the present fish species or on incoming materials) could modify the fundamental niche

(Gore et al., 2001). As far as we know, in the Iberian context no multivariate

microhabitat suitability model has been developed for macroinvertebrates.

This document includes the steps and results of the models developed to be applied in

the study sites of the SCARCE WP3. Specifically, at the univariate approach the brown

trout Habitat Suitability Curves (HSCs) to be applied in the Cabriel River (Jucar River

Basin). The HSCs for Iberian redfin barbel were developed in the Mijares River to be

applied in the Siurana River (Ebro River Basin) and the brown trout HSCs to be applied

in the Ésera River (Ebro River Basin). Additionally, at the multivariate level the present

document includes the Data-driven and Expert-knowledge fuzzy models for brown trout


7

WP3: MORPH

to be applied in the Cabriel River (Jucar River Basin), the Probabilistic Neural Network

study to be applied in the Cabriel River (Jucar River Basin), the Data-driven fuzzy

model to be applied in the Siurana River (Ebro River Basin), and the Expert-knowledge

fuzzy model to be applied in the Ésera River (Ebro River Basin). Additionally Data-

driven fuzzy models were developed for some of the most important macroinvertebrate

orders, specifically Ephemeroptera, Plecoptera and Trichoptera (EPT).

1.2 Habitat suitability modelling for Riparian vegetation

The knowledge of the riparian forests, regarding their distribution and composition has

been largely expanded since the 1980’s (Hupp and Osterkamp, 1985; Nilsson, 1986;

Salo et al., 1986; Slater et al., 1987). Also, the relations between riparian species and

stream hydrology have been developed (Stromberg, 1993; Mahoney and Rood, 1998;

Shafroth et al., 1998). Among such studies, some used variables related to river

morphology to explain species distribution (Stromberg and Patten, 1996; Nakamura et

al., 1997; Van Coller et al., 1997), others developed dynamic models to simulate

growth and processes of succession/retrogression in the riparian ecosystem (Inamdar

et al., 1999; Glenz, 2005; Egger et al., 2007; Benjankar et al., 2011).

Fluvial disturbance and water stress vary along the lateral and longitudinal gradient of

the river (Lite et al., 2005), and therefore, are considered as two of the primary factors

that influence spatial vegetation patterns (Malanson, 1993). Furthermore, the

vegetation patterns are time-dependent and change with temporal variation in the

primary factors (Nilsson et al., 1994). Looking in detail laterally, i.e., along the

transverse profile of the riparian zone, the intensity and frequency of flood disturbance

typically diminish with increasing distance from (and above) the active channel. Along

this same gradient, water stress may increase, due to increases in groundwater depth.

In the context of semi-arid regions, and especially in Mediterranean rivers, due to water

scarcity, riparian vegetation patterns are highly influenced by factors such as flow

regime, soil moisture and groundwater availability (Tabacchi et al., 1996). In addition,

flood disturbance and water availability are interrelated, and both vary greatly over

time. For instance, wet years produce conditions of both low water stress and high

disturbance, and dry years result in high stress and low disturbance (Lite et al., 2005).


8

WP3: MORPH

Besides, the intensity of the disturbance regime can influence plant species diversity at

local scales (Ali et al., 2000).

In the transverse profile, other important factors are the sedimentation rate and grain-

size, which depend largely on the distance from the active channel and on ground

elevation. Soil water is very important to the entire soil system as well, not only

because it is necessary for plant growth, but because the nutrients required for plant

growth are also present in the soil solution. Most of the important soil reactions

(weathering, cation exchange, organic matter decomposition, fertilization) take place in

the context of the soil solution. Thus, it is evident that not only the texture, but also the

moisture status of a soil is a key property.

Figure 2. Importance of different environmental variables across the transverse gradient from

the channel to the upland forest. Different water levels are indicated (low: LW; mean: MW; high: HW). Figure modified from Dr. Peter J. Horchler (pers. com.).

As can be seen in Figure 2, different environmental variables become important as we

move farther in the transverse gradient. Firstly, water level and flow velocity are the key

factors for plant establishment and development in the proximities of the aquatic zone

(base flow or low water level conditions). When we pass to the bank zone, besides the

previous factors, the soil texture plays its role. As the influence of the water surface

reduces and distance and elevation increase, soil moisture becomes a limiting factor

for plant development. Farther, in the floodplain zone, the vegetation succession via


9

WP3: MORPH

autogenic processes (internal mechanisms) is able to induce changes in the vegetation

patterns with processes such as competition. Finally, in the transitional zone or ecotone

between floodplain and upland, where only high flows of low recurrence take place, the

land uses, increased soil depth profiles and nutrient pools become more important,

promoting a more robust and well-developed vegetation.

Another important factor to take into consideration in this hierarchy is the status of plant

development. Herbaceous species may be more sensitive than trees and shrubs to

disturbance, drought stress or geomorphic heterogeneity (Lyon and Sagers, 1998).

According to Decocq (2002), tree and shrub layers may respond to soil moisture and

nutrient gradients, whereas the herbaceous communities may be most related to

gradients of soil nutrients and light availability. In the same way, plants with different

life-history strategies may respond differently to disturbances and stressors

(Karrenberg et al., 2002).

Most studies relating riparian communities with abiotic factors were focused in a few

species, and their conclusions have not been easily transferable to other rivers in other

geographic regions. However, species tend to associate according to similarity in

response to environmental conditions (Simberloff and Dayan, 1991; Austen et al.,

1994). These associations of species are known as functional groups or guilds.

According to Bejarano et al. (2012), the guild approach provides information about the

general trends in plant populations and assemblage structures and allows

generalization and comparisons among different fluvial systems. The guild approach,

for example, allows the analysis of the vegetation response to changing river hydrology

and geomorphology (Bejarano et al., 2012). In fact, the identification of riparian

vegetation-flow response guilds has been proposed by Merritt et al. (Merritt et al.,

2010) as a tool for determining environmental flows.

In order to apply the guild approach, it is necessary to distinguish if different species

are affected in the same way by certain environmental variables, and which species

show significant differences. Another question arising is the applicability of the

empirical results of a given species (for example, in tolerance to inundation) in rivers

with diverse physical conditions or the comparison of the performance of some species


10

WP3: MORPH

in rivers with different conditions. This is essential for the validation of models in other

rivers than the ones where the models were calibrated.

In consequence with these ideas, a pilot study was developed in a free-flowing

segment of a Mediterranean river within the Júcar River Basin District (Eastern Spain).

The general objective of this study was to describe the riparian vegetation community

using different indicators and to analyse the preferences of their main woody riparian

species in terms of positional patterns (through two key morphological variables:

distance and elevation to/above thalweg) and soil characteristics along the transverse

gradient. More specifically, this study was developed in order to address the following

questions:

• How do riparian woody species vary laterally across the floodplain in rivers with

Mediterranean influence, in relation to distance and elevation to/above thalweg?

and in relation to the soil properties (texture, organic matter and moisture

content)?

• Is it possible to define guilds of species with similar response to the physical

habitat conditions in a multivariate way?


11

WP3: MORPH

2 METHODS

2.1 Habitat suitability modelling for fish and macroinvertebrates - Field data collection

2.1.1 Cabriel River

The Mediterranean brown trout (Salmo trutta fario L.) was the target species of the

microhabitat study. The surveyed individuals of this species were divided in three size

stages; fry (< 10 cm), juvenile (10-20 cm) and adult (>20 cm). The surveys were carried

out in four rivers comprised in the Jucar River Basin District (JRBD); Senia, Turia,

Jucar and its principal tributary, the Cabriel River (Figure 3). In order to get a larger

database for the models, other two rivers were included in the Tagus River Basin (TB),

Guadiela and Cuervo. The surveys were carried out at low flows during late spring,

summer and early autumn in the period 2005-2009. The data to develop the spawning

models were collected in several surveys in rivers from the Jucar River Basin District

during the spawning period (late autumn and winter). The rivers included were

Guadalaviar, Valbona, Cabriel, Jucar, Villahermosa and Ebron, draining from the

Montes Universales (Figure 3); the sites had order one or two (Strahler, 1952).

Figure 3. Location of the Tagus Basin (TB) and Jucar River Basin District (JRBD) in the Iberian

Peninsula, and detail of sites where microhabitat surveys were carried out (2005 – 2009).


12

WP3: MORPH

The microhabitat study for fry, juvenile and adult was done in complete and connected

HydroMorphological Units (hereafter HMUs) classified as: pool, glide, riffle, and rapid

(Alcaraz-Hernández et al., 2011). The equal effort approach was applied (Johnson,

1980), with the selection of equal areas of slow and fast water HMUs, grouping pools

with glides (slow) and riffles with rapids (fast). Underwater observation by snorkelling

was the selected methodology to carry out the microhabitat surveys. The variables

defining microhabitat were depth, mean water columns velocity and substrate type, as

defined below. Each HMU was surveyed during daylight, with minimum disturbance to

the fish according to standard procedures (Heggenes, 1991). This technique allows the

observation of fish behaviour and position in the water column, even in relatively

adverse surveying conditions (Martínez-Capel and García de Jalón, 1999; Martínez-

Capel, 2008). The direct underwater observation has been demonstrated as more

reliable than electrofishing for fish location (Bovee and Cochnauer., 1977; Bovee,

1986; Gatz Jr et al., 1987; Heggenes, 1991). Direct underwater observation does not

frighten trout and does not produce galvanotaxis displacement.

The microhabitat conditions over the entire HMUs were measured in cross-sections

with a minimum amount of 300 points of unoccupied locations per survey, hereafter

Availability records, to ensure the applicability of the transferability tests (Thomas and

Bovee, 1993). This methodology produced a variable density from 1.23-7.96 m2 per

record. Table 1 shows a summary of the sample sizes. All the measurements in the

Availability survey (i.e. locations where no fish were observed) and in the Use survey

(i.e. locations where fish were observed) were taken with the same methods. Single

fish observations were considered and Presence/Absence modelling was applied.

Three variables were measured: flow velocity, water depth and substrate type, usually

considered the most relevant variables for fish species distribution at this scale

(Waters, 1976; Bovee, 1986; Heggenes, 1990; Gibson, 1993; Bovee, 1998). Velocity

was measured with an electromagnetic current metre (Valeport®) and depth was

measured with a wading rod at the nearest cm. The percentages of each substrate

class were visually estimated within 15 cm around the sampling point or fish location

(Bovee and Zuboy, 1988).


13

WP3: MORPH

Table 1. Summary of the sample sizes for the three life stages of microhabitat data. JRBD

means Jucar River Basin District and TB means Tagus Basin.

Fry (0 - 10 cm) Juvenile (10- 20 cm) Adult (>20 cm) River Use Availability Use Availability Use Availability

Jucar (JRBD) 19 735 38 735 7 735 Guadiela (TB) 51 455 Turia (JRBD) 25 379

Cabriel (JRBD) 68 532 Senia (JRBD) 34 711 11 714 Cuervo (TB) 29 385

Total 44 1114 140 1978 98 2289

The classification was simplified from the American Geophysical Union size scale in:

bedrock, boulders (>256 mm), cobbles (64–256 mm), gravel (8–64 mm), fine gravel (2–

8 mm), sand (62 mm–2 mm), silt (< 62 mm) and vegetated soil (i.e. substrate covered

by macrophytes), similarly to previous works made by snorkelling in Iberian rivers

(Martínez-Capel and García de Jalón, 1999; Martínez-Capel et al., 2009b). To agree

with previous studies substrate composition was converted into a single Substrate

index (S) by summing weighed percentages of each substrate type. The weights used

were: S = 0.08 x bedrock + 0.07 x boulder + 0.06 x cobble + 0.05 x gravel + 0.04 x fine

gravel + 0.03 x sand (Mouton et al., 2011).

Regarding the fish spawning, the surveyor walked along the river and the hydraulic conditions

were measured on the redds and in the surrounding area. On each redd (Use data) the

microhabitat variables were recorded in two points, the redd head (i.e. upstream conditions) and

the redd centre; four points in the surrounding area (absence or Available data) were also

surveyed; two locations at both sides, 1.5 m away from the redd, and other two upstream and

downstream, also 1.5 m away. The surveys reached close to 16 Km along the rivers, 356 Use

records and 712 Availability records were collected (

Table 2).


14

WP3: MORPH

Table 2. Summary of the sample sizes for spawning brow trout. JRBD means Jucar River Basin

District. In total, 356 data of habitat use, 712 for habitat availability.

B. trout spawning River Use Availability

Cabriel (JRBD) 54 108 Ebrón (JRBD) 58 116

Villahermosa (JRBD) 60 120 Jucar (JRBD) 54 108

Guadalaviar (JRBD) 60 120 Valbona (JRBD) 70 140

2.1.1.1 Hydraulic modelling and validation

A 2D hydraulic simulation with River-2D© (University of Alberta 2002) was done in a

300 m long reach of the Cabriel River, 9 km downstream from the locations of the

microhabitat survey for brown trout, and close to the spawning survey locations [see

more details in Muñoz-Mas et al., 2012]. The topographic data of the river channel and

banks were collected using a Leyca© total station and the substrate composition was

visually estimated. Eleven cross-sections from three different flow rates: 0.54, 1.04 and

2.75 m3/s were used to calibrate the model in terms of water depth and velocity pattern,

accordingly to previous studies (Jowett and Duncan, 2011).

The target species was the brown trout. The survey was conducted in a single week in

the early summer 2012, with a steady flow rate of 0.89 m3/s corresponding to the Q85.

Unlike previous biological surveys, the surveyor did not snorkel the entire HMUs.

Instead, the survey was done covering the whole area included in the hydraulic model,

which presents a stable trout population (Martínez-Capel et al., 2009a). The survey

was done in similar standards (Heggenes, 1991) as the microhabitat surveys presented

in the previous section, but instead of collecting data on velocity, depth and substrate,

the coordinates (X,Y,Z) of each observed trout were measured with a FOIF© Total

Station. This information was also used to check the model reliability in terms of

morphologic changes on the river bed, which showed no lateral displacement, and

minor adjustments of elevation, in the range of 0.04 ± 0.12 m. This was considered


15

WP3: MORPH

acceptable, and thus the entire model suitable to be used in the present study. The

information to validate the models in terms of velocity, depth and substrate was,

therefore, based on the hydraulic simulation. The sample sizes for adult, juvenile and

fry life stages were 31, 30 and 79 respectively. That information was also used to

calculate the fish density using the tool kernel density in ESRI® ArcMapTM 9.3, with a

radius equal to the mean distance to the nearest trout of the corresponding size. This

density information was standardized between 0 and 1 and discretized in five intervals

(Very Low density, Low density, Medium density, High density and Very high density).

The average suitability corresponding to these five categories was used to check

positive correlations suitability and trout density. Finally, this information was used to

modify the rules that yielded the maximum error in order to adjust the Data-driven fuzzy

models to the observed trout preferences.

2.1.2 Siurana River and Mijares River

The target species in the Siurana River was the Iberian redfin barbel, (Barbus haasi

Mertens, 1925). The species is present in the upper part of most rivers in the North-

East of the Iberian Peninsula (Perea et al., 2011). However, a preliminary study in the

Siurana River showed that summer low flows were too low because depth in the

remaining pools was very small, limiting fish electivity in a serious manner. Thus, we

performed the microhabitat survey in the Mijares River (Valencian Country) during early

summer 2012. The microhabitat study locations covered alternatively most of the

quasi-unregulated length of the upper Mijares River (14 km) (Figure 4).


16

WP3: MORPH

Figure 4. Map of the sites where redfin barbel microhabitat surveys were carried out during the summer 2012, in the Mijares River.

The microhabitat study was done in complete HydroMorphological Units. In the river

reach, we detected these HMU: run, riffle, deep run, riffle, rapid and pool. The equal

effort approach was applied (Johnson, 1980), with the selection of equal areas of slow

and fast water HMUs. Underwater observation by snorkelling was the selected

methodology. Each HMU was surveyed during daylight, with minimum disturbance to

the fish according to standard procedures (Heggenes, 1991). The specimens were

visually divided in three categories: fry (0-5 cm), juvenile (5-10 cm), and adults (>10

cm). This division agreed electrofishing data provided by (Alcaráz-Hernández, 2011).

Fish were observed in 101 locations (Use database) whereas the habitat in the

surrounding area (Availability database) was measured in 341 locations randomly

placed along the HMU, but far enough from fish location to avoid any overlapping

between the Use and the Availability databases. The survey provided 9 data for fry, 68

for juvenile and 24 for adult with variable amount of fish at each location, but with a

maximum of 9 individuals in a single location (Figure 5). Therefore, to carry out further

analysis the juvenile and adult databases were joined and the fry data discarded.

Figure 5. Frequency analysis of the number of individuals per size category at each fish

location.

The frequency analysis showed that most of the data came from single fish

observations. Therefore, Presence/Absence was considered the proper modelling

approach. Three microhabitat variables were measured (flow velocity, water depth and

substrate), with the methodology previously described for trout. They are considered

the most relevant variables for fish species distribution at this scale (Waters, 1976;


17

WP3: MORPH

Bovee, 1986; Heggenes, 1990; Gibson, 1993; Bovee, 1998). In order to use a single

numerical index, substrate composition was transformed into a Substrate index (S), but

we modified previous index (Mouton et al., 2011) in order to use more information

available about substrate (dominant, subdominant). The dominant substrate type was

considered as 75% of the sample, while the subdominant was the 25%. The weighing

categories were 0.08, 0.07, 0.06 ... 0.01, 0.00 for bedrock, boulder, cobble, gravel, fine

gravel, sand, silt and macrophytes respectively. Therefore, the developed index ranged

from 8 to 0. The maximum mean velocity was 2.13 m/s and maximum depth 2.75 m.

2.1.3 Ésera River

The study site at the Ésera River was surveyed during the summer 2012. The main

objective was to collect data for developing microhabitat models without a clear target

species. The Ésera River often is turbid, and thus, unsuitable for snorkelling. The

survey was thus carried out using electric traps consisting in a metallic frame that

conducted the rectified electricity provided by a generator. A set of three traps

distributed along the study reach in order to avoid frightening fish. The traps remained

undisturbed for at least 15 minutes before they were activated for 20 seconds. The

captures were measured and weighed and the hydraulic conditions measured. Velocity

was measured with an electromagnetic current metre (Valeport®) and depth was

measured with the wading rod at the nearest cm both, velocity and depth at the four

corners of the trap. The dominant and codominant substrate classes were visually

estimated around the sampling point (Bovee and Zuboy, 1988) considering the area

within the metallic cylinder used for sampling.

The amount of captures was extremely low, thus it was impossible to develop any

model based on them (Table 3). Therefore, to assess habitat changes in the Ésera

study site we adapted models from the literature about similar basins. However, the

collected data were used to calibrate qualitatively the developed models for brown trout

fry. The available information to develop the Expert-knowledge fuzzy models was

drawn from several sources depending on the life stage and activity, with the brown

trout as the target species, since it plays a crucial role in the generation of ecosystem

services (Schindler et al., 2010). The main information for adult and juvenile brown

trout come from Habitat Suitability Curves by Bovee (unpublished 1995). These curves


18

WP3: MORPH

were developed downstream of the Cheesman reservoir, South Platte River, in

Colorado State (USA). Two sets of velocity and depth curves were developed for adult

(n = 74) and juvenile (n = 153) brown trout. Nowadays there are curves developed in

the Iberian Peninsula (Muñoz-Mas et al. unpublished), but the surveys were conducted

under lower flows than those registered at our study site. However, the lack of curves

for substrate in Bovee's study force us to include the Iberian curves for that variable.

The substrate curves developed in the Iberian Peninsula include surveys carried out in

the Tagus Basin (TB) and the Jucar River Basin District (JRBD) in stretches of order

two and three (Strahler, 1952), with sample sizes of 98 and 140 records for adult and

juvenile respectively.

Table 3. Summary of the captures carried out in the study site of the Ésera River during the

summer 2012.

Trap number Species Length (mm) Weight (g) Category 1 Salmo trutta 86 6.9 Fry 1 Luciobarbus graellsii 98 12.9 Fry 2 Barbus haasi 61 3.6 Fry 2 Barbus haasi 69 4.9 Fry 2 Luciobarbus graellsii 84 6.4 Fry 3 Barbus haasi 111 17.4 Juvenile 6 Barbus haasi 70 3.2 Fry 6 Barbus haasi 71 5.3 Fry

12 Barbus haasi 70 2.9 Fry 13 Luciobarbus graellsii 232 150 Juvenile 17 Salmo trutta 84 8.4 Fry 21 Luciobarbus graellsii 365 405 Juvenile/Adult 35 Barbus haasi 57 2.4 Fry 39 Barbus haasi 300 230 Juvenile 54 Barbus haasi 154 55 Fry 58 Luciobarbus graellsii 130 34.1 Juvenile 59 Barbus haasi 54 2.4 Fry 64 Salmo trutta 72 4.2 Fry 70 Salmo trutta 86 6.7 Fry 74 Barbus haasi 123 23.9 Juvenile/Adult 93 Luciobarbus graellsii - - Adult 101 Barbus haasi 120 19.3 Juvenile/Adult

In the fry case the opposite occurred: most curves were developed in the Iberian

Peninsula but substrate information that came from American studies. Among the


19

WP3: MORPH

Iberian information, we selected those geographically closer, which corresponded to

curves developed in the Jucar River Basin District (43 records). In that set of curves, fry

showed maximum suitability for substrate over Vegetation/Silt (corresponding to

depositional areas), but the Ésera survey detected fry trout on cobbles and boulders.

Therefore, we decided to use other univariate Habitat Suitability Curves for substrate. A

search among the published literature was carried out selecting curves that showed

appreciable suitability over coarse substrata. We found the curves developed by

Raleigh (1984) and by Bovee (1978), both in the USA match our necessities. The

Raleigh curves were derived from the previous work by Gosse (1977; 1981) in Utah

rivers. However, running flows were measured only in 2 out of 3 rivers studied. The

curves developed for each river were joined in a unique curve which was used in the

Expert-knowledge fuzzy models for fry. The Bovee's curves were also included in the

analysis of the substrate suitability and its discretization in the form of fuzzy sets. The

Bovee's curves were developed based on expert judgment and literature, and are

theoretically suitable to be applied all along the USA, but no extra information was

available in the consulted document.

2.1.4 Macroinvertebrates

The macroinvertebrate study was conducted in summer 2003-2006. The study sites

were located in the headwaters of the Ebrón and Vallanca Rivers (Turia River

tributaries), Palancia River and the Villahermosa River (Mijares River Tributary) (Figure

6). The study sites (4 per river) were placed in unregulated streams providing a wide

range of flows from 0.02 to 1.84 m3/s. A complete climatic description of the involved

rivers is shown in Alcaraz-Hernández et al. (2011) and Mouton et al. (2011). The

selected reaches were at least 300 m long and the different Hydromorphological Units

(HMU) which were visually stratified accordingly to their different biotopes

(mesohabitats) in 4 different categories; Pool, Glide, Riffle or Rapid (see Alcaraz-

Hernández et al., 2011 for further details). Pools were characterized by large depth

(>0.6 m), generally associated with erosion phenomena, low flow velocity and very low

gradient. Glides were characterized by large depth (>0.6 m), low flow velocity and

constant section. Riffles were characterized by surface ripples, velocity smaller than

0.4 m/s, constant section and mean depth similar to mean substrate size. Rapids were

characterized by shallow depth, abundant surface turbulence and macro-roughness


20

WP3: MORPH

elements with dominance of supercritical flow. The HMUs were grouped in two types;

(i) slow flow, which included pools and glides and (ii) fast flow, which included riffles

and rapids. The grouping was based on previous studies on the physical dissimilarity of

the HMUs in these rivers (see Alcaraz-Hernández et al., 2011 for further details). In

each study site two HMUs were selected and the macroinvertebrate survey was carried

out, one from the fast group and another from the slow group with two replicates in

each HMU.

Figure 6. Locations of the microhabitat survey in the tributaries of the Mijares, Palancia and

Turia Rivers.

Sampling was carried out according to the International Standard ISO 8265:1988,

official version of the European Standard EN 29265, dated January 1994, using a Hess

cylinder. This is a stainless steel cylinder with both ends open and handles in the sides

to help pushing the sampler in the stream bed. The top edge is coated with plastic to

protect the operator. An oval hole in one side of the cylinder covered with 1 mm mesh

allows water to enter, whereas precluding the entrance of drifting invertebrates. A

second hole is connected to the collection net, with 0.5 mm mesh size and 500 mm

deep. The area surface of the cylinder is 0.05 m2 and it has a height of 450 mm and a

weight of 3 kg. The cylinder is introduced firmly on the bed of the river about 70 mm

deep, the sampler removes the substrate by hand to a depth of approximately 50 mm,

repeating the operation to ensure that all organisms are removed and transferred to the


21

WP3: MORPH

net. The organisms so collected are placed in containers, and preserved immediately

with 4% formaldehyde following ISO 5667-3.

Mean flow velocity, water depth and substrate type were the selected input variables, as they

yield similar information than more complex sets of variables such as Reynolds number of

Froude number (Jowett, 2003; Brooks et al., 2005). The percentages of each substrate class

were visually estimated within the sampled area (Bovee and Zuboy, 1988). To agree with

previous studies substrate composition was also converted into a single Substrate index (S) by

summing weighted percentages of each substrate type (Mouton et al., 2011). Back to the

laboratory, the samples were sieved to eliminate the formaldehyde, divided in Petri dishes and

cleaned with distilled water. Then, with the help of a binocular microscope, macroinvertebrates

were sorted and identified to the family level, genus level when possible. 109 samples were

finally studied. Table 4 summarizes the average density. However the present study was mainly

focused on three orders; Ephemeroptera, Plecoptera and Trichoptera usually termed as EPTs.

Table 4 Summary the collected macroinvertebrates. Only Order data were considered for further

analyses.

Type Name Aver. density (ind/m2) Prevalence Phylum Nematoda 70.18 0.18 Class Turbellaria 57.89 0.48 Class Gastropoda 1470.35 0.89 Class Bivalvia 3.33 0.08 Class Ostracoda 4.91 0.07 Class Hirudinea 3.86 0.08

Subclass Collembola 7.02 0.06 Subclass Oligochaeta 190.35 0.38 Order Amphipoda 734.04 0.68 Order Isopoda 0.35 0.02 Order Hymenoptera 1.40 0.05 Order Coleoptera 677.54 0.89 Order Diptera 2224.39 0.89 Order Trichoptera 221.93 0.75 Order Plecoptera 83.33 0.42 Order Ephemeroptera 344.39 0.78 Order Odonata 22.11 0.39

Suborder Heteroptera 1.93 0.05 Family Hydraenidae 250.88 0.66

The maximum surveyed mean velocity was 1.05 m/s and the maximum depth was 0.75

m, whereas the most common substrate type was cobble.


22

WP3: MORPH

2.2 Habitat suitability modelling for riparian vegetation - Riverine data collection

A study site was selected in a free-flowing segment of the Cabriel River river in the

Júcar River Basin District (Eastern Spain) (Figure 7). The study site, known as ‘Rabo

del Batán’, near Carboneras de Guadazaón (Cuenca province), had a good

conservation status (nearly-natural conditions), notable scenic value and relatively low

impact of agricultural activities. The coordinates at the centre of the reach are X

609259 – Y 4420615 (European Datum 1950, UTM-Zone 30N). According to the Júcar

River Basin Authority (CHJ, 2005; CHJ, 2009), the site is located in the water body

18.21.01.04, classified as natural. According to the ecological typologies defined by the

Spanish Ministry of Environment (CEDEX, 2004; CHJ, 2009), the study site belong to

the ecotype 12, named as ‘Mediterranean calcareous mountain river’. The Strahler

stream order ws 3.

Figure 7. Location of the study reach ‘Rabo del Batán’ within the Cabriel River Basin and the

Júcar River Basin District (Eastern Spain). The twenty cross-sections are shown in red.

2.2.1 Hydrological characterization

The Cabriel River is one of the most important rivers in Eastern Spain. It rises at 1620

m above sea level in Muela de San Juan (Teruel province) and flows 220 km in a

north-south direction to its confluence with the Júcar River in Cofrentes (Valencia

province). Its basin area is 4754 km². It shows low magnitude floods in spring and


23

WP3: MORPH

larger floods in autumn, usually between October and November. The summer is

normally a period of low flow.

Figure 8. Daily river discharges (m3/s) in the Cabriel-Rabo del Batán reach. Period: 1949-2009.

The Cabriel-Rabo del Batán reach is located at 925 m above sea level. It is 910 m long

and its catchment area is 962.60 km2. The distance from river source to the site is

93.47 km. Its mean annual discharge is 6.19 m3/s (Francés et al., 2009). The site is

located downstream from the Cristina weir. The gauging data in this site comes from

the station of Pajaroncillo (CEDEX), located only 3421 m upstream of the study site.

The daily river discharges are represented in Figure 8, and the characteristics of the

gauging station and flow time series are listed in Tables 5 and 6, respectively

(information available online at the CEDEX’s website: http://hercules.cedex.es/

anuarioaforos).

Table 5. Main characteristics of the Pajaroncillo flow gauging station at the Cabriel River.

Code X-Y UTM Catchment area (km2)

Level (m)

Location (from site) Period

Pajaroncillo 8090 610230- 829 940 Upstream 1949-

Table 6. Characteristics of the flow time series (m3/s) of the gauging station at the Cabriel River.

Mean annual flow

Min. annual flow

Max. annual flow

Min. monthly flow

Max. monthly flow

Annual coefficient of variation

Pajaroncillo 5.091 0.91 13.67 0.52 39.11 0.54

2.2.2 Riverine characterization


24

WP3: MORPH

The site selection was based on aerial photographs and field observations of plant

diversity and riparian quality. Currents and glides are the dominant hydromorphological

units in Cabriel-Rabo del Batán study reach (hereafter, CRA), although a big pool is

also present in the upstream section of the reach. The substrate of both riverbed and

floodplain is mainly composed of sand. The deposition areas, only covered with water

during floods, are composed by coarser substrate, mainly cobbles in a matrix of gravel

and fine gravel. In two sections of the reach (right bank downstream and left bank

upstream) secondary channels appear where water only flows in exceptional floods.

The riparian forest is dominated by specimens of willow, both in the shrub and tree

layer that densely cover both riverbanks, with some alternating habitats dominated by

reeds. The shrub layer dominating the bank zone is composed by Salix purpurea, Salix

eleagnos and Salix triandra. The tree layer dominating the bank and floodplain zone is

composed by Fraxinus angustifolia, Salix alba, Populus alba and Populus nigra,

accompanied by exemplars of Crataegus monogyna. The connectivity of the riparian

forest with the upland pine tree forest is total. In shallow margins, reeds (e.g.,

Phragmites, Sparganium, Thypa, etc.) dominate. These natural riparian formations are

protected by the regional government of Castilla La Mancha (Ley 9/1999, de 26 de

mayo, de Conservación de la Naturaleza), in particular the white poplar formations

(Rubio tinctorum-Populetum albae) and willow formations (Salicetum neotrichae).

Figure 9. Photographs illustrating the upstream (a) and downstream (b) sections of the Cabriel-

Rabo del Batán study reach.

Field data collection


25

WP3: MORPH

2.2.2.1 Vegetation transects

The first part of this study consisted on a survey of the riparian plant species (woody

shrubs and trees) along transects perpendicular to the stream flow and valley axis

(cross-sectional transects; Figure 7, Figure 10, Figure 11b and Figure 16), following the

‘line intercept method’ (first described by Canfield, 1941). Transects encompassed the

river channel and the complete floodplain, which is the zone built of sediments

deposited by the river and vegetated by riparian species. Transects finished upslope,

when stands dominated by terrestrial vegetation or agricultural uses were dominant.

Therefore, their length was variable, depending on the river valley width and land uses.

Figure 10. Cross-sections and steel rods in the Cabriel-Rabo del Batán study reach.

The cross-sections were marked out with one steel rod in each bank (Figure 11 a).

Absolute coordinates (X, Y and Z) were assigned to the top of each steel rod using a

GPS, model Leica® 1200 RTK. For those rods located in densely vegetated areas

where the GPS signal was weak, a total station FOIF® (Figure 11b) was used instead

to geo-reference them. In Figure 10 can be seen the aerial photograph of the reach

and the location of the cross-sections and steel rods.

Not all tree and shrub individuals have the same probability to be recorded, depending

on their size and shape. For this reason, we used strip transects, which are more

effective than line transects (Greenwood and Robinson, 2006). Sampling vegetation

within a certain distance of the line increases the number of plants sampled per unit


26

WP3: MORPH

length. In our case, we recorded the vegetation units intercepting with the transect, as

well as those located two meters apart from the line.

Figure 11. Example of a steel rod marking the location of a cross-section (a) and geo-

referenced survey of vegetation units using a total station FOIF® (b).

The vegetation unit consists of just one plant or a distinct group of plants (several

stems) of the same species and age, located in a reduced place and under the same

conditions (rather uniform soil characteristics, distance and elevation above water’s

edge). In this sense, we identified the vegetation units touching the transect and those

close to it, and recorded two points per vegetation unit, one at the beginning and

another at the end of the vertical projection of the crown of the plant on the transect line

(Figure 16). For each vegetation unit sampled, the number of stems was counted and

noted down. Two codes were assigned to each vegetation unit. For example, 02PA1i

and 02PA1f, meaning the beginning (i) and end (f) points of the first vegetation unit of

white poplar (Populus alba L.) recorded in transect number two. This type of transects

could be considered as small gradsects (gradient-directed transects), which are laid out

to sample intentionally the full range of floristic variation over a study area.

2.2.2.2 Soil sampling

Soil samples were collected within the study reach (Figure 10). The surveyors walked

along the site drawing a sketch of the different types of soils and writing a short

description of them. After that, all the information recorded about the site was

summarized and several points were sampled. The points selected covered the lateral

and transverse gradient of the river, i.e., the points were localized in representative soil

types from near to the water edge to the upland, and from upstream to downstream. In


27

WP3: MORPH

addition, a sketch of the soil types present in each cross-section was done. A soil type

(and hence, its characteristics) was assigned to each vegetation unit sampled in each

cross-section.

Figure 12. Appearance of some of the soil samples taken in the Cabriel-Rabo del Batán study

reach (a, soil 3; b, soil 2; c, soil 7; d, soil 6; e, soil 1; f, soil 12).

2.3 Habitat suitability modelling for fish and macroinvertebrates

2.3.1 Univariate Habitat Suitability Curves (HSCs)

Habitat Suitability Curves of Category II ½ and Category III (Bovee, 1986) were

developed, on purpose, for target species and life stages, including a curve for velocity,

depth and substrate. The curve generation procedure followed the usual standards

(Bovee, 1986). During the development of curves, each case was weighed by the

surveyed area of its corresponding river in order to equal the degree of influence of

each river on the resulting curve. This was followed by a frequency analysis of each

separate variable. The intervals used in the frequency analysis varied depending on

the involved species. In the brown trout case the frequency analysis involved several


28

WP3: MORPH

rivers. Therefore, the intervals were common to all rivers: 5 cm/s for velocity and 5 cm

for depth, whereas the intervals for substrate corresponded to the nine whole numbers

of the substrate index. However, in the redfin barbel case there was just one river.

Therefore, the selected interval consisted in the maximum value divided by the root of

the sample size for velocity and depth, whereas the substrate frequency was analysed

using the mentioned nine classes.

This frequency analysis was carried out for the Use data and the Availability data, then

both plots were standardized between zero and one. This analysis on the Use data

was used to develop Category II curves. The Category III curves were developed

applying the forage ratio (Savage, 1931; Cock, 1978). For the curves a smoothing

technique in R environment was applied, specifically the smooth.spline function in the

stats package (R Development Core Team, 2012). This procedure was applied to get a

unimodal curve because increases and decreases within the curve lack of ecological

sense. Expert-knowledge fuzzy modelling

The Expert-knowledge fuzzy models were based on the aforementioned univariate

Habitat Suitability Curves (HSCs) and the expert judgment of the corresponding

authors. The Expert-knowledge fuzzy models present the usual elements in a fuzzy

inference system, the input variables in the form of categories defined by fuzzy sets

and their membership functions (Zadeh, 1965), and the set of rules relating each

combination of the categories of the input variables with the corresponding output. Both

elements present their own development methodology.

The fuzzy set geometry selected was the trapezoidal, which showed successful in

previous studies (Mouton et al., 2007; Mouton et al., 2008; García et al., 2011). A

trapezoidal fuzzy set is defined by four parameters; am, bm, cm and dm which determine

the degree of membership of a given value to that fuzzy set. The membership degree

to a given fuzzy set increases from zero to one between am and bm, is equal to one

from bm to cm and decrease from one to zero from cm to dm. The region between am and

bm is shared with the adjacent, if exists, becoming a fuzzy region. The region between

bm and cm belongs, 'purely', to that fuzzy set whereas the region between cm and dm is

also shared with the adjacent, becoming a fuzzy region as well. Therefore the

discretization of the variables consists in a sequence of clearly defined regions that


29

WP3: MORPH

belong 'purely' to a given category with fuzzy defined regions inserted among them.

These fuzzy regions belong to both of the adjacent categories, thus to both of the

adjacent fuzzy sets. The determination of the parameters am, bm, cm and dm for each

fuzzy set of the three involved variables was based on the analysis of concavity and

convexity changes in the corresponding Habitat Suitability Curve. The values which

define each of these parameters were placed at the minimum and the maximum values

of the range covered by the curve, and in each point of the HSCs presenting any

change in the slope trend in term of changes in concavity and convexity of the curve.

All of those values hereafter called breaks of the curve.

The construction of fuzzy sets sequence to discretize the variable not started defining

the am parameter of the first fuzzy set, usually corresponding to the category Low,

because the maximum suitability of the curve, which tends to be a 'plateau' in the

curve, was intended to be included in a defined interval. Therefore, firstly the bm and cm

of the central fuzzy set was defined by the values of the breaks in the curve comprising

the maximum suitability for the involved variable, the aforementioned 'plateau'. These

breaks also defined the dm-1 and the am+1 parameters of the adjacent fuzzy sets. The am

and the dm of the central fuzzy set were defined by the following breaks both the right

and the left, which also correspond to the parameters cm-1 and bm+1. The further breaks

in both sides, the left and the right, corresponding then to the bm-1 and cm+1 and so on.

The procedure to define the fuzzy sets parameters continued until no more breaks

were defined for that curve (Figure 13 A). This procedure was applied for the remaining

HSCs constituting the first part of the fuzzy inference system for the three life stages.


30

WP3: MORPH

Figure 13. A) Example of discretization for depth in 3 fuzzy sets for adult brown trout. B) Expert-knowledge Output discretization. C) Data-driven Output discretization. Notice no overlapping,

but still providing a continuous output within the range from 0 to 1.

Instead of the procedure to determine the number, shape and parameters of the fuzzy

sets carried out for the input variables, the output variable was theoretically

determined. Te main objective was to produce a model whose output ranges, like the

Habitat Suitability Index, from zero to one. Accordingly, zero means unsuitable

whereas one means maximum suitability. The output variable was also discretized in

fuzzy sets selecting a three fuzzy set discretization corresponding to the categories

Low, Medium and High. The output variable was divided in five uniform intervals of 0.2

suitability length and the values of the limits of each interval defined the breaks to

construct the fuzzy sets like the input variable case. The first interval belong 'purely' to

the category Low followed by a fuzzy interval between Low and Medium suitability then

'purely' to Medium and so on (Figure 13 B).

Once the discretization of the input and the output variables, was done for each life

stage, the fuzzy rules were defined.

The fuzzy rules are constructed in an if-then sequence. Therefore, the number of rules

is always equal to the product of the number of fuzzy sets per variable and the main

goal at this point was to determine the proper output corresponding to each of these

combinations. The determination of the corresponding output was based on the

information derived from the habitat suitability curves following the next procedure.

First, the partial output for each separate variable and fuzzy sets was determined

analysing the values of the curve in the range covered by each fuzzy set. Generally,

the partial suitability output of the extreme fuzzy sets was always determined as 'Low'

in both extremes. The partial suitability output for the fuzzy set covering the highest

suitability of the HSC, usually with the shape of a 'plateau', was determined as 'High'

and the remaining fuzzy sets, with intermediate suitability, as 'Medium'. The combined

suitability output for each rule was determined following the next criteria. Generally, if

the depth was extremely low or extremely high the output of the rule including that

fuzzy set was always considered as Low, and could not be compensated by any better

suitability output from the remaining variables. If the velocity was high (i.e. the fuzzy set


31

WP3: MORPH

which included the maximum value), the combined suitability output was always Low

and neither could be compensated by any better partial suitability output from the

remaining variables. The combined suitability output for the remaining rules was

determined as the maximum appearance independently of any better remaining partial

suitability output. For instance, if velocity suitability was High, depth suitability was

Medium and substrate suitability was Medium the combined suitability was medium

neglecting the high suitability from the velocity variable. If a tie appeared, the suitability

for that rule was determined as Medium in any case not in conflict with the initial

assumptions: the lowest or highest depth and the highest velocity mean always Low

suitability. However, in the adult brown trout the previous rule was modified in the case

of depth, because habitat suitability has been shown to be independent of larger depth

(Bovee, 1977; Jowett and Davey, 2007; Ayllón et al., 2010). Therefore, for that case

the partial suitability for the largest depth was considered Medium. The remaining rules

followed the aforementioned criteria.

2.3.2 Data-driven fuzzy modelling

2.3.2.1 Sub-sampling

Prevalence can have a strong effect on model performance (Manel, 2001), and affects

the Data-driven models. Then, a sub-sampling procedure was recommended to obtain

a new Availability database which provides a 0.5 prevalence when compared with the

corresponding Use database. The sub-sampling methodology presented in Muñoz-Mas

et al. (2012) was applied to the Availability dataset extracting the same amount of

representative cases than the corresponding Use database. The procedure follows the

next steps: first the Euclidean distance of each data contained in the Availability

dataset to its centre of gravity was calculated, after:

Centre of gravity = (Average Velocity, Average Depth, Average Substrate)

These distances summarize the three microhabitat variables in a unique index. The

small-distance cases generally have more frequent values of velocity, depth and

substrate index, consequently are closer to the centre of gravity, than the large-

distance which corresponds to less frequent values, usually extremes. Then, the

accumulated frequency distribution of these distances was generated and the proper


32

WP3: MORPH

cases were selected (sub-sampled) following a systematic sampling procedure based

on that accumulated frequency distribution. The single variable distribution of each sub-

sample should be similar to the original Availability dataset. However, two statistical

tests were applied to check differences between each original dataset of Availability

and its respective sub-sample. These tests were applied to the three microhabitat

variables separately, and were a robust generalization of the Welch test (Welch, 1951)

and a robust generalization of the Kruskal-Wallis test (Rust and Filgner, 1984). After

the creation of the new Availability datasets based on sub-sampling, all of them were

pooled if necessary in the corresponding database to be used in the generation of the

Data-driven fuzzy models.

2.3.2.2 Model training

The development of the Data-driven fuzzy models followed the methodology presented

by Mouton (2008). Their development involves two main procedures similar to the

steps followed to develop the previously explained Expert-knowledge fuzzy models.

First the optimization of the fuzzy sets is carried out, and then the fuzzy rules are

developed.

The optimization of the fuzzy sets searches the optimal discretization of the input

variables in categories; Low, Medium, High etc. based on the Shannon–Weaver

entropy (Shannon and Weaver, 1963). The main goal was to obtain a uniform

discretization based on the number of cases included in each fuzzy set to improve the

results in the optimization of the fuzzy rules, otherwise if a fuzzy set of an input variable

contains very little training instances or no one, rules that apply this fuzzy set will be

trained inadequately. In the Data-driven fuzzy models the fuzzy set geometry is also

likewise defined by its membership function and, following similar criteria, the selected

geometry was trapezoidal. Like the Expert-knowledge methodology, four parameters

(am, bm, cm and dm) determine, the degree of membership of a given value to that fuzzy

set with similar meaning. The optimization of the fuzzy sets consists in the slight

modification of these parameters step by step. After each modification the Shannon–

Weaver entropy following (1) is calculated:


33

WP3: MORPH

(1)

where n is the number of classes and pi the proportion of cases belonging to the

category i. This modification is carried out while an improvement of the entropy is

obtained until the entropy threshold or the maximum possible entropy is achieved (see

Mouton et al. 2008 for further details). Once the fuzzy sets were optimized, they were

used in the optimization of their corresponding fuzzy rules.

Regarding the fish databases the Presence/Absence was the selected output

discretization. Although more gradual discretizations are preferred, the survey

methodology impeded any other discretization of the output variable because the

observed fishes were mostly individually spotted. The suitability output was then

discretized in two fuzzy sets with no overlapping areas (Figure 13 C). The absence of

overlapping areas do not provide always an integer suitability, zero for absence or one

for presence, because the final suitability obtained in the assessing of a given condition

depend on the centre of gravity of both fuzzy sets considering the areas under their

respective degrees of fulfilment. Therefore, a smooth transition from presence to

absence classification is possible. However, the integer zero and one values are the

most frequent outputs. The macroinvertebrate case presented differences, as the

orders with high prevalence (i.e. approximately 0.75) allowed the optimization of the

output likewise the input variables. Therefore, density models were developed for

Ephemeroptera and Trichoptera.

The Data-driven fuzzy models also present a set of rules that relate the input variables

with the output variable. The fuzzy rules optimization was carried out based on the

information contained in the pooled database for each life stage (i.e. the pooled data

from the Use dataset and the sub-sampled Availability) in the brown trout cases or

simply from the sub-sampled database of the redfin barbel. For macroinvertebrates no

subsampling was applied. The optimization was done with the software FISH© (Mouton

2010). During the optimization process FISH executes a defuzzyfication procedure

(Mouton et al., 2008) generating a fuzzy classification. Therefore, the observed and

modelled values become comparable, which is necessary to calculate a performance


34

WP3: MORPH

criteria. The entire optimization was based on Cohen's Kappa (Cohen, 1960).

However, in the optimization of habitat suitability for brown trout spawning, the True

Skill Statistic (TSS) was selected as performance criterion because it resulted to be

good in modelling from unbalanced prevalence databases (Allouche et al., 2006). The

fuzzy rules were optimized based on the hill-climbing algorithm (Michalewicz and

Fogel, 2000) in FISH. For each fuzzy rule (or set of antecedents) the process starts at

one random consequent (for example Low habitat suitability), then this consequent is

changed to its adjacent category (for example, Medium habitat suitability) and the

performance criterion is calculated. If the model performance increases in the current

step, the algorithm continues with the adjusted rule; if not, it retains the previous one

repeating the process with a new output category. To assess the model convergence

and robustness, 10 times three-fold cross validations were done, with five iterations

each. The value of a performance criterion was then calculated as the average value of

such criterion in the 150 resulting confusion matrices. The optimal consequent of a rule

was the consequent that occurred with the highest frequency in the optimizations.

However, in some cases does not exist any case to train a given rule (the so called

uncovered rules). Then Expert-knowledge or the information derived from validation

data were used to assess these improperly trained rules.

2.3.3 Probabilistic neural network modelling

2.3.3.1 Probabilistic neural network - Theory

PNNs are pattern classification radial-basis neural networks based on a Bayes-Parzen

classifier (Specht, 1990). PNNs basically compare the inputs with each of the

measurements included in the training database, and determine the probability of

membership of that combination of inputs to each one of the categories present in the

training database. To deal with differences in the intensity of the output, the weight of

each pattern is inversely proportional to the number of training patterns in the

corresponding category. Thus, the classification of certain conditions within a given

category depends on the values of the inputs, but not on the number of training

patterns included in that category.

In the present classification problem where two categories were considered (i.e. adult

brown trout presence or absence), the Bayes theorem considers a sample x=[x1, x2, x3,


35

WP3: MORPH

..., xp] which will be classified, for instance, in the category presence if the following

inequality is fulfilled: hP·iP·fP (x) > hA·iA·fA (x) where hi is the a priori probability of

occurrence, ii is the cost associated with misclassification and fi is the probability

density function of the corresponding category. The aggregation of these three

parameters defines the membership function. The Bayes theorem favours a class that

has a high density in the vicinity of the unknown sample (fi(x)) if the cost of

misclassification (ii) or prior probability (hi) are high (Hajmeer and Basheer, 2002). The

cost of misclassification (ii) and the prior probability (hi) allow the development of over-

predictive models where false positives are preferred, for instance in cancer diagnosis

(Berrar et al., 2003). In our study, the a priori probability of occurrence (hi) was

considered 0.5 and no misclassification costs (ii) were applied, thus, both factors were

neglected. In this case, the training samples must provide the information to estimate

the underlying multivariate probability density functions (f(x)) (Specht, 1990), after this

expression:

(1)

where x is the pattern to be classified and Xn is the ith training pattern, σ1, σ2, ..., σj are

the smoothing parameters that represent the standard deviation around the mean of

the p random variables, x1, x2, ..., xp, in the present study p=3, corresponding to

velocity, depth and substrate. The σ1, σ2, ..., σj control the degree of influence in the

vicinity of each training pattern. n corresponds to the total number of training patterns in

the considered category (1457 absences and 98 presences). Finally the present study

considered a single smoothing parameter, thus resulting in σ1=σ2=σ3.

The smoothing parameter (σ) has a decisive impact on the PNN performance (Figure

15). Therefore, its optimisation is recommended to obtain an optimal PNN (Hajmeer

and Basheer, 2002). If the smoothing parameter is too small, the multivariate

probability density function would be highly over-fitted to the training patterns, thus

reducing the network's capacity to generalize. However, if the smoothing parameter is

too large, the output value would be almost constant and proportional to the number of

training patterns in the considered class (presence or absence). Then, the values of the


36

WP3: MORPH

inputs would not play any role in the assessment of a given pattern (Zhong et al.,

2005).

2.3.3.2 Network operation and optimisation

PNN architecture differs from other ANNs like the Multilayer Perceptron (hereafter

MLP) (Figure 14). In the case of presence-absence classification, the PNN calculates

two multivariate probability density functions (f(x)) in parallel, one for each output

category. The input pattern (e.g. every pixel in an hydraulic model) will be classified in

the category of the output node that produces the most intense signal.

The first layer (the input layer) is a distributing layer, where x is the input pattern (i.e. a

combination of velocity, depth and substrate), connected to every node in the second

layer (the hidden layer). The hidden layer has an equal number of neurons as there are

training patterns (i.e. the 1555 collected patterns; n=98 presences, m=1457 absences).

In the hidden layer, the 'difference' between the input pattern and each training pattern

(XP1, XP2,..., XPn, and XA1, XA2, ..., XAm) is calculated. The third layer executes the

summation of the signals produced in the previous layer, but each category has an

independent summation of signals. This means that the absence output node is

connected only with the absence patterns (1457 connections; Figure 14) and the

presence output node only with the presence patterns (98 connections; Figure 14).

Once the sigma parameters are selected, the network is already prepared to assess

any pattern. That is the main reason why PNNs are considered a one-pass learning

method, because they are automatically trained by the patterns in the training database

(Specht, 1989).

Finally, the output of both nodes is standardized between 0 and 1, by dividing the

results with the sum of both outputs in order to agree with other habitat suitability

models. The optimisation of the PNN trained with the complete database (hereafter

PNNC) was carried out through the following steps. The three input variables (velocity,

depth and substrate) were normalized and the model was developed and optimised in

the R environment (R Development Core Team, 2012) by leave-one-out cross-

validation. Waters (1976) introduced the use of univariate Habitat Suitability Curves

(HSCs) assessing the degree of suitability of the usual microhabitat variables, such as

depth or velocity, ranging from 0 and 1. Accordingly, several studies, comprising


37

WP3: MORPH

difference techniques, ranged the habitat suitability between 0 and 1 (Jowett and

Davey, 2007; Ayllón et al., 2010).

Figure 14. General architecture of a Probabilistic Neural Network (PNN) in a presence-absence classification problem. x corresponds to the assessed pattern. XP correspond to the presence

training patterns (n=98) whereas XA to the absence training patterns (m=1457).

Figure 15. Effect of the selection of different values of a single smoothing parameter (σ) in the habitat assessment of a pool at the Cabriel River (Z=elevation). The larger the σ, the smoother

the classification. Large values of the smoothing parameter sigma (0.25) do not provide the extremes of the output range (0-1) whereas lower values (0.025) provide sharper transitions

achieving the extremes of the output range.

However, large values of σ usually do not provide the extreme feasible outputs (Figure

15). Therefore, two main goals were included in the objective function, the

maximization of the classification strength and the maximization of the output range.

The classification strength was quantified by means of the True Skill Statistic (TSS)


38

WP3: MORPH

because it has been demonstrated suitable in the modelling of unbalanced prevalence

databases (Allouche et al., 2006), and it favours a good balance between sensitivity

(Sn) and specificity (Sp), whereas the output range was considered by subtracting the

minimum output to the maximum output. The objective function, finally, aggregated

both values. The subplex algorithm proposed by Rowan (1990) and implemented in the

R environment (R Development Core Team, 2012) by King (2008) was used to

optimise the smoothing parameter σ.

2.4 Habitat suitability modelling for riparian vegetation

2.4.1 Laboratory analysis

2.4.1.1 Soil analysis

Soil samples were analysed in the Soil Laboratory of the Universitat Politècnica de

València. Three analyses were conducted to define the texture and organic content:

• Particle size analysis by sieves (125-0.08 mm, UNE 103.101-95).

• Particle size analysis of very fine materials (0.08-0.001 mm) by sedimentation:

Hydrometer method (UNE 103.102-95).

• Determination of the oxidisable organic matter content by Potassium

permanganate method (UNE 103.204-93).

Thus, the following parameters were determined:

• Percentage of gravel (> 2 mm)

• Percentage of sand (0.05 mm - 2 mm)

• Percentage of silt (0.002 mm - 0.05 mm)

• Percentage of clay (< 0.002 mm)

• Percentage of organic matter


39

WP3: MORPH

Additionally, the moisture content was determined for each soil sample by a gravimetric

method, where the sample was weighed before and after oven-drying at 115 ºC during

approximately 24 hours.

2.4.2 Data analysis

2.4.3 Definition of vegetation indicators

Transects are commonly used to survey changes in vegetation along an environmental

gradient, as well as to estimate overall density or cover of species in an area (Bullock,

2006). Based on the data recorded in the field by transects, we calculated the

following vegetation variables:

• Richness. Total number of species sampled in the study reach.

• Abundance. Number of individuals (stems) of each species touching the

transect lines.

• Cover. Proportion of ground within the transects occupied by the vertical

projection of all individuals of a species. Because the vegetation may be

layered, the cover of all species often sums to > 100%. In this study, we

estimated the percentage cover of each species in relation to the total cover

sampled.

• Density. Number of stems in a prescribed area. Further, we estimated the

number of stems of each species (in relation to the total length of all transects).

Therefore, we obtained two indicators of density: m of cover of a certain species

per m of transect and, number of stems of a certain species per m of transect.

2.4.3.1 Definition of morphological variables

All the vegetation points recorded in the field with the total station were imported into

ArcGISTM version 9.3 (ESRI, Redlands, California, 2009). A straight line was drawn

joining the steel rods of each transect. Then, the vegetation points were projected in

this line in order to correct small deviations from the line during the surveys.


40

WP3: MORPH

Figure 16. Cross-sectional view of a theoretical transect. The position of the points recorded in

the field is shown: beginning and end of the vertical projection (to estimate cover) and predicted ground surface level of the main stem (to estimate distance and elevation to/above thalweg).

From the topographic information of each transect, the thalweg was identified as the

deepest point of the channel bed. Then, the coordinates of the centre of each

vegetation unit were obtained (using for this the beginning and end points of each unit).

Finally, we calculated the location relative to the thalweg for all species, i.e., relative

linear distance from the vegetation unit to thalweg and relative elevation (vertical

distance) from the vegetation unit above thalweg (Figure 16). In addition, dominance

curves were developed for all species based on the morphological variables, in order to

assess the abundance of each species (frequency) and their preference for specific

places along the entire lateral gradient.

2.4.3.2 Positional patterns and definition of guilds

The location of the riparian species recorded was compared based on distance and

elevation above thalweg using box-plots, and the Huber estimator was calculated for

each species. The Huber estimator belongs to a class of estimators called M

estimators that represent a compromise between the mean and the median. It is

obtained by minimising a term involving the sum of errors. Therefore, it could be

defined as a robust estimate of location (i.e., population mean) that reduces the effects

of outliers in the data (García-Pérez, 2005).


41

WP3: MORPH

Robust and non-parametric statistics were applied for most of the calculations because

the distribution of species in the riparian corridor did not follow a normal distribution.

Furthermore, the presence of outliers in these distributions made necessary the use of

robust statistics, which reduce the influence of outliers in the rest of the dataset. In this

sense, a robust analysis of variance was carried out (robust Welch one-way ANOVA; R

function t1way) to contrast the null hypothesis of equality in the parameters of position

of the r independent populations, because the usual hypotheses of normality and

homoscedasticity (equal variances) could not be assumed. We wanted to contrast the

null hypothesis (H0) of equality of means for all the riparian woody species in relation to

the two morphological variables (distance and elevation above thalweg),

H0 : μα,Salix purpurea (SP) = μα,Populus alba (PA) = … = μα,r

H1 : μα,Salix purpurea (SP) ≠ μα,Populus alba (PA) ≠ … ≠ μα,r

versus the alternative hypothesis (H1) of not being all the means the same. In all cases,

these comparisons make use of sample α-trimmed means , , …, ,

obtained after sampling in the r populations to be compared, and using random

samples of size nSP, nPA, … nr. The amount of trimming α and the significance level

were the same in the r populations (10 % and 0.05 in most cases). The null hypothesis

(H0) was rejected for large values of the FWe statistic, which follows an F Snedecor

distribution if H0 is true.

When the null hypothesis was rejected, a test for multiple comparisons (R function

lincon) was applied to define groups of species with the same response in relation to

the variables analyzed. The significance level applied was 0.05 and the trimming

percentage 10%.

2.4.3.3 Soil characteristics

Barplots with error bars were drawn to illustrate the differences in the soil

characteristics for each species. Statistical program R and the package Hmisc (Harrell,

2010) was used for it. The error bars represented the boundaries for the 95%

confidence interval.


42

WP3: MORPH

2.4.3.4 Multivariate analysis

A Principal Components Analysis (hereafter, PCA), followed by Cluster Analysis was

performed to define riparian guilds, i.e. non-phylogenetic groups of species that

respond in a similar way to the measured environmental variables, which ultimately

depend on the flow regime. Although the variables did not follow the normal

distribution, the PCA was performed because when the purpose is just exploratory, the

normality of the variables is not a strict requirement. The variables used in this analysis

were: DIST_THW, distance to the thalweg; ELEV_THW, elevation above thalweg;

GRAVEL, percentage of gravel; SAND, percentage of sand; SILT, percentage of silt;

CLAY, percentage of clay; OM, organic matter content; SM, soil moisture content. The

PCA ºwas carried out using the function prcomp from the package stats of R (R

Development Core Team, 2010). The average values of the scores of the specimens of

each species on each principal component were calculated. The values belonging to

the first two components were overlaid onto the PCA plot, which represented the

average value of each species. For better interpretation of the association among

species to certain physical variables, such values were scaled.

Then, a Hierarchical Cluster Analysis was conducted using for each species, the

average values of the scores of the first three principal components extracted from the

PCA, which summarized most of the total variance. This was useful for a better

definition of the association among species. Firstly, a proximity matrix using Euclidean

distance was generated from the PCA scores information. Then, the function hclust,

from the package stats of R (R Development Core Team, 2010), performed a

hierarchical cluster analysis using a set of dissimilarities for the n species being

clustered. Initially each species is assigned to its own cluster and then the algorithm

proceeds iteratively, at each stage joining the two most similar clusters, continuing until

there is just a single cluster. Average linkage clustering was selected as agglomeration

method. It is the method recommended in community ecology because it seems to be

more neutral in grouping and a compromise between the single (nearest neighbour)

and the complete (further neighbour) linkages (Oksanen et al., 2011). It makes fusions

between group centre points, and its fusion levels are between the previous two trees.


43

WP3: MORPH

The resultant dendrogram was cut at some fusion level in order to have a certain

number of clusters (specifically, in this analysis the tree was cut at height =1). This

value is associated with the particular clustering method being used and must

represent a good compromise between loss of information and interpretability of the

similarities among species being agglomerated. Additionally, the clustering results were

displayed in ordination diagrams. The function ordicluster (within package vegan and

compatible with hclust) overlays a cluster dendrogram onto ordination (based on the

proximity matrix). It combines points and cluster midpoints similarly as in the original

cluster dendrogram, i.e., it connects cluster centroids to each other with line segments.

According to Oksanen et al. (2011), overlaying classification in ordination can be used

as a cross-check: if the clusters look distinct in the ordination diagram, both analyses

probably were adequate.


44

WP3: MORPH

3 RESULTS



3.1.1.1 Salmo trutta in the Cabriel River

Figure 17. Habitat Suitability Curves for adult brown trout. Upper sequence Category II ½

curves (Nuse = 98). Lower sequence Category III curves (Nuse = 98,Navailability = 2289).

The generated Category II ½ curves for the adult brown trout provided unimodal

pointed curves with maxima in 0.125 m/s for velocity, 0.35 m for depth, and 6 s for

substrate (corresponding to cobble) (Figure 17). The Category III curves showed also

uni-modality, presenting a clear displacement of the highest suitability to higher velocity

and larger depth, 0.8 m/s and 1.3 m respectively, whereas the highest suitability for the

variable substrate was reduced to 4.5 s (corresponding to gravel). The depth curve

showed a clear pointed shape, while the remaining were wider than the corresponding

Category II ½ (Figure 17).


45

WP3: MORPH

Figure 18. Habitat Suitability Curves for juvenile brown trout. Upper sequence Category II ½

curves (Nuse = 140). Lower sequence Category III curves (Nuse = 140,Navailability = 1978).

Figure 19. Habitat Suitability Curves for fry brown trout. Upper sequence Category II ½ curves (Nuse = 44). Lower sequence Category III curves (Nuse = 44,Navailability = 1114).


46

WP3: MORPH

The juvenile life stage provided in the Category II ½ case unimodal curves with maxima

in 0.175 m/s, 0.35 m and 6 s (Figure 18). The Category III curves also showed uni-

modality, and presented a certain displacement of the highest suitability to higher

velocity and larger depth, 0.55 m/s and 0.55 m respectively, whereas the highest

suitability for substrate was reduced to 4 s (corresponding to gravel). The difference

between Category II ½ and Category III curves was smaller for juveniles than for adults

(Figure 18).

In the Category II ½ the fry life stage showed unimodal curves with maxima at 0.12

m/s, 0.25 m and 0 s (corresponding to silt and vegetation) (Figure 19). The Category III

curves showed uni-modality, and presented certain displacement of the highest

suitability to higher velocity 0.55 m/s and larger depth 0.57 m, whereas the highest

suitability for the variable substrate remained constant (Figure 19).

3.1.1.2 Barbus haasi in the Mijares River and Siurana River

The generated Category II ½ curves for redfin barbel (Barbus haasi) provided unimodal

curves with maxima in 0.0 m/s, 0.75 m and 5 s (gravel) (Figure 20). The Category III

curves showed also uni-modality. The velocity variable showed no clear displacement,

whereas depth presented a small displacement of the highest suitability to larger depth;

1.2 m. The highest suitability for substrate was completely switched, showing the

maximum suitability for 1 s (silt), followed by the 8 s (bedrock) (Figure 20).


47

WP3: MORPH

Figure 20. Habitat Suitability Curves for redfin barbel. Upper sequence Category II ½ curves

(Nuse = 91). Lower sequence Category III curves (Nuse = 91, Navailability = 345).

3.1.1.3 Salmo trutta in the Ésera River

The selected Category II curves for the adult brown trout provided unimodal curves

with maxima in 0.5 m/s, 1.0 m and 6 s (cobble) (Figure 21). Juvenile brown trout

yielded Category II unimodal curves with maxima in 0.5 m/s, 0.5 m and 6 s (cobble)

(Figure 21). Notice that the depth curve showed a narrow suitable interval (Figure 21).

In contrast with the previous cases, the fry life stage was based on data from the Jucar

River Basin district, but the substrate curve was based on the literature. The curves

were also unimodal with maxima in 0.25 m/s, 0.35 m and 5 s (gravel) (Figure 21).


48

WP3: MORPH

Figure 21. Habitat Suitability Curves for adult, juvenile and fry brown trout. Adult and juvenile

based on Bovee (unpublished 1995) (Nadult = 74, Njuvenile = 153). Fry based on data collected in the Jucar River Basin District (Nfry = 44).

3.1.2 Multivariate Habitat Suitability Models

3.1.2.1 Salmo trutta in the Cabriel river

Fuzzy modelling

The Expert-knowledge approach for adult brown trout yielded for both, the models

derived from Category II ½ and Category III curves, three Fuzzy Sets for velocity, four

Fuzzy Sets for depth and two Fuzzy Sets for substrate, presenting certain similarity in

the shapes but strongly differing on the partial suitability (i.e. the suitability of the

covered area of the HSC) (Figure 22).

As the sub-sampling methodology showed no statistical differences between the

original database and the extracted sub-sample, the sub-sampled database was

considered adequate to develop habitat suitability models for adult brown trout. The

Data-driven approach discretized velocity in five Fuzzy Sets, depth in two Fuzzy Sets


49

WP3: MORPH

and substrate also in two Fuzzy Sets, achieving the Shannon–Weaver entropy values

of 0.84, 0.94 and 0.94, respectively (Figure 22).

Figure 22. Category II ½ and Category III Habitat Suitability Curves and their corresponding

fuzzy sets for adult brown trout. The last sequence corresponds to the Data-driven fuzzy sets obtained from the Shannon-Waver entropy based optimization.

The Expert-knowledge approach provided different discretization in the models derived

from Category II ½ and Category III for the juvenile brown trout. The derived from the

Category II ½ curves provided two, four and two Fuzzy sets for velocity, depth and

substrate, respectively, whereas the Category III derived presented similar

discretization than the adult case with three, three and two Fuzzy sets (Figure 23).


50

WP3: MORPH

As the sub-sampling methodology again showed no statistical differences between the

original dataset and the subsample, the resulting database was considered suitable for

further analysis. Under the Data driven approach, five Fuzzy Sets were obtained for

velocity, 3 for depth and 2 for substrate, achieving the Shannon–Weaver entropy

values of 0.93, 0.98 and 0.87, respectively (Figure 23).

The Expert-knowledge approach led to these numbers of Fuzzy sets: Velocity was

discretized in 2 Fuzzy Sets, depth in 3, substrate in 2. The discretization derived for the

Category III curves involved 3 Fuzzy sets for three variables (Figure 24).

The sub-sampling methodology showed no statistical differences. The Data driven

approach discretized velocity in five Fuzzy Sets, depth in three and substrate in two

Fuzzy Sets, achieving the Shannon–Weaver entropy values of 0.92, 0.98 and 0.87,

respectively (Figure 24). Notice that the variable substrate presented almost no

overlapping.


51

WP3: MORPH


fuzzy sets for juvenile brown trout. The last sequence corresponds to the Data-driven fuzzy sets obtained from the Shannon-Waver entropy based optimization.

The Expert-knowledge Fuzzy Rules are summarized in the Table 7. The Data-driven

optimization of the fuzzy rules achieved the values 0.36 ± 0.4, 0.3 ± 0.4 and 0.5 ± 0.9 of the

Cohen's Kappa for adult, juvenile and fry, respectively. The training for the adult case presented

three rules with no cases to be trained, and thus Expert-knowledge was used in their

determination. The juvenile case presented five rules with no cases to train them and were also

assessed using Expert-knowledge. Finally the fry cases presented the major number of

undefined rules; six rules and similarly Expert-knowledge approach was used to their definition

(Table 7).


52

WP3: MORPH

Figure 24. Category II ½ and Category III Habitat Suitability Curves and their corresponding fuzzy sets for brown trout fry. The last sequence corresponds to the Data-driven fuzzy sets

obtained from the Shannon-Waver entropy based optimization.


53

WP3: MORPH

Table 7 Summary of the Fuzzy Rules for the models derived by Expert-knowledge (EK), either from Category II ½ or Category III curves, and the Fuzzy Rules obtained by Data-driven approach (DD). The table also includes the rules adjusted after their validation in the study site. Asterisk (*) means adjusted rule. Double asterisk (**) means uncovered rule. The output was determined through authors’ consensus. Suitability in 5 categories corresponds to Very Low (VL), Low (L), Medium (M), High (H) and Very High (VH).

Rule Nº Velo

city

(m/s

)

Dep

th (m

)

Subs

trat

e (s

)

EK A

du C

at II

EK A

du C

at II

I

DD

Adu

lt

EK J

uv C

at II

EK J

uv C

at II

I

DD

Juv

EK F

ry C

at II

EK F

ry C

at II

I

DD

Fry

B.t.

Spa

wni

ng

1 L VL L L L L L L 2 M VL L L L L L 3 H VL L L L L L L 4 VL L L L L H 5 L L L H M L H H H H L L L 6 M L L M M L M M H L H L 7 H L L L L H L L L L L L L 8 VH L L H L L** L** 9 VL M L L H

10 L M L M M M M L H H L/H* L 11 M M L M H M M L H L L 12 H M L L L L L H L L L L 13 VH M L L** L** L** 14 VL H L L L H 15 L H L L L H L L H L L H L 16 M H L L L H L L H L H L 17 H H L L L H** L L H** L L L** L 18 VH H L L** L** L** L** 19 L VH L L** 20 M VH L L** 21 H VH L L** 22 VH VH L L** 23 L L M L H 24 M L M L H 25 H L M L L 26 VH L M L** 27 L M M H H 28 M M M M H 29 H M M L H 30 VH M M L** 31 L H M L H 32 M H M L H 33 H H M L L


54

WP3: MORPH

Table 7. (continues)

Rule Nº

Velo

city

(m/s

)

Dep

th (m

)

Subs

trat

e (s

)

EK A

du C

at II

EK A

du C

at II

I

DD

Adu

lt

EK J

uv C

at II

EK J

uv C

at II

I

DD

Juv

EK F

ry C

at II

EK F

ry C

at II

I

DD

Fry

B.t.

Spa

wni

ng

34 VH H M L** 35 L VH M L** 36 M VH M L** 37 H VH M L** 38 VH VH M L** 39 L VL H L L L L L 40 M VL H L L L L 41 H VL H L L L L L 42 VL L H L/H L L 43 L L H H M L/H H H H H L H L 44 M L H H H H H H H L L L 45 H L H L L H L L H L L L L** 46 VH L H L L L** L** 47 VL M H H H** 48 L M H H H H H H M H L/H L 49 M M H M H M M H M L L 50 H M H L L L L H L L L L 51 VH M H L H L** 52 VL H H L L L/H 53 L H H L L H L L L/H L L L/H L 54 M H H L L H L L H L H L 55 H H H L L H** L L H** L L L L** 56 VH H H H L** L L** 57 L VH H L** 58 M VH H L** 59 H VH H L** 60 VH VH H L**

The assessment of the modelled study site showed different results for the adult life

stage. The Expert-knowledge model based on Category II ½ curves assessed most of

the reach as highly suitable, except the shores and the deep areas corresponding to

the northern and central areas of the study site (Figure 25 - EK type A). The Expert-

knowledge model based on Category III presented higher variability and the suitability

gradually increased from the shallower areas to the deeper (Figure 25 - EK type B).

The unmodified Data-driven model assessed most of the reach as unsuitable for adult

brown trout, except the deeper areas, that were assessed with High suitability (Figure


55

WP3: MORPH

25 - DD type A). Two rules were modified in order to maximize the number of fish

location assessed with the maximum suitability. The assessment showed most of the

reach as highly suitable, except the shores, which remained unsuitable (Figure 25 - DD

type B).

Figure 25. Habitat assessment for adult brown trout carried out with the four generated models. EK type A means Expert-knowledge based on Category II ½ curves, EK type B means Expert-knowledge based on Category III curves, DD type A means Unmodified Data-driven model and

DD type B means Modified Data-driven model.

The assessment of the modelled study site for the juvenile life stage showed

apparently similar results for both models within the Expert-knowledge approach, but

differences respect to the Data-driven approach. The Expert-knowledge model based

on Category II ½ curves assessed most of the reach with High suitability, except the

shores and the deep areas, the latter especially significant in the northern area (Figure

26 - EK type A). The Expert-knowledge model based on Category III curves presented

higher suitability all along the surface, except a narrow line close to the shores (Figure

26 - EK type B). The unmodified Data-driven model assessed most of the reach as

highly suitable, except the deep areas, which presented null suitability (Figure 26 - DD

type A). One rule was modified in order to maximize the number of fish location

assessed with the maximum suitability. The assessment showed variable suitability

from shallow, unsuitable areas to deeper areas with the highest suitability (Figure 26 -

DD type B).


56

WP3: MORPH

Figure 26. Habitat assessment for juvenile brown trout carried out with the four generated

models. EK type A means Expert-knowledge based on Category II ½ curves, EK type B means Expert-knowledge based on Category III curves, DD type A means Unmodified Data-driven

model and DD type B means Modified Data-driven model.

The assessment of the modelled study site for fry showed differences between models.

The Expert-knowledge model based on Category II ½ curves assessed most of the

reach as highly suitable, except the shores and the deep areas, which were assessed

as unsuitable (Figure 27 - EK type A). The Expert-knowledge model based on

Category III presented high suitability all along the surface, but keeping certain

decrease on the deeper areas (Figure 27 - EK type B). The unmodified Data-driven

model presented a messy pattern, assessing shallow and deep areas as highly suitable

but intermediate conditions as unsuitable (Figure 27 - DD type A). The fry case

presented the maximum amount of modified rules in order to maximize the number of

fish location assessed with the maximum suitability. Four rules were modified to

achieve the objective and the messy pattern was reduced, although it appeared in two

areas in the middle part of the model, which also presented high concentrations of fry

(Figure 27 - DD type B).


57

WP3: MORPH

Figure 27. Habitat assessment for brown trout fry carried out with the four generated models.

EK type A means Expert-knowledge based on Category II ½ curves, EK type B means Expert-knowledge based on Category III curves, DD type A means Unmodified Data-driven model and

DD type B means Modified Data-driven model.

The frequency analysis of the habitat assessment over the simulated validation site

showed that, disregarding Data-driven modifications, the Expert-knowledge models

were in general more optimistic (higher suitability values) than the Data-driven models

(Figure 28).

The Expert-knowledge fuzzy models based on Category II ½ curves indicated, for the

three size classes, that most of the reach had the maximum suitability; thus the

sensitivity was high but not the specificity (Figure 28, upper sequence). The models

presented large frequency of the higher suitability (0.6 to 1.0) within the entire reach

(Figure 28 black bars), similar to the frequency analysis of the assessment of trout

locations (i.e. habitat Use) (Figure 28 grey bars). The frequency analysis for the Expert-

knowledge fuzzy models derived from Category III HSCs (Figure 28 second sequence)

presented most of the area with maximum suitability (0.8 - 1.0) for juvenile and fry

(Figure 28 black bars), and all the observed trout locations in both size classes were

assessed with the maximum suitability (Figure 28 grey bars). In that class of maximum

suitability, there is an over-proportion of habitat Use in relation to Availability.

Therefore, the models for juvenile and fry based on Category III HSCs presented

perfect sensitivity but low specificity.


58

WP3: MORPH

Figure 28 Frequency analysis of the habitat assessment carried out with the four models and

the three size classes (black bars), over the entire simulation reach (Availability data). Frequency analysis of the habitat assessment carried out with the four generated models and the three size classes (grey bars) over the corresponding size class locations (Use data). EK

mean Expert-knowledge and DD Data-driven.

The adult model presented the habitat assessment spread along the considered

categories in comparison with the previous models. The largest frequency of

Availability (Figure 28 black bars) appeared for the middle ranged suitability (i.e.

suitability from 0.4 to 0.6) and the frequency decreased in both sides (Figure 28 black

bars). However, there was no habitat Use in the lower suitability intervals (Figure 28


59

WP3: MORPH

grey bars), and an appreciable amount of trout locations were assessed within the

higher suitability intervals (i.e. suitability ranging from 0.4 to 1.0). As in the two previous

models, there was an over-proportion of habitat Use in relation to the Availability; this

means that there was a positive selection of the fish towards the microhabitats of

higher suitability, as expected. Therefore it presented good sensitivity but the better

specificity among the Expert-knowledge fuzzy models (Figure 28 first and second

sequences).

The unmodified Data-driven fuzzy models presented major disparity among results

(Figure 28 third sequence). The adult case presented a good trade-off between

sensitivity and specificity (Figure 28 third sequence - Adult). The largest frequency of

the assessment of the entire reach appeared in the lowest suitability interval (i.e.

suitability ranging from 0.0 to 0.2) (Figure 28 black bars) whereas the maximum

frequency of adult locations corresponded to the highest suitability interval (i.e. 0.8 -

1.0) (Figure 28 grey bars). However, according to the premise that overprediction is not

necessarily an ecological error, the Data-driven model for adults was modified by

enlarging the adult locations assessed with the maximum suitability (Figure 28 last

sequence - Adult). It maximized the trout locations assessed with the maximum

suitability (i.e. suitability ranging from 0.8 to 1.0) (Figure 28 grey bars), and thus,

maximized the sensitivity and retained certain specificity, because the Availability data

presented a relevant proportion with the lowest suitability (i.e. suitability ranging from

0.0 to 0.2) (Figure 28 black bars).

The frequency analysis of the unmodified Data-driven model for juveniles showed most

of the reach and most of the habitat use at microhabitats with the highest suitability (i.e.

suitability ranging from 0.8 to 1.0) (Figure 28 third sequence - Juvenile). Thus, it

presented a high sensitivity and relatively low specificity (Figure 28). However, an

appreciable number of trout were located in areas assessed as unsuitable (i.e.

suitability 0.0 - 0.2). The modification of a single rule (Rule 4, Table 7) displaced the

assessment (Figure 28 last sequence - Juvenile) of that juvenile locations to higher

suitability values (Figure 28 grey bars), but the specificity remained almost constant

(Figure 28 black bars).


60

WP3: MORPH

The unmodified Data-driven fuzzy model for fry presented the worst results (Figure 28

third sequence - Fry). The maximum frequency appeared for the lowest suitability (i.e.

suitability 0.0 - 0.2) and most of the fish locations were assessed as unsuitable (Figure

28 grey bars). Thus, the sensitivity of the model was low. The modification of a single

rule (Rule 15, Table 7) improved the results (Figure 28 last sequence - Fry),

maximizing the trout locations with maximum suitability (i.e. 0.8 - 1.8) (Figure 28 grey

bars), but keeping an acceptable trade-off between sensitivity and specificity.

The density analysis showed that adult brown trout appeared sparsely distributed in the

reach, but with a peak of density in the northern part of the validation site (Figure 29).

Although the juvenile brown trout appeared all along the study site (Figure 29), density

peaked close to the adults, which conditioned the generated density categories (Figure

29). Both cases presented the maximum density in areas with a relatively large depth

and in an area where the flow was concentrated downstream from a relatively fast

habitat. The fry appeared more sparsely distributed, with several areas of very high

density; in general, they were far from the elder individuals (Figure 29).

The correlation of the assessed suitability and the observed density showed different

patterns depending on the methodology and the size class (Figure 30). The Expert-

knowledge model based on Category II ½ HSCs for adults showed a positive trend

between average suitability and density, but a decrease on the average suitability in

the most densely populated interval (Figure 30 upper plot). The Expert-knowledge

fuzzy model based on Category III HSCs presented a positive correlation between

density and average suitability, achieving the maximum suitability in the most highly

populated areas (Figure 30 upper plot).


61

WP3: MORPH

Figure 29. Density map of brown trout in the Cabriel River evaluation site.

The unmodified Data-driven presented an increasing trend as density increased,

whereas the modified Data-driven presented similar pattern, but with higher average

values than in the unmodified model. The Expert-knowledge fuzzy model based on

Category II ½ HSCs for juveniles presented a flat trend as the density increased from

very low density to medium density, and decreased for larger densities. The most

densely populated areas did not coincide with any area with the highest suitability

(Figure 30 middle plot). The Expert-knowledge fuzzy model based on Category III

HSCs presented high average suitability for the lowest density interval, achieving the

maximum suitability for the remaining density intervals, regardless of the observed

density interval. The unmodified Data-driven fuzzy model for juveniles presented

medium average suitability with a slight increasing trend as the density increased,

although with some irregularities (Figure 30 middle plot). The modified Data-driven

fuzzy model presented an increment of the average suitability at the lower density

intervals, presenting the maximum average suitability for the remaining intervals.


62

WP3: MORPH

Figure 30. Average suitability (black lines) for the five generated density categories, by fish size

class. The frequency distribution of the assessed suitability on each density area was very skewed, thus the range of the habitat assessment per density category were plotted. These values corresponded to the maximum and the minimum values assessed in the study area.

The Expert-knowledge fuzzy models based on Category II ½ and in the Category III

HSCs presented high average suitability regardless the density interval (Figure 30

lower plot). The unmodified Data-driven fuzzy model presented positive correlation

between average suitability and fry density (Figure 30 lower plot). However, as the


63

WP3: MORPH

juvenile, it also presented some irregularities. The modified Data-driven fuzzy model

presented and improvement compared to the unmodified Data-driven fuzzy model, with

higher values of the average suitability at any considered density interval, and not

presenting pronounced irregularities.

The prevalence of the spawning database was 0.33. Therefore, sub-sampling was not

considered necessary for model optimization.


optimization for brown trout spawning. An extra fuzzy set corresponding to extreme conditions was added for velocity and depth. The variable substrate was discretized in two Fuzzy Sets, but the fuzzy region covered the suitable substrate. The Expert-knowledge approach was used to

discretize the variable substrate properly in three Fuzzy sets.

The Data-driven approach discretized velocity and depth in three Fuzzy Sets, and

substrate in two Fuzzy Sets, achieving the Shannon–Weaver entropy values of 0.93,

0.92 and 1.00 respectively (Figure 31). The survey methodology recommended the

addition of an extra fuzzy set for velocity and depth. The surveyor measured redd

conditions (two points) at redd head and redd centre, and the surrounding area (four

points) upstream, downstream and left and redd right, but redds did not appear in fast

flow or deep areas. Therefore, the database presented certain but well known bias. In

order to avoid miscalculations of habitat suitability (i.e. model calculus could

extrapolate extreme conditions and show them as suitable areas, but in fact

corresponding to areas which never could contain trout redds), an extra Fuzzy Set was

added and any rule including that fuzzy set was assessed as unsuitable.


64

WP3: MORPH

The optimization of Fuzzy Rules of brown trout spawning fuzzy model achieved the

values 0.40 ± 0.018 of the TSS performance criterion. The training presented two rules

with no cases to be trained, and Expert-knowledge was used in their determination (

Table 7). The spawning model was not validated with independent data, and since no

sub-sampling was carried out, the calculated TSS assesses the performance over the

entire database. The rules including the extra Fuzzy Sets were assessed in any case

as unsuitable (

Table 7).


65

WP3: MORPH

Probabilistic Neural Networks

The optimization of the PNN yielded acceptable results. The model resulted slightly

overpredictive, Sn=0.82 and Sp=0.58 (Table 8).

Table 8 Summary TSS, accuracy, sensitivity and specificity values achieved in the optimization

of the smoothing parameters and the values of the corresponding parameters.

Smoothing parameter (σ) TSS Accuracy Sensitivity (Sn) Specificity (Sp)

0.117 0.39 0.59 0.82 0.58

The assessment carried out using a single smoothing parameter did not produce any

area with a suitability higher than 0.8 (¡Error! No se encuentra el origen de la referencia.). However, the frequency analysis showed clear crossed tendencies. The

maximum frequency in Availability was present in the interval of suitability 0.2-0.4,

which means absence, whereas the maximum frequency in Use was present in the

interval of suitability 0.6-0.8, which means presence. The frequency analysis of Use

showed an increasing trend from the lower (0-0.2) to the higher values (0.6-0.8),

whereas the frequency analysis of Availability showed a decreasing trend from the

lower interval (0.2-0.4) to the higher one (0.6-0.8). In addition, 57% of the observed

trout locations were classified with suitability higher than 0.5 (i.e. they were classified

as presence).

The synthetic database was assessed to check model performance (Figure 33). It showed close

to no suitability over fine substrates (i.e from substrate index=0 to substrate index=2) whereas

the brown trout suitability increased widely with the substrate index (Figure 33).


66

WP3: MORPH

Figure 32. LEFT: comparison of the habitat assessment of the flow surveyed (Q=0.89 m³/s) during the biological validation using the developed PNN. Red areas mean unsuitable locations and dark green mean areas with perfect suitability for adult trout. Dots are adult trout positions

at the moment of the validation survey. RIGHT: frequency histogram of the assessment of PNN. Availability (assessment of the entire simulated reach) is represented by black bars and Use

(assessment of fish locations) by grey bars. Notice that the assessment carried out through the PNN did not provide the maximum suitability (range 0.8-1).

The substrate interval between two and five showed indistinct habitat preferences and

the PNN classified as presence most of the training patterns. This interval did not

present any training pattern with velocity higher than 1 m/s. Up to this point, the brown

trout preference for deeper areas increased as the substrate index increased until the

maximum over bedrock, where suitable areas were calculated as deeper than 1 m.

There were no training patterns where the sum of depth and velocity exceed a value of

two; then, this value was considered the threshold of the range of applicability of the

generated model. The assessment of the synthetic database did not achieve the

maximum possible suitability (i.e. one) within the acceptable range of application of the

PNN, and this should condition the interest about its application. Notice that the PNN


67

WP3: MORPH

connects every assessed condition with each of the cases comprised in the training

database, Thus, the training patterns assessed through the leave-one-out procedure

are also plotted in Figure 33.

Figure 33. Slices of the assessment of the synthetic database based on substrate types for the optimization carried out with a single sigma. Red colour means unsuitable, yellow the highest suitability. Dots correspond to the training dataset. Colours represent the assessment through

the leave-one-out optimization. Dark green: presences classified as presence; light grey: presences classified as absences; light green: absences classified as presences; dark grey:

absences classified as absences.


68

WP3: MORPH

3.1.2.2 Barbus haasi in the Mijares river and Siurana River

The sub-sampled database showed similar distribution but lower frequency in the three

surveyed variables. Therefore, the application of statistical tests was considered

unnecessary (Figure 34).

Figure 34. Frequency distribution of the original Availability database and of the Sub-sample database. Notice that distributions are quite similar, but that frequencies are lower in the sub-

sample case.

The number and shape of the fuzzy sets are fundamental issues determining the

accuracy of a fuzzy logic inference system. Therefore, an experimental approach was

carried out merging the Expert-knowledge and the Data driven methods. Thus, in

addition to the Expert-knowledge approach, the fuzzy rules were trained by means of

the Data-driven approach but the considered fuzzy sets were derived from both

methods.

The discretization of input variables provided three fuzzy sets in each case when based

on Habitat Suitability Curves Category II ½ and Category III. Additionally, the Data-

driven fuzzy sets were produced following two approaches in order to assess the


69

WP3: MORPH

impact of variable discretization on model performance. The first one tried to force the

entropy-based optimization to provide the same amount of fuzzy sets present in the

HSC-based cases, providing three fuzzy sets for velocity and depth, whereas in the two

fuzzy sets were obtained for substrate. The unconstrained entropy-based optimization

increased the amount of fuzzy sets for velocity and depth to five, but the substrate

discretization remained in two sets (Figure 35). The Cohen's Kappa values achieved

during the optimization of the fuzzy rules were 0.26 ± 0.05, 0.41 ± 0.04, 0.36 ± 0.04

and 0.51 ± 0.04 for the Category II ½-based, the Category III-based, the constrained

entropy-based and the unconstrained entropy-based methods, respectively. The

number of uncovered rules was relatively high in the four cases.

The Category II ½-based case presented 60% of uncovered rules. The Category III-

based case presented 52% of uncovered rules, whereas the entropy-based presented

50% of uncovered rules (Table 9). The unconstrained entropy-based optimization

showed the lowest amount of uncovered rules (44%). Nevertheless, it must be stressed

that the Table 9 should be viewed just as a rule summary, not as a comparison, since

discretization could strongly differ, thus changing the meaning of each category.


70

WP3: MORPH


fuzzy sets for Barbus haasi. Data-driven fuzzy sets obtained from the Shannon-Waver entropy based optimization with similar amount than the HSC-based and Data-driven fuzzy sets

unconstrained.


71

WP3: MORPH

Table 9. Summary of the Fuzzy Rules for the generated models. Asterisk (*) means uncovered rule and the proper output was based on the corresponding rule in the Expert-knowledge approach, if it exists of in the Data-driven model with different substrate. Suitability in 5 categories corresponds to Very Low (VL), Low (L), Medium (M), High (H) and Very High (VH).

Data-driven approach Expert-knowledge approachRule Velocity Depth Substrate Cat. II ½ Cat. III Entropy Ent. Uncons. Cat. II ½ Cat. III1 VL VL L - - - L - - 2 L VL L - - - L* - - 3 M VL L - - - L* - - 4 H VL L - - - L* - - 5 VH VL L - - - L* - - 6 VL L L - - - H - - 7 L L L L* L H L L L 8 M L L L* L* L* L* L L 9 H L L L* L* L* L* L L

10 VH L L - - - L* - - 11 VL M L - - - H - - 12 L M L H H H H* H H 13 M M L H* H* H* H* M H 14 H M L L* L* L* H* L L 15 VH M L - - - L* - - 16 VL H L - - - H - - 17 L H L L* L* L* H L L 18 M H L L* L* L* H* L L 19 H H L L* L* L* H* L L 20 VH H L - - - L* - - 21 VL VH L - - - L* - - 22 L VH L - - - L* - - 23 M VH L - - - L* - - 24 H VH L - - - L* - - 25 VH VH L - - - L* - - 26 VL L M - - - L - - 27 L L M L* L - L L L 28 M L M L L - L L L 29 H L M L* L - L L L 30 VH L M - - - L - - 31 VL M M - - - H - - 32 L M M H H - H H H 33 M M M L H - L H M 34 H M M L L* - L L L 35 VH M M - - - L - - 36 VL H M - - - H - - 37 L H M H H - H L L 38 M H M L L* - H L L 39 H H M L* L* - H L L 40 VH H M - - - L - - 41 VL VL H - - - H - - 42 L VL H - - - H - - 43 M VL H - - - H - - 44 H VL H - - - H - - 45 VH VL H - - - L - - 46 VL L H - - - L - - 47 L L H L* L L H L L 48 M L H L* L* L L* L L 49 H L H L* L L L* L L


72

WP3: MORPH

Table 9. (continues)

Data-driven approach Expert-knowledge approach

Rule Velocity Depth Substrate Cat. II ½ Cat. III Entropy Ent. Uncons. Cat. II ½ Cat. III

50 VH L H - - - L* - - 51 VL M H - - - L - - 52 L M H H H H L* H H 53 M M H H H H L* M M 54 H M H L L* H L* L L 55 VH M H - - - L* - - 56 VL H H - - - H - - 57 L H H L L L L L L 58 M H H L* L* L* L* L L 59 H H H L* L* L* L* L L 60 VH H H - - - L* - - 61 VL VH H - - - H - - 62 L VH H - - - H* - - 63 M VH H - - - H* - - 64 H VH H - - - H* - - 65 VH VH H - - - L* - -

The assessment of the training database with the developed models showed certain

disparity. The Data-driven approach showed underpredictive when trained with fuzzy

sets based on Category II ½ and the entropy-based with similar amount of fuzzy sets,

whereas the Category III-based and the unconstrained entropy-based showed

overpredictive. The Expert knowledge approach showed overpredictive in both cases,

but the Category II ½ showed extremely overpredictive, with a specificity close to zero

(Table 10). The model with the best overall results was the one developed with the

unconstrained entropy-based, with a TSS of 0.39, the highest (Table 10).

Table 10. Assessment of the training database through the models. The sensitivity (sn), the

specificity (sp) and the true skill statistic (TSS) were calculated.

Data‐driven approach Expert‐knowledge approach

Cat. II ½ Cat. III Entropy Entropy

Unconstrained Cat. II ½ Cat. III

sn 0.55 0.82 0.57 0.79 0.96 0.81 sp 0.66 0.50 0.75 0.59 0.08 0.50 TSS 0.21 0.32 0.33 0.39 0.04 0.32


73

WP3: MORPH


Figure 36. Category II ½ Habitat Suitability Curves and their corresponding fuzzy sets for adult,

juvenile and fry brown trout based on Bovee (1995 unpublished).

The Expert-knowledge fuzzy models were based on the Category II curves developed

by Bovee (1995, unpublished) providing a simple discretization in the adult brown trout

case. Velocity was discretized in two Fuzzy Sets, depth in three and substrate also in

two (Figure 36). The juveniles presented discretization similar to adults. Velocity was

discretized in two Fuzzy Sets, depth in three and substrate also in two (Figure 36).

Moreover, the ranges of the developed Fuzzy Sets presented high similarity with those

generated for adults. Fry presented the most detailed discretization: velocity was

discretized in two Fuzzy Sets, depth in four Fuzzy Sets (both based on the curves


74

WP3: MORPH

developed in the Jucar River Basin District), finally, substrate (based on American

curves (Bovee, 1978; Raleigh, 1984)), presented three Fuzzy Sets (Figure 36).

The Expert-knowledge Fuzzy Rules generated are summarized in Table 11. Due to the

low discretization, the amount of generated rules was the lowest among the generated

models.

Table 11. Summary of the Fuzzy Rules for the generated fuzzy models including a model for

adult, juvenile and fry brown trout based on Bovee (1995 unpublished). Suitability in 5

categories corresponds to Very Low (VL), Low (L), Medium (M), High (H) and Very High (VH).

Rule Nº Velocity (m/s)

Depth (m)

Substrate (s) EK Adult EK

Juvenile EK Fly

1 L VL L L 2 L L L L L H 3 L M L H H M 4 L H L M M L 5 H VL L L 6 H L L L L L 7 H M L L L L 8 H H L L L L 9 L VL M L 10 L L M H 11 L M M H 12 L H M L 13 H VL M L 14 H L M L 15 H M M L 16 H H M L 17 L VL H L 18 L L H L L H 19 L M H H H M 20 L H H H H L 21 H VL H L 22 H L H L L L 23 H M H L L L 24 H H H L L L


75

WP3: MORPH

3.1.2.4 Macroinvertebrates

The Data-driven approach discretized velocity and depth in three Fuzzy Sets but

substrate in two Fuzzy Sets, achieving Shannon–Weaver entropy values of 0.97, 0.95

and 0.84, respectively (Figure 37). However, an extra fuzzy set was added to include

the bedrock substrate separately. In the optimization of the output fuzzy sets for

Ephemeroptera, Trichoptera and EPTs the Shannon–Weaver entropy values of 0.5, 0.5

and 0.63 were obtained. Surprisingly, the algorithm provided similar discretization but

with different meaning, thus the original densities were different. In both cases, three

fuzzy sets were provided. Finally, an extra fuzzy set was added to cover the null

density cases (Figure 38).


optimization. Substrate was discretized in two Fuzzy Sets, but an extra set was included to cover the bedrock substrate.


76

WP3: MORPH

Figure 38. A) Data-driven fuzzy sets obtained from the Shannon-Waver entropy-based

optimization for the standardized density in Ephemeroptera, Trichoptera, and EPT. The Data-driven approach provided exactly the same discretization on the three cases, but differed on the

represented density values B) Presence/Absence output fuzzy sets for Plecoptera.

The Data-driven optimization of the fuzzy rules achieved the values 0.41 ± 0.04, 0.47 ±

0.07, 0.44 ± 0.04 and 0.41 ± 0.07 of the Cohen's Kappa for Ephemeroptera,

Plecoptera, Trichoptera and the EPT aggregation, respectively. The training for all of

them presented 8 uncovered rules and Expert-knowledge and the analysis of the

trained rules was used in their determination (Table 12).


77

WP3: MORPH

Table 12 Summary of the Fuzzy Rules for the macroinvertebrates models. Asterisk (*)

means uncovered rule (no data for training); in such cases, the outputs were

determined through authors’ consensus, that is, expert-knowledge approach. Suitability

in 5 categories corresponds to Very Low (VL), Low (L), Medium (M), High (H) and Very

High (VH).

Velocity Depth Substrate Ephemeroptera Plecoptera Trichoptera EPT L L L L L VL L M L L L L L L H L L VL* L* VL* VL* L M L L H L L M M L L L M M H M L M* L* L* L* L H L L L L L M H L VL L VL VL H H L L* L* VL* L* L L M VL L L L M L M VL L L VL H L M L H L L L M M L H H L M M M L H L L H M M H L L M L H M L H VL L M H M H H M M H H M M L VL M L L H L L VL L M L H VL* L* L* VL* H L H L* L* L* L* L M H L* L* M* L* M M H VL L VL VL H M H L L M M L H H VL* L* VL* L* M H H L L L L H H H L* L* VL* L*


78

WP3: MORPH

3.2 Habitat suitability modelling for riparian vegetation

The riparian vegetation at the Cabriel-Rabo del Batán study reach is described and

species are grouped in relation to their positional patterns (distance and elevation

to/above thalweg) and soil preferences (texture, organic matter and moisture content).

Table 13. Summary of the plant species surveyed at the Cabriel-Rabo del Batán study reach.

The scientific and common names of each species are indicated, as well as the code assigned

to each species. In the last column appears the number of specimens surveyed.

Scientific name Common name Code Family Sampling sizeSalix alba L. White willow SL Salicaceae 34 Salix eleagnos Scop. Rosemary willow SE Salicaceae 75 Salix purpurea L. Purple willow SP Salicaceae 76 Salix triandra L. Almond willow ST Salicaceae 6 Populus alba L. White poplar PA Salicaceae 293 Populus nigra L. Black poplar PN Salicaceae 116 Fraxinus angustifolia Vahl Narrow-leaved ash FA Oleaceae 61 Crataegus monogyna Jacq. Common hawthorn CR Rosaceae 16 Cornus sanguinea L. Common dogwood CS Cornaceae 8 Pinus spp. Pine PC Pinaceae 154

Total Sampling Size 839

3.2.1 Vegetation description

Twenty transects were surveyed at Cabriel-Rabo del Batán reach (Figure 7 and Figure

10). Their mean length was 69 m and their range from 30 to 106 m. Plant richness was

10 (five tree species and five shrub species) and 839 plants were recorded (Table 13).

The total length of transects surveyed was 1380 m and the length of plant cover 1934

m, which depicted a total cover of 140%, meaning that several strata or vegetation

layers were overlapped in parts. More specifically, 32% of the total cover was

composed of shrubs and 108% of trees. As can be seen in Figure 39, the most

abundant species were Populus alba (PA), Pinus spp. (PC) and Populus nigra (PN),

with 35%, 19% and 14% of the total number of stems, respectively, indicating that this

site has a mature riparian forest. The percentages of stems and cover were similar for


79

WP3: MORPH

all species, except for Salix alba (SL) and PA. In the case of SL, 4% was the

percentage of stems and 6% the percentage of cover. This revealed that this species

presented few and old individuals. In contrast, PA exhibited larger percentage of stems

than cover (35 respect to 30%), which meant that despite being the most abundant tree

species, many of its individuals were small.

Figure 39. Descriptors of the riparian vegetation species sampled by transects at the Cabriel-

Rabo del Batán study reach. LEFT: abundance (percentage of stems of each species respect to the total number of plants sampled), cover (percentage of cover of each species respect to the

total cover value in the site). RIGHT: density (meter of cover of each species per meter of transect, in bars; number of stems of each species per meter of transect, in solid line). Species appear in ascendant ranking based on the Huber estimator of elevation above thalweg. Species abbreviations: SP, Salix purpurea L.; ST, Salix triandra L.; CS, Cornus sanguinea L.; SL, Salix alba L.; FA, Fraxinus angustifolia Vahl; PN, Populus nigra L.; SE, Salix eleagnos Scop.; CR,

Crataegus monogyna Jacq.; PA, Populus alba L.; PC, Pinus spp.

In relation to the density, PA showed the highest values (for each m of transect

sampled, 0.4 m were covered by this species and additionally, this corresponded to 0.2

stems, i.e., in every ten m of transect two stem of PA were present with a total cover of

four m). Additionally, both PC and PN had approximately 0.1 stems, with 0.25 and 0.2

m of cover per m of transect, respectively. Salix purpurea (SP), Salix eleagnos (SE),

Fraxinus angustifolia (FA) and SL had similar density in the study reach (0.1 m of cover

per m of transect). Nevertheless, FA and SL showed lower values of stems/m than SP

and SE, probably because the first two species are trees and the second two shrubs.

Finally, Salix triandra (ST), Cornus sanguinea (CS) and Crataegus monogyna (CR)


80

WP3: MORPH

were the species with smaller abundance, cover and density in the study reach. ST and

CR had similar cover values, but ST had smaller number of stems.

3.2.2 Positional patterns and riparian guilds

Differences in the positional pattern of the riparian species were detected at the Cabriel

River. In relation to distance to thalweg (Figure 40 - left), the robust test of Welch gave

a value of FWe = 51.14 (p-value = 0), therefore we rejected the null hypothesis of equal

0.1-trimmed means. The test for multiple comparisons showed four homogeneous

groups of species in relation to their distance to thalweg, considering 10% as trimming

percentage and 0.05 as significance level. These were: [CS, ST, SL, SP] close to the

active channel and with a narrow distribution range, then [ST, SL, FA, SP], followed

by [SP, PN, CR] in intermediate positions and finally [PA, CR, SE, PC] further away

from the channel and showing a much wider distribution range. However,

considering only the species with larger sample size and tr = 20 % and α = 0.10,

the groups obtained were the following: [SL, FA, SP, PN], [SL, FA, SP], [SL, SP,

PN] and [PA, SE, PC]. The species FA and PN did not appear together.

Figure 40. Boxplots of the distribution (left plot, distance to thalweg; right plot, elevation above

thalweg) of the woody riparian species sampled at the Cabriel-Rabo del Batán study reach. Species are in ascendant order according to their median values. Red dots represent the Huber

estimator.

In relation to elevation above thalweg, the robust test of Welch gave a value of FWe =

38.35 (p-value = 0), therefore, we rejected the null hypothesis of equal 0.1-trimmed

means. Considering only the species with larger sampling size, a trimming percentage


81

WP3: MORPH

of 20% and 0.10 as significance level, the test for multiple comparisons showed four

homogeneous groups of species, from lower to higher elevation: [SP, SL], [SL, FA,

PN], [PN, PA, SE], and finally, [PC].

3.2.3 Dominance curves

The dominance curves revealed the locations where each species was most abundant,

and therefore, suggested their highest preference for those places. In relation to

distance to thalweg (Figure 41), CS had its highest abundance values in the range 0-5

m from thalweg; ST, SL, SP, FA and PN in the range 5-10 m; SE and PA in the range

10-15 m and PC in the range 20-25 m. CR appeared under tree cover in a wide

distance range from the thalweg, but was specially abundant in the range 10-30 m.

However, CS had a more reduce preference for places at less than 10 m from thalweg.

Figure 41. Dominance curves of the riparian species respect to the distance to thalweg at the

Cabriel-Rabo del Batán study reach. Left, absolute frequency. Right, normalized giving a value of 1 to the peak frequency of each species.

FA showed a disperse distribution along the entire riparian zone, but it was particularly

abundant in the range of 5-10 m from thalweg. Along with FA, four species more (PA,

PC, SE and PN) occupied the entire distance gradient. PA was abundant from 0 to 25

m and had a second peak in the range 55-60 m due to the presence of a secondary

channel only flooded in highly rainy periods; PN and SP were also present (only a few


82

WP3: MORPH

plants) near to this secondary small channel. SE showed a second peak at the range

30-35 m, in an area between both channels with a coarse substrate.

In relation to elevation above thalweg (Figure 42), the vast majority of plants were

located from 0 to 3 m. SP and PN, two of the most pioneer riparian species, started at

the lowest locations (0-0.5 m above thalweg). The highest abundance of SP was

registered in the range 0.5-1.5 m, dominating in the range 0.5-1 m. ST, SL, PN, FA and

CS had their largest abundance values in the range 1-1.5 m, being at that band PA,

followed by PN and FA the species dominating in number. PA showed the widest

distribution along the elevation gradient, but was dominant in the range 1.5-2 m,

accompanied by PN and PC. In the next range (2-2.5 m), SE and CR were the species

showing their highest values for their own species, but the dominant species were PA,

SE and PC. Finally, PC dominated and presented its highest abundance values in the

range 2.5-3 m, along with PA and PC.

Figure 42. Dominance curves of the riparian species respect to the elevation above thalweg at the Cabriel-Rabo del Batán study reach. Left, absolute frequency. Right, normalized giving a

value of 1 to the peak frequency of each species.


83

WP3: MORPH

3.2.4 Relationship with soil characteristics

Twelve soil samples were analysed from the study reach. They differed in their grain

size distribution (Figure 43).

Figure 43. Grain size distribution of the soil samples of the Cabriel-Rabo del Batán study reach.

Samples 2, 5, 7, 10 and 12 were classified as gravel soils (gravel content > 68%);

sample 1 as gravel-sandy soil (50% sand and 30% gravel); samples 4 and 9 as sandy

soils (sand content > 80), and samples 3, 6, 8 and 11 as silty-sandy soils (20-50%

sand, 30-60% silt). No clayey soils were found in the reach. In all soils the organic

matter was always lower than 5% and the moisture in the range 3-60%. Soils with

larger content of organic matter also had larger moisture (Figure 44).


84

WP3: MORPH

Figure 44. Percentages of organic matter and moisture content in the twelve soil samples

analyzed in the Cabriel-Rabo del Batán study reach.

In relation to the soil preferences (Figure 15 and Figure 16), although groups of species

can be identified according to their association to specific soil characteristics, most

species showed a clear overlap, except PC, which appeared in most of the cases

clearly differentiated from the rest of species. For soil comparisons, the tests were

performed using 10% of trimming and 0.05 as significance level. The species with

small sampling size (ST, CS and CR) were not considered, although they were kept in

the graphs.

Soil gravel content was one of the variables structuring plant communities, PC being

found in coarser soils, whereas SP, SL and PN, appeared in soils with lower gravel

content. The rest of the species showed intermediate values, but also a high variation

in their values. SL and SP had a standard deviation of 26 and 29% respectively,

whereas the rest of species had values in the range 33-37%.


85

WP3: MORPH

Figure 45. Barplots with standard errors for the percentage of gravel, sand, silt and clay for each

species at the Cabriel-Rabo del Batán study reach. Species appear in ascendant order according to the ranking defined by the Huber estimator in the elevation gradient above

thalweg.

On the contrary, PC presented the smallest percentage of sand (25 %) and showed

significant differences with the rest of species, which had mean values in the range 32-

39%. PN had the smallest standard deviation (17%) and SE the highest (24%). In

relation to the silt percentage, significant differences were found only between PN and

SE. In fact, SE had the smallest values (19% of silt). Looking at the entire set of

species, two groups could be defined, a first group with lower silt content formed by

SP-SE-PA-PC (range of mean values: 19-30%; range of median values: 10-14%) and

a second group with higher silt content by SL-FA-PN (range of mean values: 30-35%;

median value 51%). The standard deviation was in the range 22-26%.

In general, the percentage of clay was small in all soil samples (16% was the maximum

value obtained; soil 8). Visually, two groups could be distinguished: SP-SE-PA-PC with

lower content of clay (range of mean values: 7-9%, and of median values: 4-6%), and

SL-FA-PN with higher content of clay (range of mean values: 9-10%; median: 14%),

however significant differences were found only between PN and SE. The standard

deviation was 5-6%.


86

WP3: MORPH

The percentage of organic matter content was high in SP-FA-SE (3-3.3%),

intermediate in SL-PN-PA (2.5-2.7%) and low in PC (1.9%). Significant differences

were found between PC and the rest of species and also between FA and PN, SE and

PN, and SE and PA. The standard deviation was in the range 0.8-1.4%.

Figure 46. Barplots with standard error for the percentage of organic matter and moisture content for each species at the Cabriel-Rabo del Batán study reach. Species appear in

ascendant order according to the ranking defined by the Huber estimator in the elevation gradient above thalweg.

The moisture content showed similar groups as the texture. Using the same test

parameters (10% trimming and 0.05 significance level), three groups could be defined:

SP-FA-SE-PA on soils with higher moisture content (range: 21-31%) respect to PN-SL

(14 and 16%, respectively). PC obtained the lower moisture value (10%). SE had the

largest standard deviation (24%).

Finally, a few lines for the species with lower abundance. CS and ST showed similar

patterns, close to the water’s edge, in soils with fine texture, high moisture content and

organic matter. CR showed a completely different distribution, more similar to PN, PA

and PC, i.e., higher grain size (basically gravel and sand).

3.2.5 Multivariate interpretation

A Principal Component Analysis (hereafter, PCA) and a Cluster Analysis were

performed in order to analyze the joint variation of all physical variables (morphology

and soil characteristics) and to obtain groups of species with a similar response

(guilds). This interaction between species and physical variables is illustrated in Figure

47, where the first two axes of the PCA explained 66.1% of the total variance. The first

axis was defined positively by the distance to thalweg and gravel content, and


87

WP3: MORPH

negatively by the clay and silt content. The second axis was defined positively by the

organic matter content and negatively by the elevation above thalweg.

The classification analysis performed on the coordinates of the first three PCA

components explained 85.8% of the total variability (Figure 48). In this sense, ST and

CS were located in areas of low distance and elevation above thalweg (i.e. bank zone)

with high content of sand and organic matter. In the centre of the plot appears a group

of riparian species that could be divided into two groups within the floodplain. A first

band would be composed of SP and FA, close to the water’s edge and with high

organic content and soil moisture. They would be followed by a second band formed by

PA and CR, farther from thalweg, and on coarser substrate. Finally, as a differentiated

guild appear PC, in areas of high elevation and coarse substrate (mainly gravel) and

with low soil moisture and organic matter content, revealing a transitional change from

riparian to terrestrial conditions. SL and PN could be defined as the species marking

the transition between the aforementioned guilds within the floodplain. SE would be

located near PA-CR but in coarser areas (with low content of silt and clay) and high soil

moisture and organic matter content.


88

WP3: MORPH

Figure 47. PCA diagram for the Cabriel-Rabo del Batán study reach. Riparian woody species (dots in black colour) are abbreviated as: ST, Salix triandra; CS, Cornus sanguinea; SL, Salix alba; SP, Salix purpurea; FA, Fraxinus angustifolia; PN, Populus nigra; PA, Populus alba; CR,

Crataegus monogyna; SE, Salix eleagnos; PC, Pinus spp.

In short, considering the information given by the three analyses (PCA, cluster and

ordicluster), seven riparian guilds could be defined at the CRA, as follow: [CS-ST] [SP-

FA], [SL], [PN], [PA-CR], [SE], [PC]. Grouping them into a smaller number of guilds,

these would be defined as: [CS-ST], [SP-FA-SL-PN], [PA-CR-SE], [PC].

[CS-ST] would be located close to the active channel in distance and elevation (i.e.

bank zone), consequently it could be defined as ‘highly tolerant to inundation’, typical of

positions close to the water’s edge, able to withstand the waterlooging and burial by

sediments. These species had their highest abundances at distances smaller than 5

and 10 m from thalweg, respectively, which corresponded with fine-textured soils with

high moisture and organic matter content.


89

WP3: MORPH

Figure 48. Left: hierarchical cluster dendrogram (average clustering with Euclidean distance)

based on the first three PCA scores at the Cabriel-Rabo del Batán. The vertical axis shows the level of fusion and is based on the dissimilarities between species. Right: clustering results

overlaid onto an ordination diagram based on the proximity matrix. Cluster centroids are connected to each other with line segments similarly as in the original cluster dendrogram.

[SL-SP-FA- PN] would be located in an intermediate position, covering the banks and

part of the most dynamic floodplain; therefore this guild could be defined as ‘high-

intermediate tolerant to inundation’. They appeared mostly in the range 5-10 m. The

highest abundance of SP took place in the range 0.5-1.5 m above thalweg, dominating

in the range 0.5-1 m. SL, PN and FA had their largest abundance values in the range

1-1.5 m. These species were located in soils with low gravel and high silt content (in

comparison with the rest of species). In addition, SP and FA were located in areas with

higher moisture and organic matter content than SL and PN. Therefore, PN could be

defined as the species marking the transition between this guild and the next one within

the floodplain.

[SE-PA-CR] would be located further from the channel, followed by [PC]. In this sense,

this guild could be defined as ‘transitional between floodplain and terrestrial’. SE and

PA appeared mostly at distances of 10-15 m, CR at 10-30 m and PC in the range 20-

25. PA showed the widest distribution along the elevation gradient, but it was dominant

in the range 1.5-2 m above thalweg, followed by PC in the range 2.5-3 m. This guild

would correspond mostly with gravel soils, with low percentages of sand, silt, clay and

organic matter content, except for SE that appeared in areas with higher moisture and

organic matter content.


90

WP3: MORPH

4 DISCUSION



4.1.1.1 Salmo trutta in the Cabriel River

The developing of the Cat.II ½ Habitat Suitability Curves (HSCs) followed usual

standards (Bovee, 1986). These curves differ from those in the literature in most of the

cases. The adult case (Figure 17) presents the greatest number of possible

comparisons. In general, the Cat.II ½ HSCs for adults generated in the present study

showed the highest suitability for lower values (velocity and depth) than some of the

most relevant studies about adult brown trout which include Cat.I, Cat.II and Cat.III

HSCs. (Bovee, 1978; Raleigh, 1984; Hayes and Jowett, 1994; Heggenes, 1996; Ayllón

et al., 2010). Only the velocity was similar to that from Bovee and Raleigh (1984) and

the depth compared with the Ayllón's (2010) and the corresponding of Hayes and

Jowett (1994). The latter study presented both, Cat.II and Cat.III HSCs. The substrate

showed similar suitability over coarse substrates than those from literature, and did not

present remarkable differences. In general those differences with previous studies

could be mainly related to several factors such as differences in fish size or river types

(Jowett and Davey, 2007) or in the selected sampling protocol (Heggenes et al., 1990).

The observed differences could be due to the selected modelling approach. A recent

study corroborated that each modelling technique could be focused on different

aspects of the training database, even predicting different habitat suitability (Fukuda et

al., 2013). However, habitat selection patterns of brown trout are well established in

broad terms (Ayllón et al., 2010). Thus, concerning near-natural rivers, trout has been

reported to dwell preferably in relatively deep pools occupying near-bottom locations

with slow flow and medium-to-coarse substrate (Heggenes, 1996; Moyle, 2002;

Armstrong et al., 2003; Ayllón et al., 2010), and our results were certainly in

accordance with these statements. This agreement was specially remarkable regarding

the Cat.III HSCs.


91

WP3: MORPH

The generated Cat.II ½ HSCs for juvenile (Figure 18) had less possibilities of

comparison. The velocity curve presented a pattern similar to the one shown by

Raleigh (1984) and Bovee (1978), but Ayllón et al. (2010) differs strongly, as it on

average suggests the highest suitability to faster flow. The depth curve presented the

highest suitability for shallower areas than that by Raleigh, whereas the one by Bovee

covers entirely the generated Cat.II ½ HSC, thus becoming wider. Finally the Ayllón

curves on average presents the highest suitability for deeper areas. The substrate

suitability presents a pattern similar to those from the literature, with the highest

suitability over coarse medium-to-substrates.

The Cat.II ½ curves for fry (Figure 19) presented the highest suitability for lower

velocities than the the curves by Bovee (1978), Raleigh (1984) and Ayllón et al. (2010),

whereas the depth was similar to these shown by Ayllón, lower than Raleigh and like

the previous case, the curve by Bovee covers entirely ours. A great discrepancy

appears with the substrate suitability: the aforementioned studies showed the highest

suitability to be similar to that for elder life stages, over medium-to-coarse substrate,

whereas our curves showed suitability to peak on silt. This discrepancy could condition

strongly the results derived from our model. The Cat.III curves developed in the present

study showed relatively major similarities with those from literature, in contrast to the

Cat.II ½ HSCs.

The adult case presented the highest suitability in faster areas than Bovee, Raleigh

and Ayllón, but similar to Hayes and Jowett (Figure 17). The Cat.III HSC for depth

developed for adults had the same pattern of similarities; the highest suitability was

presented for deeper areas in comparison with all of the aforementioned studies,

except Hayes and Jowett and Ayllón, which presented a constant and high suitability,

regardless the depth.

The juvenile case (Figure 18) presented the Cat.III curves with the largest number of

coincidences with those from the literature. The generated Cat.III were different from

the results provided by Raleigh and Bovee, presenting the highest suitability for faster

flows, but similar to those by Ayllón. In addition, the Cat.III curve for depth showed

clear similarities with Bovee, Raleigh and Ayllón. The substrate did not differ

substantially, and the highest suitability was just slightly displaced one Cat.toward finer


92

WP3: MORPH

substrates in comparison with the Cat.II ½ counterpart. However, an appreciable

suitability over coarse substrates (cobble/boulder) remained in accordance with the

literature.

The fry case (Figure 19) presented a pattern similar to that reported by Ayllón and

Raleigh for velocity, whereas the Bovee curve involves the fry Cat.III curve for the

variable velocity. The depth curve presented a pattern similar to those by Bovee and

Raleigh, but the highest suitability for deeper areas than Ayllón. The adjustments

carried out for the substrate did not generate strong changes in the Cat.III curve if

compared with the Cat.II ½, thus the curve presented the highest suitability also over

silt, contrasting with the literature, which presents the optimum over medium-to-coarse

substrata.

Generally, the generated Cat.III curves presented all the inconveniences compiled by

Payne (2009). The application of the forage ratio produced a displacement of the curve

toward the higher values of velocity and depth, whereas the substrate curves remained

almost constant. In the past this displacement discouraged their application (Bovee,

1996). However, upholding their consideration, some authors have demonstrated that

habitat availability affects habitat use, and even habitat selection by brown trout

(Heggenes, 1991; Rincon and Lobon-Cervia, 1993; Grossman and De Sostoa, 1994b),

thus recommending certain correction based on habitat availability. However, the

overcorrection produced through the application of the forage ration is still unsolved

(Payne, 2009). Hayes and Jowett (1994) pointed out that the ratio of used and

available habitat is particularly sensitive to extreme values, and it does not account for

habitat that was not available at the time or place of sampling, but due to the limited

amount of data the curved were not trimmed in any case. The Cat.III Habitat Suitability

Curves agree with previous studies where the Cat.III HSCs for adults move to deeper

and faster areas compared with Cat.II's (Bovee and Zuboy, 1988). In contrast, the

curves for fry presented similar behaviour, whereas the literature suggests it should

move to lowed depth and velocity. In general, according to the results achieved in the

validation of the fuzzy models, the performance of the Cat.III curves is expected to be

better than the Cat.II ½ for brown trout in the Cabriel River.


93

WP3: MORPH

4.1.1.2 Barbus haasi in the Mijares River and Siurana River

The survey was considered acceptable as about 100 fish locations were obtained, and

the surveying methodology provided a relatively high prevalence (0.21). The developed

habitat suitability curves showed in both cases, Cat.II ½ and Cat.III, a decrease on the

habitat suitability as the velocity increases from null velocity and tailed off towards 1.5

m/s. The depth maximum depended on the observed curve. The Cat.II ½ showed a

maximum around 0.75 m, whereas the Cat.III HSC peaked at 1.5 m. In both cases, the

deformation of the curve through the application of the forage ration was not

considered dramatic. Then, although Cat.III curves have been demoted (Bovee, 1996),

our results suggest that both are valid. However, contrasting with velocity and depth,

the substrate curve showed dramatically all the drawbacks compiled by Payne (2009);

Cat.II ½ showed the maximum suitability for gravels, the most abundant substrate, but

the Cat.III HSC showed the opposite, a maximum for silt and macrophites, being the

minimum suitability for gravel. (Figure 20). This contrasting result could discourage the

selection of Cat.III HSC, since large deformation is considered inappropriate. However,

the Cat.III HSCs agreed with the results presented by Grossman et al. and Grossman

and De Sostoa (1987; 1994a), according to which Barbus haasi showed preference for

low velocity, relatively high depth and fine substrate. Nevertheless, caution should be

taken when extrapolating data from these studies, as they are based on the monitoring

of just 30 m. Therefore, we suggest the Cat.II ½ curves to be used for further analysis.


The development of the Cat.II HSCs followed the usual standards (Bovee, 1986).

Additionally, these HSCs did not differ substantially from those from literature in most of

the cases.

As mentioned before, the adults (Figure 21) provide the greatest number of possible

comparisons. In general the unpublished Cat.II curves for adults generated by Bovee

showed the optimal for higher values than the Cat.II ½ developed in the Cabriel river

but roughly agreed the most relevant studies on brown trout microhabitat suitability in

the literature (Bovee, 1978; Raleigh, 1984; Hayes and Jowett, 1994; Heggenes, 1996;

Ayllón et al., 2010). However, the unpublished HSCs did not consider substrate.

Therefore, we followed the most conservative approach and the Cabriel HSCs for


94

WP3: MORPH

substrate were selected, which presented similar suitability over medium-to-coarse

substrates than those from literature.

The generated Cat.II ½ curves for juvenile (Figure 21) present less possibilities for

comparison. The velocity curve was wider than those presented by Raleigh and Bovee,

with the optimum in a faster velocity (0.5 m/s), and the depth curve was narrower that

the one by Bovee, which certainly involves the entire curve. Additionally. although this

HSC was similar to that by Raleigh, it presented diverging maxima, lower in the

unpublished curve. On the other hand, the unpublished curves presented similar

patterns than those by Ayllón, with a wide range for velocity with the optimum around

0.5 m/s, and a narrow range for depth, with the optimum around 0.5 m. The substrate

presented similar suitability distribution than the curves from the literature, with the

maximum over coarse substrate.

The Cat.II ½ curves for fry (Figure 21) were based on those developed in the Cabriel

River since the available curves from literature were not in rivers with large flow. The

Cabriel curves presented, in general, higher suitability for lower velocities and depth

than those from the literature. A great discrepancy appeared with the substrate

suitability. Therefore, accordingly with the data collected at the study site, a

combination of Bovee and Raleigh curves was considered appropriate. This

corroborates the need to validate models (Guisan and Zimmermann, 2000).

4.1.2 Multivariate Habitat Suitability Models

4.1.2.1 Cabriel river

Fuzzy modelling

In the present study, we applied the minimum smoothing technique to get unimodal

curves, although it was certainly subjective. In addition, expert judgment was also

applied in the combinations of partial suitability and in the determination of the

controlling variables under some conditions (i.e. depth was considered a controlling

variable on the extreme values and velocity was considered for the higher flows).

Therefore, certain a priori knowledge is necessary. However, the entire procedure

showed simple enough to be applied in further studies.


95

WP3: MORPH

The fuzzy models derived from the Cat.II ½ HSCs present good performance in terms

of assessment of the fish locations for the three life stages considered in the present

study. The fish locations were assessed mostly over 0.8 suitability. Although the

specificity was low because most of the area was also assessed over 0.8 suitability, it

was not considered wrong since the study reach hosts a stable trout population

(Martínez-Capel et al., 2009a). In addition, it has been reported that over-prediction is

preferable to under-prediction (Mouton et al., 2010); that is, the absence of the target

species in a suitable area may be due to the unbalanced colonization of habitats in the

study area given the presence of barriers, temporal population variations or sampling

inefficiencies (MacKenzie et al., 2003). The limited amount of trout could condition the

colonization of all the places which present suitable conditions. However, the pattern of

the assessment over the study site showed a decay on the suitability over the deeper

areas (Figure 25 and Figure 26). specially significant for the adult and juvenile life

stages, which coincided with the most densely populated areas (Figure 29) for both life

stages. The flow when the validation was conducted was 0.89 m3/s representing the

Q85, then larger depths were easily expected. The trout can be considered a territorial

fish (Chapman and Bjonn, 1969; Titus, 1990; Johnsson et al., 2000) and this

territoriality is related with food availability (Brännäs et al., 2003). Trout is a drift-feeding

strategist (Elliott, 1973; Bachman, 1984) holding stations in slow water, but close to a

fast current (Wańkowski and Thorpe, 1979; Bachman, 1984), and recently some

models included macroinvertebrate drift to improve fish habitat modelling with

promising results (Hauer et al., 2012). All these suggested to improve the suitability for

trout in these deeper and more densely populated areas. Therefore, the Cat.III HSCs

were considered in order to test the impact of their selection in the development of

Expert-knowledge fuzzy models. The results were considered in general more

satisfactory than those based on Cat.II ½.

The frequency analysis of the general assessment over the entire reach and over fish

locations showed good agreement in terms of assessment of fish locations (Figure 28),

and assessment over the most densely populated areas (Figure 30). Although the

results for adult brown trout could be criticized because most of the trout locations were

assessed as medium, 0.4 to 0.6 suitability does not mean absence but presence.

Indeed, these values were assigned to the intervals of the curves that presented

certain suitability, which means that trout were observe but not abundantly. Therefore,


96

WP3: MORPH

it was considered a satisfactory result. On the other hand, that model can be

considered the most specific for adults because the trout would appear more

intensively in the areas with high suitability than according to the remaining models

(Figure 25), and the density and average suitability were clearly correlated (Figure 30).

The juvenile model based on Cat.III curves was the less specific and most of the area

was assessed with the maximum suitability (Figure 26), but the less densely populated

areas presented the lowest average suitability (Figure 30). The model generated for the

fry life stage presented the lower improvement, but while the Cat.II ½-based presented

a slight decrease for the more densely populated areas (Figure 30), the Cat.III-based

presented an slight increase (Figure 30).

Regarding the Data-driven models, the values of the performance criteria obtained can

be considered acceptable for adult, poor for juvenile, and good for fry in comparison

with previous studies that used similar training strategies (Mouton, 2008; Muñoz-Mas et

al., 2012). The results derived from the assessment carried out with the Data-driven

models presented, in general, poorer results before their modification than the obtained

with the Expert-knowledge approach. These poor results could be attributed to the sub-

sampling procedure applied (Muñoz-Mas et al., 2012), but the models generated were

in every case under-predictive. The statistical tests showed no differences between the

original databases and the corresponding sub-samples. Therefore, the addition of more

cases corresponding to the Availability database will discourage the optimization

algorithm to determine a given rule as High because the amount of cases

corresponding to the Low Cat.will be larger. Therefore, the disagreements derive from

imperfections in the database, not from procedure error.

The Data-driven fuzzy models achieved values of the Cohen's Kappa that can be

considered acceptable in comparison with previous studies that used similar training

strategies (Mouton, 2008; Muñoz-Mas et al., 2012). Considering the amount of trout

locations at highly suitable microhabitats, the Data-driven fuzzy models (unmodified)

showed a poorer performance than the Expert-knowledge approach (Figure 28), but

they presented a positive correlation between the assessed suitability and the trout

density (Figure 30). The observed deficiencies were improved by modifying the

corresponding rules, thus providing finally satisfactory results.


97

WP3: MORPH

The Data-driven fuzzy model for the adult size class was considered satisfactory despite the

relatively low value of the performance criterion (Kappa = 0.31). The model presented the best

trade-off between sensitivity and specificity (Figure 28), in addition to a positive trend between

average suitability and trout density (Figure 30). These results agreed with previous studies

(Heggenes, 1996; Ayllón et al., 2010) that assigned maximum suitability to the deeper areas.

The modification of this Data-driven fuzzy model was mainly considered in the development of

the proposed methodology. In order to maximize the model agreement one rule was modified

(Rule 14, Table 11). The modified Data-driven model improved the habitat assessment over the

trout locations, displacing all of them to the interval comprising the maximum suitability, but

keeping certain specificity (Figure 28). Therefore, the model was considered suitable for further

analysis. However, the rule that includes the most extreme condition -High velocity, High depth,

High substrate- was determined as High. This will provide an increasing average suitability as

the flow increases, and it should be taken into account if this model is applied in further studies,

Our results indicate that the limiting conditions for adult brown trout had not been surveyed. One

of the advantages of the fuzzy approach is its versatility and easy adjustment. Therefore, this

deficiency could be fixed adding an extra Fuzzy Set to depth or velocity and determining the

rule consequent in those new cases as Low.

The Data-driven model for the juvenile class was also considered satisfactory, although it

achieved the lowest Kappa value (0.21). The frequency analysis over the validation site and the

assessment at the trout locations presented an acceptable trade-off between the sensitivity and

the specificity (Figure 28). The modification of this model implied a single rule (Rule 4, Table

11), displacing the habitat assessment over the trout locations close to the maximum suitability

for all the observed individuals; the frequencies of the available unsuitable pixels remained

almost unaltered (Figure 28).

The analysis of the consequences showed certain discrepancy with the literature,

because the habitat suitability has been considered to decrease beyond certain depth,

whereas our Data-driven model determined the inclusion of High depth as High

(Bovee, 1978; Raleigh, 1984; Ayllón et al., 2010). However, in contrast to the adult

model, the Fuzzy Rules including the maximum velocity were assessed as Low in any

case; this allows applying the model in larger flows. On the other hand, a careful

application is necessary in the habitat assessments when considering Low velocity,

Low dept and High substrate, since it has been assessed as Highly suitable, and could

provide the maximum suitability over too shallow areas. In that sense, a narrow Fuzzy

Set covering the extremely shallow depth would be preferable.


98

WP3: MORPH

The data-driven models for the fry class were clearly the worst, despite the highest

value of the performance criteria (Kappa = 0.37). The frequency analysis in validation

determined the maximum frequency of fish in the unsuitable microhabitats. These

results were not surprising because the database for fry had the lowest sample size (N

= 44), which hardly could cover the whole conditions found in the study site, even

taking into account that the validation sample was almost twice as large (N = 79). The

fry model considered that small fish prefer pools in accordance to previous studies, but

disagreeing with the distribution pattern in the spatial validation (Figure 27). The

observed pattern (Figure 27), accordingly with several authors, could be produced by

the exclusion produced by the presence of larger trout, which also prefer these areas

(Raleigh, 1984). However, the Expert-knowledge model did not show that pattern, thus

these conclusions could derive from imperfections on the database, or from the

exclusion of the better habitats derived from the presence of older individuals.

Therefore, further effort should be placed in the improvement of the fry database.

The habitat suitability for the brown trout spawning lacked of validation with

independent data. However the sample size (Nuse = 353) or the achieved values of the

performance criterion (TSS = 0.40 ± 0.018) encourage its acceptance for further

applications. Our results mostly agreed with previous studies (Bovee, 1978; Louhi et

al., 2008), with maximum suitability around 0.4 m depth and velocity lower than 0.5

m/s. However, it is remarkable the present study provided wider suitable range than

those from literature. In addition, the automatic discretization through the Shannon-

Weaver entropy resulted questionable and strongly dependent on data distribution.

This was one of the main reasons not to include its optimization as a default option in

the FISH software (Mouton personal communication). However, at this point the

versatility of fuzzy modelling was underlined because this methodology allowed the

modification of the generated model altering the Fuzzy Sets with a more widely

accepted discretization, and even adding an extra Fuzzy Sets covering the unsampled

habitat conditions.

The Data-driven approach resulted highly dependent on the training database, but with

model validation its adjustment can be easily carried out. In addition, this kind of

models inform about the relations between input variables and how their combination

conditions the output, thus becoming a preliminary step in the development of any


99

WP3: MORPH

Expert-knowledge model. However, in rare species field data collection could be

unavoidable and the proper training of the Data-driven models unaffordable. Therefore,

the Expert-knowledge approach and the Data-driven approach are cursed to

understand each other. This has been specially demonstrated in the spawning case.

Expert-knowledge approach based on Habitat Suitability Curves present the ability to

fill gaps in the database, so this is the standard procedure in the development of

habitat suitability curves becoming more robust models, because best meet with the

ecological gradient theory.

On the other hand, the general overview on the results corroborate the premise

proposed by Guisan and Zimmermann (2000) about the necessity of validating habitat

models with independent data in order to test their reliability and applicability

(transferability). In the present study there were several issues to miss-select the

proper model. In general, there should be no doubts about the validity of the data

collected; rivers were from the same region and dimensions, and the sample size was

acceptable, especially for adult and juvenile (but not for fry). The flow assessed in the

Cabriel River (Q=0.89 m3/s) did not provide hydraulic conditions out of the surveyed

extremes; the simulated flow produced a maximum average velocity of 0.53 m/s and a

maximum depth of 1.4 m, both quite close to the maximum surveyed values in the Use

datasets, and clearly comprised in the Availability datasets. Therefore. both curves

Cat.II ½ and Cat.III, were within the range to develop fuzzy habitat suitability models.

The literature on the topic has reprobated Cat.III curves (Bovee, 1996). Therefore, the

most probable choice would be a miss-selection of the proper model selecting the

Cat.II ½ derived models. However, according to our results, the models derived from

Cat.III curves were recommended for further analysis of adult and fry. The miss-

selection could also be due the acceptance of the results from the Data-driven

approach, since it has been widely and successfully applied (Mouton, 2008; Mouton et

al., 2009a; Mouton et al., 2011), but in this case, cross validation resulted insufficient to

get the proper generalization capability if the database presents a limited amount of

training instances. Nevertheless, we recommend the application of the juvenile Data-

driven fuzzy model in further studies.


100

WP3: MORPH

Probabilistic Neural Networks

The PNN approach yielded acceptable results, comparable with similar studies

(Mouton et al., 2008), taking into account that the TSS parameter produced the same

results as Cohen's Kappa (Cohen, 1960) parameter in 0.5 prevalence databases

(Allouche et al., 2006). This also demonstrates the ability of PNNs to deal with low

prevalence databases producing accurate models. The networks showed slightly

overpredictive. Sensitivity was certainly higher than specificity (Table 8). Previous

research demonstrated that overprediction does not necessarily imply a model error,

but underprediction usually means a model failure (Mouton et al., 2008; Mouton et al.,

2009b). In contrast, PNN presented poor results in the assessment of the hydraulic

model and the trout locations at the validation site, because the maximum suitability

was not achieved.

According to the classification nature of this kind of networks, 57% of the trout locations

were assessed as 'presence'. However, the results presented certain sensitivity and

specificity, since trends in the frequency analysis of the Availability (i.e. assessment of

the entire simulated reach) and Use (assessment of trout locations) were inversely

correlated (¡Error! No se encuentra el origen de la referencia.). The PNN showed

generalization capability (¡Error! No se encuentra el origen de la referencia.): 85%

of the trout were observed over cobble, boulders and bedrock, which also could

provide high suitability, but the assessment of the synthetic database showed that the

achieving of high suitability values over these substrates needed an increase of water

depth as the substrate index increases (i.e. from cobble to bedrock) (Figure 33). The

maximum simulated depth was 1.4 m, but the average depth of the hydraulic model

was 0.35 m. As a consequence, to reach higher values of the suitability index, depth

should be larger. Depth could thus be considered the bottleneck in the correct

assessment of the hydraulic model and the trout locations. The flow observed during

the validation survey represented the Q85. Therefore, larger depth was easily expected

and, as the model did not produce suitability higher than 0.811, it could be considered

that the PNN was correctly assessing with higher values the residual suitable habitat. It

was considered that trout were occupying the better but not the best habitat, since this

one was unavailable in the validation site. However, during the leave-one-out

procedure, a maximum suitability value of 0.811 was reached. This could discourage


101

WP3: MORPH

the application of this kind of networks in terms of probability, considering only their

classification ability. However, upholding its applicability, the assessment of a given

flow in a river reach is usually summarized in a single index, the Weighted Usable Area

(WUA) which also ranges from 0 to 1 (Bovee and Cochnauer, 1977). The WUA for the

range of feasible flows is usually calculated, generating the WUA-flow curve. The

Spanish norm for hydrological planning (Instrucción de Planificación Hidrológica (IPH)

Ministry of Environment and Rural and Marine Affairs; MAGRAMA 2008) established

that the minimum flow should be selected within the range of 50–80% of the maximum

WUA. Therefore, this value is related with the relative maximum suitability, but it is not

mandatory to achieve the absolute maximum suitability (i.e. suitability = 1).

A previous study demonstrated that the flow that represents the 50–80% of the

maximum WUA for a single model is independent from the values of the WUA under

different hydraulic conditions (Muñoz-Mas et al., 2012), and that it rather is related to

the evolution of the habitat under different hydraulic conditions (i.e. different running

flows). Then, to improve the accuracy of the calculation of that minimum flow, a

continuous output, (i.e. the suitability index ranging from 0 to 1) would be preferred

over a dichotomous output (i.e. Presence/Absence). However, certainly the

dichotomous output could also provide a good basis to calculate the minimum flow, and

further analyses could be taken to corroborate that issue. Therefore, PNN provide

relatively promising results to be applied in fish habitat modelling.

4.1.2.2 Siurana River

The achieved prevalence (0.21) was relatively high. However, this value still

recommends the execution of a sub-sampling procedure to avoid the undesirable

effects of the unbalanced prevalence. Although no statistical tests were carried out, the

direct observation of the frequency distribution of the involved variables in both the sub-

sampled and the original databases showed no strong differences, and both databases

were pooled to optimize the system (Figure 34). Following the Data-driven approach,

the achieved values of the performance criteria were similar to previous studies that

used similar training strategies (Mouton, 2008; Muñoz-Mas et al., 2012). However, the

amount of uncovered rules was higher in our case because of the high gradient and

velocity of the Mijares River. The maximum frequency was present for the interval


102

WP3: MORPH

around 0.75 m/s. This relatively high velocity removed most of the fine substrate, and

therefore, several rules had no training instances (Table 9), mostly those with Low

substrate class. This fact does not influence the assessment of the original database

through the generated models, since those rules were not applied in that case. No

training samples means mostly not cases in the original database. However, it is

possible that the sub-sampling procedure exclude some samples with rare

combinations of the variable values. On the contrary, this lack of training samples with

Low substrate will condition the applicability of the presented models on different

locations, and agreeing with previous studies that recommend model validation

(Guisan and Zimmermann, 2000), we strongly recommend it before their transferability.

The assessment of the entire database through the developed models showed

disparity of results. The Data-driven model based on Cat.II ½ HSCs showed to be

surprisingly underpredictive, compared both to entropy-based with similar amount of

fuzzy sets and to the Expert-knowledge approach. The Data-driven approach based on

Cat.III curves resulted overpredictive and similar to the Expert-knowledge based on

Cat.III curves. The results in terms of specificity and TSS were quite similar (Table 10),

and the trained rules differed in just one consequent with those generated in the

Expert-knowledge approach (Table 9). The unconstrained entropy-based approach

improved those models based on Cat.III curves, achieving higher performance (Table

10), and presenting less uncovered rules. The Expert-knowledge approach based on

Cat.II ½ resulted extremely overpredictive, with close-to-zero specificity (Table 10). The

Expert-knowledge approach based on Cat.III showed good performance, being slightly

overpredictive (Table 10).

Accordingly to the obtained results, the selected model to carry out further

assessments would be the unconstrained entropy-based model. It yielded the highest

performance and the lowest number of uncovered rules. In addition, it was

overpredictive but not as dramatically as the Cat.II ½-based models. Certain

overprediction is desirable since it does not necessarily imply a model error, in contrast

to underprediction (Lütolf et al., 2006; Mouton et al., 2008; Mouton et al., 2009b).


103

WP3: MORPH

Further research will be focused in transferability of the generated models in order to

know their generalization capability. Specifically on the Siurana River, this is a target

river in the SCARCE project.

4.1.2.3 Ésera River

The Expert-knowledge fuzzy models were based on the Cat.II HSCs developed by

Bovee (1995 unpublished), providing the simplest discretization for adults and

juveniles, and were developed according to the Expert-knowledge fuzzy modelling.

That simplest discretization forced to alter the standard procedure to develop the Fuzzy

Rules, and then constrains were slightly relaxed, permitting to enlarge the suitable

conditions. The substrate was in any case based on literature or field work derived from

other basins, demonstrating the previously mentioned versatility of the fuzzy logic

approach. The results pointed out the necessity of model validation, agreeing with

previous works (Guisan and Zimmermann, 2000). Unfortunately, the campaign carried

out on the Ésera study site just provided data for a qualitative validation of the fry life

stage (N = 4), recommending to modify the curves developed in the Iberian Peninsula,

making coarse substrate as the most suitable. The Ésera river is by far the largest river

studied in the MORPH package (Mean annual flow = 16.85 m3/s around the study site),

and thus, the Cat.III models developed in the Cabriel River were considered

inappropriate. These HSCs provide the maximum suitability for a faster velocity and

larger depth, with a relatively constant suitability over coarse substrate except for fry.

However, the use of preference curves has been strongly discouraged (Payne, 2009),

and because no validation was possible, the most conservative choice was the

selection of Cat.II HSCs, but selecting the ones developed in the largest river as

possible.

The largest river presenting a set of curves for brown trout was the South Platte River,

where flows ranged from 7 to 18 m3/s during the survey. These curves agreed with the

most relevant studies on brown trout microhabitat suitability (Bovee, 1978; Raleigh,

1984; Hayes and Jowett, 1994; Heggenes, 1996; Ayllón et al., 2010). The similar

results across rivers indicate the consistency of the habitat selection by the same

species, when the bioenergetic conditions (indirectly indicated by stream flow, food

availability and physical conditions) are similar; therefore, it is coherent to compare with


104

WP3: MORPH

other rivers in other countries, as it has been done in diverse papers in the literature

about habitat suitability models for fish. The validation carried out in the Cabriel River

demonstrated that the Expert-knowledge approach can easily generate accurate

models to assess the habitat suitability of a river reach. Besides, the value of the fuzzy

logic approach has been emphasised, due to its capability to deal with incomplete

information (Van Broekhoven et al., 2006). The Ésera River is impacted by

hydropeaking. Therefore, in a frame of global change, alterations on the operation rules

could be expected; then, a compromise solution joining energy needs with legal

limitations would be mandatory. The ecological flow (e-flow) regime (Poff et al., 1997)

would be certainly unattainable. However, from the legal point of view, the assessment

of a minimum flow is necessary (Instrucción de Planificación Hidrológica (IPH) Ministry

of Environment and Rural and Marine Affairs; MAGRAMA 2008). As it has been

mentioned before, that minimum flow does not differ according to the values of the

available habitat in a given condition rather than in its evolution (Muñoz-Mas et al.,

2012). Therefore, even if an slight decrease on model accuracy occurs, it does not

automatically result in a inappropriate e-flow assessment. Therefore, based on the

literature, expert-knowledge and legal implications, we considered the developed

models suitable to be applied in further analysis.

4.1.2.4 Macroinvertebrate model

The optimization of the fuzzy models was good and the results consistent for further

applications. However, the developed models presented several limitations. The

microhabitat suitability models produced kappa values similar to freshwater fish models

trained following a similar strategy (Mouton et al., 2008; Muñoz-Mas et al., 2012).

Although the Shannon-weaver entropy values for the outputs were not so large (Figure

38), the amount of fuzzy sets was considered acceptable, and a larger amount could

provide unreliable results, especially considering the sample size (109 samples).

Most macroinvertebrates microhabitat studies focus on depth, velocity, shear-stress,

shear velocity, Reynols number and Froude number (Brooks et al., 2005). Recently,

the use of complex hydraulic variables such as the shear stress has been strongly

recommended (Mérigoux et al., 2009). However, in some cases the approach was

univariate, and considering the multivariate aspect of the fuzzy logic, it was considered


105

WP3: MORPH

that most variables were redundant, and the three variables considered were sufficient.

Besides, the multivariate approach has been also suggested as a better procedure to

analyze this kind of data (Brooks et al., 2005)

Despite the inherent differences in the traits of species aggregated at the order level,

the models developed showed a clear pattern. EPTs in general appeared more often

(larger density was observed) for large depth and coarse substrate, whereas the

velocity showed less clear pattern. The present study showed discrepant results with

previous works, although certain agreement was also observed. It has been reported

that some macroinvertebrates inhabit depositional zones (i.e. silt), but most of the

species in mountain streams inhabit coarse substrate (Jowett, 2003). This could be the

main reason to the observed increment over the density or presence in the rules that

cover the coarser substrate (Table 12). Quinn and Hickey (1994) suggested that the

coarser the substrate the larger the difficulty to hold position, but this asseveration

disagree our results. In our case, the decrease on density or presence was observed

over bedrock instead over coarse substrate, what could reflect differences in survey

conditions rather than a real preference regarding fine substrate. In contrast, our

results agreed with several studies (Growns and Davis, 1994; Gore et al., 2001; Jowett

and Davey, 2007) who showed preference for coarser substrates.

Depth showed a positive effect on macroinvertebrate density, in accordance with

previous studies (Quinn and Hickey, 1994; Gore et al., 2001; Jowett and Davey, 2007),

but contrasting with others (Brooks et al., 2005). Velocity showed no clear pattern, in

contrast to the aforementioned studies. Nevertheless, Ephemeroptera showed certain

tendency to prefer faster flows. Previous studies considered that Ephemeroptera and

Plecoptera prefer relatively high flows, in contrast to Trichoptera, which were more

abundant in slow. (Pastuchová et al., 2008). However, our results not agree at all that

asseverations. We do not know whether these inconsistencies with the literature result

from the aggregation of species differing in their traits, or from the fact that the velocity

plays a smaller role than depth or substrate. It has been reported that

macroinvertebrate abundance decreases as the Reynolds number increases (Brooks

et al., 2005). In that sense, we cannot consider velocity a good surrogate to this

variable. Therefore, it could be interesting to consider also the Reynolds number, but

the fuzzy logic approach presents the disadvantage of requiring cases to train each


106

WP3: MORPH

rule, and due to the limited amount of cases, the addition of an extra variable could

seriously limit the modelling.

More complex variables (i.e. shear stress and Froude number) can outperform simpler

ones (i.e. velocity, depth or substrate) (Mérigoux and Dolédec, 2004; Dolédec et al.,

2007; Mérigoux et al., 2009). Therefore, further research should focus on collecting

extra data about hydraulic characteristics, for instance applying the hemisphere-

method to measure near-bed flow forces (Statzner and Müller, 1989).

Mérigoux and Dolédec (2004) observed seasonal changes on habitat preferences. This

was also suggested by Jowett (2003). Because most of our samples were collected in

summer, our models should only be applied to summer periods. Unfortunately, the

uncertainty about the transferability (Moyle and Baltz, 1985; Gore and Nestler, 1988;

Fukuda, 2010) led to recommend site-specific criteria. The rivers included in the

models did not present a large variability. Therefore, it is strongly recommended to

validate the model before any application in further studies. The generation of habitat

suitability criteria at the microscale for macroinvertebrates species has proved to be

difficult, and inconsistencies were observed. Thus, several non-hydraulic factors can

affect the development of populations (Gore et al., 2001), what suggests the necessity

of increasing the training datasets in order to shed some light over the confounded

effect of hydraulics and other environmental factors. Therefore, we do not consider

these models suitable for studies generated out of the study area.

4.2 Habitat suitability modelling riparian vegetation

4.2.1 Composition

A free-flowing study reach with good riparian quality status was surveyed in the Jucar

River Basin District (East Spain) to implement this study. The site had a heterogeneous

shape, showed high species richness, and a balance between tree and shrubs species.

In total, 839 specimens of 10 species were recorded within the site. The small sample

size presented by some species was more related to phytogeographic reasons than to

human-induced alterations, i.e., their scarcity might be consequence of being in the

outer range of their optimal distribution.


107

WP3: MORPH

4.2.2 Transverse variation across the floodplain: definition of guilds

Four guilds were found in terms of distance and elevation, but differences were clearer

in elevation. In order of elevation, these guilds were [SP, SL], [SL, FA, PN], [PN, PA,

SE] and [PC]. The terrestrial PC was established as a completely different group,

although this was not evident during the field surveys. SL and PN were the species that

intermediated the change from one guild to the other.

Regarding the soil trends, the sand content was very high in the entire site. The gravel

content increased from channel edge to the upland zone. On the contrary, the silt and

clay content decreased with elevation. Both morphological variables produced different

rankings, mainly due to the high complexity (several channels) and natural dynamism

of the site, but both variables were able to define a gradual transition of species in the

riparian zone.

4.2.3 Importance and uncertainty of the variables in the definition of guilds

The use of the central point of the vertical projection of any plant was considered as a

reasonably valid simplification to implement this study. However, in some cases, the

predicted surface could be displaced respect to the observed in the field, for example in

trees with a highly heterogeneous crown or because of the presence of physical

barriers, such as bedrock outcrops. Nevertheless, this displacement would take place

in a small proportion of cases over the entire database, with no consequences in the

results if sampling size is large.

The two morphological variables studied, i.e., distance and elevation to/above thalweg,

were considered as an easy way to measure indirectly other factors such as tolerance

to inundation (Harris, 1986) and anoxia, or capability to resist the drag forces caused

by floods. Both variables led us to compare the species with a ‘static’ point of view

based on the current morphology. In this way, the surveys showed the result of many

hydrological cycles and repeated processes of vegetation succession and

retrogression. According to Lite et al. (2005), both variables are negatively correlated


108

WP3: MORPH

with inundation frequency. Hence, species located closer to the thalweg tend to have

higher tolerance to inundation.

Between both variables, elevation above thalweg was considered to better distinguish

guilds, and other studies showed elevation to be more correlated than distance with

inundation frequency (Lyon and Sagers, 1998). In addition, according to Lite et al.,

(2005), elevation above thalweg showed a negatively relation with ground-water depth

as well. Given that elevation is less related to changes in morphology than distance,

the conclusions from this variable are more transferable between rivers. In our case,

distance and elevation performed similarly when the channel was single, but when it

was more complex the elevation performed better in separating guilds. Furthermore,

the ordination of species produced by elevation was more gradual, what helped to

discriminate better the response of each species.

According to Lite et al. (2005), water availability may be a stronger limiting factor than

disturbance along semi-arid region rivers. Apart from water availability, other variables

that could determine the lateral vegetation pattern are related to the soil type and

canopy (dense canopy influences the litter content and reduces light levels).

Depending on the soil properties, the communities are composed of plants tolerating

episodic to permanent inundations. Plants tolerating inundation are bound to soils with

low permeability. More permeable soils are the habitat of plants tolerating the often

rapidly changing extremes of flooding and dryness.

A decreasing trend was observed in organic matter content from the water’s edge to

the upland fringe. This could be explained because the valley is a calcareous gorge,

and there is low organic matter content in the slopes. The riparian zone has a greater

income coming from the trees than from the surrounding areas. The slopes are very

rocky and the typical Mediterranean vegetation (Pinus and Quercus) do not produce as

much organic matter as the riparian deciduous species.

The species located in a second line in the floodplain (e.g., SL and PN) experienced

less moisture content than those located in the first line. However, the standard

deviation was high, denoting variety of conditions across the floodplain and between

the specimens of a particular species. According to Bagstad et al. (2006), the


109

WP3: MORPH

differences in soil parameters may be more patent between different size/age class

than between species.

Nonetheless, the position revealed for each species and the final preferences obtained

by the species should be interpreted with caution. In our case, the most remarkable

uncertainties can be defined as follow:

• The channel morphology, which can determine the position of the thalweg. In

the case of a single channel, both distance and elevation to/above thalweg can

perform similarly, but in the case of a more complex channel, elevation could

perform better than distance to thalweg. If the cross-section is regular (case of a

rapid, current or glide) the thalweg will be approximately located in the centre of

the riverbed, but if the cross-section is more irregular (case of a pool), this will

make the thalweg to move more towards one bank than the other. In the case

of the Cabriel River, the thalweg was very divagating and it was located nearer

of one bank in most cases. These could have affected more the results

obtained from distance than from elevation above thalweg.

• The proper taxonomical identification of the species. Different species or even

subspecies, varieties or hybrids (not detected in our field surveys) could differ in

their performance and affect results. Other specimens could be of questionable

origin (planted or weed-grown). Clonal plants present another problem. They

grow as a set of connected shoots. However, the connections may be buried,

so it is impossible to distinguish the complete individual (the genet) (Bullock,

2006). In all these cases, only genetic analysis could clarify their origin, which

may lead to an extensive data collection effort. Hence, we can assume that the

level achieved in the field surveys was the best available approximation.

• Since the riparian zone is highly variable in terms of soils, the soil survey was

designed to cover and summarize as much as possible all the soil variability,

but in a simplified way. Despite this simplification, trends in preferences for the

different riparian species were found.

• Although overlapping was clear among species for all environmental variables,

it was possible to define groups of species. Nevertheless, the components of


110

WP3: MORPH

these groups depended on the trimming percentage and significance level

applied in each case.

4.2.4 Management implications and further research

The methodology applied in this study has proved to be suitable to define guilds of

species in a multivariate way, in response to the physical habitat conditions. It could be

also useful to measure the effects of regulation upon riparian woody species; therefore,

it could be applicable to inform water management decisions involving changes in flow

regime. Considering the ranges of distribution of a certain key species or guild in

natural conditions, the measurement of the deviation from those ‘reference values’

could be an useful tool to quantify the status of a water body in terms of riparian

species. Additionally, this method could help to test global change effects in

unregulated sites where reductions in the water inflow may take place. Similarly, the

riparian species, as long-term bioindicators, could respond to the changes of natural

disturbances by shifts in their positional patterns. Another possible use of the

information provided by this study could be the application of the ranges of distances

and elevation preferred by the species to design plantations in rehabilitation projects.

Although the relation between abundance and cover gave a good estimation of the

development stage of the riparian vegetation, an improvement of that estimation could

be achieved by developing growth curves for key riparian species, i.e., curves relating

height and/or diameter at breast height with the age of the plant. This

dendrochronological approach would let to determine specifically the time span of each

plant in the study reach and therefore, would let us know the hydrological time series

under which each plant was established (recruited) and developed. Using hydrological

variables instead of morphological ones could be an additional improvement.

Response curves of hydrological variables could be developed for riparian woody

species with a one-dimensional hydraulic model in each study reach and the

aforementioned dendrochronological information. This could help in the definition of

hydrological guilds of riparian species and their associated suitability curves.

Another possible improvement in this study could be achieved using the base flow as a

reference instead of the thalweg. This river stage is more ecologically meaningful,


111

WP3: MORPH

because the base flow is considered as a reference for the survival and maintenance of

the riparian vegetation, with special emphasis in Mediterranean environments.

5 CONCLUSIONS

The present deliverable comprises a variety of proficient approaches in the

development of habitat suitability models, for key species of fish, macroinvertebrates

and riparian vegetation; from univariate techniques, HSCs (Bovee, 1986), to complex

multivariate methodologies such as fuzzy inference systems (Mamdani, 1974, Zadeh,

1965) or Probabilistic Neural Networks (PNN) (Specht 1990). Nevertheless several

conclusions could be stated.

Regarding fish species, the Category III HSCs for trout, adult, juvenile and fry

suggested to be a better choice in studies following a univariate approach, in the Jucar

River Basin District (JCBD). However, very serious considerations about the planning

of field work, field methods (equal effort), sample size and data analysis (smoothing,

etc.) must be done before application of this method elsewhere.

In contrast, the observed overcorrection in the Category III HSCs for the redfin barbel,

discourage their application in further studies. Therefore the better choice would be use

of the Category II HSCs in studies following a univariate approach.

In the Siurana River, the unregulated segment under study was visited and the

hydrological information was analysed. The resulting information indicated that summer

low flows were too low, and the flow regime is very irregular and torrential, as other

Mediterranean streams. The physical habitat simulation (frequently applied on

environmental flow assessment), as well as the SCARCE project, focus on low flows

and effects of water scarcity. However, any study of habitat suitability criteria requires

some characteristics in the study sites, such as the availability of diverse habitats and

flow patterns, and habitats connectivity, to make posible the study of fish habitat

selection. In this river, the very low flow, the scarce water in the remaining pools and

the presence of only one type of mesohabitats, namely pools, limits the fish electivity in

a serious manner. Thus, we searched for a similar stream (in terms of physical

features) where we performed the microhabitat survey succesully, for the target


112

WP3: MORPH

species, the redfin barbel. The surveys took place in the Mijares River (Valencian

Country) during early summer 2012.

The Ésera River presented a larger uncertainty because the fish data were very scarce

and did not allow the development of HSC in situ. However, habitat selection patterns

of brown trout are well established in broad terms (Ayllón et al., 2010) and exist

abundant literature (Bovee, 1978; Raleigh, 1984; Hayes and Jowett, 1994; Heggenes,

1996; Ayllón et al., 2010). Therefore we presented the better choice among the

available HSCs corresponding to Bovee’s HSCs (1995) for adult and juvenile, as well

as the HSC developed in the Júcar River Basin District (JRBD) for fry fish. However, a

strong deviation between the real habitat suitability and the presented assessment

cannot be discarded. Thus, the fuzzy models should be preferable, since fuzzy

inference systems are able to deal with the inherent uncertainty in ecological data by

discretizing inputs and output in broad terms (Mouton et al., 2009a).

The expert-knowledge demonstrated to be proficient in the development of fuzzy

inference systems. However, the spatially explicit validation of our models has been

demonstrated fundamental in the selection of the best model. Furthermore, the expert-

knowledge approach presented the capability to transform dichotomous input data into

a wider range of outputs.

Concerning the JRBD, the expert-knowledge fuzzy models based on Category III

should be the better choice in studies following a multivariate approach for adult and

fry, because they presented good sensitivity and specificity, as well as a satisfactory

correlation between trout density and habitat suitability. In contrast, the data-driven

fuzzy model was the best option for juvenile fish.

Accordingly to the results, the best choice of multivariate techniques for the redfin

barbel would be the unconstrained entropy-based model. It yielded the highest

performance and the lowest number of uncovered rules. However, in concordance with

previous studies, the model validation was strongly recommended before its

transferability (Guisan and Zimmermann, 2000).

The validation carried out in the Cabriel River demonstrated that the expert-knowledge

approach can easily generate accurate models. The great value of the fuzzy logic


113

WP3: MORPH

approach has been emphasised, due to its capability to deal with incomplete

information (Van Broekhoven et al., 2006). Accordingly to the stated above the trout

models should provide a reliable overall assessment but a deviation between the real

suitability and the assessed cannot be discarded. Therefore, concerning the crucial role

of the brown trout in the generation of ecosystem services (Schindler et al., 2010), we

consider necessary the allocation of an increasingly amount of resources in order to

properly study the habitat suitability in the Ésera River basin.

The PNN (Specht 1990) has proven to be a proficient technique in studies focused on

the physical habitat simulation (Bovee, 1998). However, they presented also several

shortcomings such as the trimmed outputs. This drawback can make it difficult for

widespread applications by engineers or managers. Nevertheless, the present study

provided valuable information to the scientific community.

Regarding macroinvertebrates, the generation of fuzzy inference systems for

macroinvertebrate species has proved to be difficult, and some inconsistencies were

observed. Several factors affect the development of populations (Gore et al., 2001)

thereby, a larger dataset would be preferable. Thus, the presented results should be

considered a preliminary attempt in modelling their habitat suitability not suitable for

further applications.

In regards to riparian vegetation, the approach based on guilds was necessary and

finally successful, determining these four guilds, in order of elevation above the

thalweg: [SP, SL], [SL, FA, PN], [PN, PA, SE] and [PC]. The terrestrial PC (Pinus of

several species) was established as a completely different group. The species

abbreviations were: SP, Salix purpurea L.; ST, Salix triandra L.; CS, Cornus sanguinea

L.; SL, Salix alba L.; FA, Fraxinus angustifolia Vahl; PN, Populus nigra L.; SE, Salix

eleagnos Scop.; CR, Crataegus monogyna Jacq.; PA, Populus alba L.; PC, Pinus spp.

The use of the central point of the vertical projection of any plant was considered as a

reasonably valid simplification for a good balance between effort and

information/results. The two morphological variables considered for the guilds, i.e.,

distance and elevation to/above thalweg, were considered suitable as indicators of

some species traits, such as tolerance to inundation and anoxia, or capability to resist

the drag forces caused by floods. These variables were analysed based on the actual


114

WP3: MORPH

morphology, as a result of many hydrological cycles and repeated processes of

vegetation succession and retrogression. Beyond these results, the consideration of

the changes in fluvial forms and the changing hydrological conditions is recommended

for further studies. Given that elevation above thalweg is less related to changes in

morphology than distance, the conclusions from this variable are more transferable

between rivers.

In terms of elevation above thalweg, the best performing variable, SP and PN (pioneer

species) were located at the lowest locations (0-0.5 m above thalweg). The highest

abundance of SP was registered in the range 0.5-1.5 m. The group of PA (dominant),

PN and FA (subdominant), ST, SL and CS had their largest abundance values in the

range 1-1.5 m. PA showed the widest distribution along the elevation gradient. In the

next range (2-2.5 m) the dominant species were PA, SE and PC. Finally, PC

dominated and presented its highest abundance values in the range 2.5-3 m, along

with PA and PC.

The methodology has proved to be suitable to define guilds of species in a multivariate

way, in response to the physical habitat conditions. The guilds approach is useful to

assess the effects of river regulation on riparian woody species; therefore, it could be

applicable to inform water management decisions involving changes in flow regime.

Regarding global change, it would be also possible to test the effects of relevant

reductions of the stream flow in unregulated sites. In the same way, the riparian

species, which are long-term bioindicators, could be responding to the changes in the

regime or frequency of natural disturbances by shifts in their positional patterns.


115

WP3: MORPH

6 REFERENCES

Ahmadi-Nedushan, B., St.-Hilaire, A., Bérubé, M., Ouarda, T.B.M.J., Robichaud, É.,

2008. Instream flow determination using a multiple input fuzzy-based rule system: A

case study. River Research and Applications, 24(3): 279-292.

Alcaráz-Hernández, J.D., 2011. Estado de las poblaciones de trucha en los ríos de la

Comunidad Valenciana y caracterización de sus hábitats, València: Universitat

Politècnica de València.

Alcaraz-Hernández, J.D., Martínez-Capel, F., Peredo, M., Hernández-Mascarell, A.,

2011. Mesohabitat heterogeneity in four mediterranean streams of the Jucar river

basin (Eastern Spain). Limnetica, 30(2): 15.

Ali, M.M., Dickinson, G., Murphy, K.J., 2000. Predictors of plant diversity in a hyperarid

desert wadi ecosystem. Journal of Arid Environments, 45(3): 215-230.

Allouche, O., Tsoar, A., Kadmon, R., 2006. Assessing the accuracy of species

distribution models: prevalence, kappa and the true skill statistic (TSS). Journal of

Applied Ecology, 43(6): 1223–1232.

Almodóvar, A., Nicola, G.G., Ayllón, D., Elvira, B., 2012. Global warming threatens the

persistence of Mediterranean brown trout. Global Change Biology, 18(5): 1549-

1560.

Angermeier, P.L., Davideanu, G., 2004. Using fish communities to assess streams in

Romania: Initial development of an index of biotic integrity. Hydrobiologia, 511: 65-

78.

Armstrong, J.D., Kemp, P.S., Kennedy, G.J.A., Ladle, M., Milner, N.J., 2003. Habitat

requirements of Atlantic salmon and brown trout in rivers and streams. Fisheries

Research, 62(2): 143-170.

Austen, D.J., Bayley, P.B., Menzel, B.W., 1994. Importance of the guild concept to

fisheries research and management. Fisheries, 19(6): 12-20.

Ayllón, D., Almodóvar, A., Nicola, G.G., Elvira, B., 2009. Interactive effects of cover and

hydraulics on brown trout habitat selection patterns. River Research and

Applications, 25(8): 1051-1065.


116

WP3: MORPH

Ayllón, D., Almodóvar, A., Nicola, G.G., Elvira, B., 2010. Ontogenetic and spatial

variations in brown trout habitat selection. Ecology of Freshwater Fish, 19(3): 420–

432.

Bachman, R., 1984. Foraging Behavior of Free-Ranging Wild and Hatchery Brown

Trout in a Stream. Transactions of the American Fisheries Society, 113: 1-32.

Bagstad, K.J., Lite, S.J., Stromberg, J.C., 2006. Vegetation, soils, and

hydrogeomorphology of riparian patch types of a dryland river. Western North

American Naturalist, 66(1): 23-44.

Bejarano, M.D., Marchamalo, M., González del Tánago, M., García de Jalón, D.,

Sordo-Ward, Á., 2012. Riparian guild composition changes along a regulated

Mediterranean stream in Central-Western Spain. Proceedings of the 9th

International Symposium on Ecohydraulics, Vienna (Austria): Published by ISE 2012

Organizing Committee, p. 8.

Benaka, L.R.E., 1999. Fish habitat : Essential fish habitat and rehabilitation. Bethesda,

Md. : American Fisheries Society (Symposium 22): 419.

Benjankar, R., Egger, G., Jorde, K., Goodwin, P., Glenn, N.F., 2011. Dynamic

floodplain vegetacion model development for the Kootenai River, USA. Journal of

Environmental Management, 92: 3058-3070.

Berrar, D.P., Downes, C.S., Dubitzky, W., 2003. Multiclass cancer classification using

gene expression profiling and probabilistic neural networks. Pacific Symposium on

Biocomputing: 5-16.

Bischoff, A., Freyhof, J., 1999. Seasonal shifts in day-time resource use of 0+ barbel,

Barbus barbus. Environmental Biology of Fishes, 56(1-2): 199-212.

Bovee, K.D., 1978. Probability of use criteria for the family Salmonidae, Information

Paper 4 FWS/OBS-78/07. Washington, D.C.: Fish and Wildlife Service, U.S.

Department of the Interior., p. 53.

Bovee, K.D., 1986. Development and evaluation of habitat suitability criteria for use in

the instream flow incremental methodology. Washington, DC: USDI Fish and

Wildlife Service Instream Flow Information Paper 21 FWS/OBS-86/7.: 235.

Bovee, K.D., 1996. The compleat IFIM : a coursebook for IF 250. Washington, D.C.:

U.S. Geological Survey, Biological Resources Division.


117

WP3: MORPH

Bovee, K.D., B.L. Lamb, J.M. Bartholow, C.B. Stalnaker, J. Taylor, and J. Henriksen.,

1998. Stream habitat analysis using the instream flow incremental methodology.

U.S. Geological Survey Information and Technology.(Report 1998-0004. ): 130.

Bovee, K.D., Cochnauer., T., 1977. Development and evaluation of weighted criteria,

probability of use curves for instream flow assessments: fisheries. Washington,

D.C.: Fish and Wildlife Service, U.S. Department of the Interior Instream Flow

Information Paper 3, FWS/OBS-77/63.: 38.

Bovee, K.D., Zuboy, J.R., 1988. Proceedings of a workshop on the development and

evaluation of habitat suitability criteria: A compilation of papers and discussions

presented at Colorado State University, Fort Collins, Colorado, December 8-12,

1986., p. 407.

Brännäs, E., Jonsson, S., Lundqvist, H., 2003. Influence of food abundance on

individual behaviour strategy and growth rate in juvenile brown trout (Salmo trutta).

Canadian Journal of Zoology, 81(4): 684-691.

Brooks, A.J., Haeusler, T., Reinfelds, I., Williams, S., 2005. Hydraulic microhabitats

and the distribution of macroinvertebrate assemblages in riffles. Freshwater Biology,

50(2): 331-344.

Bullock, J.M., 2006. Plants. In: W.J. Sutherland (Editor), Ecological census techniques:

A handbook. Cambridge (UK): Second edition. Cambridge University Press, pp.

186-213.

Cacoullos, T., 1966. Estimation of a multivariate density. Annals of the Institute of

Statistical Mathematics, 18(1): 179-189.

Canfield, R.H., 1941. Application of the fine interception method in sampling range

vegetation. Journal of Forestry, 39(4): 388-394.

Carling, P.A., 1993. The nature of the fluid boundary layer and the selection of

parameters for benthic ecology: Reply to Lamouroux. Freshwater Biology, 30(2):

333-335.

CEDEX, 2004. Directiva 2000/60/CE. Caracterización de los tipos de ríos y lagos.

Versión 1.0, Madrid: Ministerio de Fomento.

Chapman, D.W., Bjonn, T.C., 1969. Distribution of Salmonids in Streams, with Special

References to Food and Feeding. In: H.R. MacMillan (Editor), Symposium on

Salmon and Trout in Streams: University of British Colombia, p. 153-176.


118

WP3: MORPH

CHJ, 2005. Informe para la Comisión Europea sobre los artículos 5 y 6 de la Directiva

Marco del Agua, Valencia: Confederación Hidrográfica del Júcar. Ministerio de

Medio Ambiente.

CHJ, 2009. Documento técnico de referencia. Identificación y delimitación de masas

de agua superficial y subterránea. Ámbito territorial de la Confederación

Hidrográfica del Júcar, Valencia: Ministerio de Medio Ambiente.

Cock, M.J.W., 1978. The Assessment of Preference. Journal of Animal Ecology, 47(3):

12.

Cohen, J., 1960. A Coefficient of Agreement for Nominal Scales. Educational and

Psychological Measurement, 20(37).

Collier, K.J., 1993. Flow preferences of larval Chironomidae (Diptera) in Tongariro

River, New Zealand. New Zealand Journal of Marine & Freshwater Research, 27(2):

219-226.

Davis, J.A., Barmuta, L.A., 1989. An ecologically useful classification of mean and

near-bed flows in streams and rivers. Freshwater Biology, 21(2): 271-282.

De Pauw, N., Gabriels, W., Goethals, P.L.M., 2006. River Monitoring and Assessment

Methods Based on Macroinvertebrates, Biological Monitoring of Rivers: John Wiley

& Sons, Ltd, pp. 111-134.

Decocq, G., 2002. Patterns of plant species and community diversity at different

organization levels in a forested riparian landscape. Journal of Vegetation Science,

13(1): 91-106.

Dolédec, S., Lamouroux, N., Fuchs, U., Mérigoux, S., 2007. Modelling the hydraulic

preferences of benthic macroinvertebrates in small European streams. Freshwater

Biology, 52(1): 145-164.

Egger, G., Benjankar, R., Davis, L., Jorde, K., 2007. Simulated effects of dam

operation and water diversion on riparian vegetation of the lower Boise River, Idaho,

USA. Proceedings of the Congress- International Association for Hydraulic

Research, Arizona, USA, p. 6.

Elliott, J.M., 1973. The food of brown and rainbow trout (Salmo trutta and S. gairdneri)

in relation to the abundance of drifting invertebrates in a mountain stream.

Oecologia, 12(4): 329-347.

Francés et al., 2009. Proyecto RIBERA: Modelación matemática de ecosistemas de

ribera para la determinación de regímenes ecológicos en el río. Informe final.


119

WP3: MORPH

Fukuda, S., 2010. Assessing transferability of genetic algorithm-optimized fuzzy habitat

preference models for Japanese Medaka (Oryzias Latipes). 4th International

Workshop on Genetic and Evolutionary Fuzzy Systems (GEFS), Mieres (Spain), pp.

57-62.

Fukuda, S., De Baets, B., Waegeman, W., Verwaeren, J., Mouton, A.M., 2013. Habitat

prediction and knowledge extraction for spawning European grayling (Thymallus

thymallus L.) using a broad range of species distribution models. Environmental

Modelling & Software, 47: 1-6.

García-Pérez, A., 2005. Métodos avanzados de estadística aplicada. Métodos

robustos y de remuestreo. Madrid: Universidad Nacional de Educación a Distancia.

García, A., Jorde, K., Habit, E., Caamaño, D., Parra, O., 2011. Downstream

environmental effects of dam operations: Changes in habitat quality for native fish

species. River Research and Applications, 27(3): 312-327.

Gatz Jr, A.J., Sale, M.J., Loar, J.M., 1987. Habitat shifts in rainbow trout: competitive

influences of brown trout. Oecologia, 74(1): 7-19.

Gegov, A., 2007. Basic types of fuzzy rule based systems, Complexity management in

fuzzy systems. Studies in fuzziness and soft computing: Springer Berlin Heidelberg,

pp. 7-16.

Gibson, R.J., 1993. The Atlantic salmon in fresh water: spawning, rearing and

production. Reviews in Fish Biology and Fisheries, 3(1): 39-73.

Glenz, C., 2005. Process-based, spatially-explicit modelling of riparian forest dynamics

in Central Europe – tool for decision-making in river restoration, Lausanne: Thèse

No. 3223. École Polytechnique Fédérale de Lausanne.

Gore, J., 1987. Development and Applications of Macroinvertebrate Instream Flow

Models for Regulated Flow Management. In: J. Craig and J.B. Kemper (Editors),

Regulated Streams: Springer US, pp. 99-115.

Gore, J.A., Judy Jr, R.D., 1981. Predictive Models of Benthic Macroinvertebrate

Density for Use in Instream Flow Studies and Regulated Flow Management.

Canadian Journal of Fisheries and Aquatic Sciences, 38(11): 1363-1370.

Gore, J.A., Layzer, J.B., Mead, J., 2001. Macroinvertebrate instream flow studies after

20 years: A role in stream management and restoration. River Research and

Applications, 17(4-5): 527-542.


120

WP3: MORPH

Gore, J.A., Nestler, J.M., 1988. Instream flow studies in perspective. Regulated Rivers:

Research & Management, 2(2): 93-101.

Gosse, J.C., 1981. Brown trout (Salmo trutta) responses to stream channel alterations,

their microhabitat requirements, and a method for determining microhabitat in lotic

systems. Ph.D. Dissertation. Thesis, Logan: Utah State University.

Gosse, J.W., R. S. Helm, W.T., 1977. Microhabitat of fish in intermountain rivers.

Logan: Utah Cooperative Fish and Wildlife Research Unit.

Greenwood, J.J.D., Robinson, R.A., 2006. General census methods. In: W.J.

Sutherland (Editor), Ecological census techniques: A handbook. Cambridge, UK:

Cambridge University Press, pp. 87-185.

Grossman, G.D., De Sostoa, A., 1994a. Microhabit use by fish in the upper Rio

Matarrana, Spain, 1984-1987. Ecology of Freshwater Fish, 3(4): 141-152.

Grossman, G.D., De Sostoa, A., 1994b. Microhabitat use by fish in the lower Rio

Matarrana, Spain, 1984-1987. Ecology of Freshwater Fish, 3(3): 123-136.

Grossman, G.D., de Sostoa, A., Freeman, M.C., Lobon-Cerviá, J., 1987. Microhabitat

use in a mediterranean riverine fish assemblage - Fishes of the upper Matarraña.

Oecologia, 73(4): 501-512.

Growns, I.O., Davis, J.A., 1994. Longitudinal changes in near-bed flows and

macroinvertebrate communities in a Western Australian stream. Journal of the North

American Benthological Society, 13(4): 417-438.

Guay, J.C., Boisclair, D., Rioux, D., Leclerc, M., Lapointe, M., Legendre, P., 2000.

Development and validation of numerical habitat models for juveniles of atlantic

salmon (Salmo salar). Canadian Journal of Fisheries and Aquatic Sciences, 57(10):

2065-2075.

Guisan, A., Zimmermann, N.E., 2000. Predictive habitat distribution models in ecology.

Ecological Modelling, 135(2-3): 147-186.

Hajmeer, M., Basheer, I., 2002. A probabilistic neural network approach for modeling

and classification of bacterial growth/no-growth data. Journal of Microbiological

Methods, 51(2): 217–226.

Harrell, F.E., 2010. Hmisc: Harrell Miscellaneous. R package version 3.8-3.

http://CRAN.R-project.org/package=Hmisc.

Harris, R.R., 1986. Occurrence patterns of riparian plants and their significance to

water resource development. Biological Conservation, 38(3): 273-286.


121

WP3: MORPH

Hauer, C., Unfer, G., Graf, W., Leitner, P., Zeiringer, B., Habersack, H., 2012. Hydro-

morphologically related variance in benthic drift and its importance for numerical

habitat modelling. Hydrobiologia, 683(1): 83-108.

Hayes, J.W., Jowett, I.G., 1994. Microhabitat models of large drift-feeding brown trout

in three New Zealand rivers. North American Journal of Fisheries Management,

14(4): 710-725.

Heggenes, J., 1990. Habitat utilization and preferences in juvenile Atlantic salmon

(Salmo salar) in streams. Regulated Rivers - Research & Management, 5(4): 341-

354.

Heggenes, J., 1991. Comparisons of habitat availability and habitat use by an allopatric

cohort of juvenile Atlantic salmon Salmo salar under conditions of low competition in

a Norwegian stream. Holarctic Ecology, 14(1): 51-62.

Heggenes, J., 1996. Habitat selection by brown trout (Salmo trutta) and young atlantic

salmon (S. salar) in streams: static and dynamic hydraulic modelling. Regulated

Rivers: Research & Management, 12(2–3): 155–169.

Heggenes, J., Brabrand, Å., Saltveit, S., 1990. Comparison of Three Methods for

Studies of Stream Habitat Use by Young Brown Trout and Atlantic Salmon.

Transactions of the American Fisheries Society, 119(1): 101-111.

Hupp, C.R., Osterkamp, W.R., 1985. Bottomland vegetation distribution along Passage

Creek, Virginia, in relation to fluvial landforms. Ecology, 66(3): 670-681.

Inamdar, S.P., Sheridan, J.M., Williams, R.G., Bosch, D.D., Lowrance, R.R., Altier,

L.S., Thomas, D.L., 1999. Riparian ecosystem management model (REMM): I.

Testing of the hydrologic component for a coastal plain riparian system. American

Society of Agricultural Engineers, 42(6): 1679-1689.

Johnson, H.D., 1980. The comparison of usage and availability measurement for

evaluating resource preference. Ecology, 61(1): 65-71.

Johnsson, J.I., Carlsson, M., Sundstrom, L.F., 2000. Habitat preference increases

territorial defence in brown trout (Salmo trutta). Behavioral Ecology and

Sociobiology, 48(5): 373-377.

Jorde, K., Schneider, M., Peter, A., Zoellner, F., 2001. Fuzzy based Models for the

Evaluation of Fish Habitat Quality and Instream Flow Assessment. Proceedings of

the 2001 International Symposium on Environmental Hydraulics. Tempe, Arizona.


122

WP3: MORPH

Jowett, I.G., 1992. Models of the Abundance of Large Brown Trout in New Zealand

Rivers. North American Journal of Fisheries Management, 12(3): 417-432.

Jowett, I.G., 2003. Hydraulic constraints on habitat suitability for benthic invertebrates

in gravel-bed rivers. River Research and Applications, 19(5-6): 495-507.

Jowett, I.G., Davey, A.J.H., 2007. A comparison of composite habitat suitability indices

and generalized additive models of invertebrate abundance and fish presence-

habitat availability. Transactions of the American Fisheries Society, 136(2): 428-444.

Jowett, I.G., Duncan, M.J., 2011. Effectiveness of 1D and 2D hydraulic models for

instream habitat analysis in a braided river. Ecological Engineering(0).

Jowett, I.G., Richardson, J., Biggs, B.J.F., Hickey, C.W., Quinn, J.M., 1991.

Microhabitat preferences of benthic invertebrates and the development of

generalised Deleatidium spp. habitat suitability curves, applied to four New Zealand

rivers. New Zealand Journal of Marine & Freshwater Research, 25(2): 187-199.

Jowett, I.G., Richardson, J., McDowall, R.M., 1996. Relative effects of in‐stream habitat

and land use on fish distribution and abundance in tributaries of the Grey River, New

Zealand. New Zealand Journal of Marine and Freshwater Research, 30(4): 463-475.

Karr, J.R., 1981. Assessment of biotic integrity using fish communities. Fisheries, 6(6):

21–27.

Karrenberg, S., Edwards, P.J., Kollmann, J., 2002. The life history of Salicaceae living

in the active zone of floodplains. Freshwater Biology, 47(4): 733-748.

Kasabov, N.K., 1998. Foundations of Neural Networks, Fuzzy Systems, and

Knowledge Engineering. Cambridge, MA: MIT Press.

King, A.A., 2008. Subplex optimization algorithm. Package 'subplex'. R package,

version 1.1–3. http://www.cran.r-project.org/web/packages/subplex/subplex.pdf.

Korman, J.P., C.J. Lekstrum, T., 1994. A Guide for the Selection of Standard Methods

for Quantifying Sportfish Habitat Capability and Suitability in Streams and Lakes of

B.C., Vancouver, B.C.: Environment, Fisheries Branch, Research and Development

Section.


123

WP3: MORPH

Lambert, T.R., Hanson, D.F., 1989. Development of habitat suitability criteria for trout in

small streams. Regulated Rivers - Research & Management, 3(1-4): 291-303.

Lek, S., Delacoste, M., Baran, P., Dimopoulos, I., Lauga, J., Aulagnier, S., 1996.

Application of neural networks to modelling nonlinear relationships in ecology.

Ecological Modelling, 90(1): 39-52.

Lite, S.J., Bagstad, K.J., Stromberg, J.C., 2005. Riparian plant species richness along

lateral and longitudinal gradients of water stress and flood disturbance, San Pedro

River, Arizona, USA. Journal of Arid Environments, 63(4): 785-813.

Louhi, P., Mäki-Petäys, A., Erkinaro, J., 2008. Spawning habitat of Atlantic salmon and

brown trout: General criteria and intragravel factors. River Research and

Applications, 24(3): 330-339.

Lütolf, M., Kienast, F., Guisan, A., 2006. The ghost of past species occurrence:

Improving species distribution models for presence-only data. Journal of Applied

Ecology, 43(4): 802–815.

Lyon, J., Sagers, C., 1998. Structure of herbaceous plant assemblages in a forested

riparian landscape. Plant Ecology, 138(1): 1-16.

MacKenzie, D.I., Nichols, J.D., Hines, J.E., Knutson, M.G., Franklin, A.B., 2003.

Estimating site occupancy, colonization, and local extinction when a species is

detected imperfectly. Ecology 84, 2200-2207.

MAGRAMA, 2008. Orden ARM/2656/2008, de 10 de septiembre, por la que se

aprueba la instrucción de planificación hidrológica.

Mahoney, J.M., Rood, S.B., 1998. Streamflow requirements for cottonwood seedling

recruitment – An integrative model. Wetlands, 18(4): 634-645.

Malanson, G.P., 1993. Riparian landscapes. Cambridge: Cambridge University Press.

Mamdani, E.H., 1974. Application of fuzzy algorithms for control of simple dynamic

plant. Proceedings of the Institution of Electrical Engineers, 121(12): 1585-1588.

Manel, S.C.W.H.C.W., H. Ormerod, S. J., 2001. Evaluating presence-absence models

in ecology: The need to account for prevalence. Journal of Applied Ecology, 38(5):

921-931.

Martínez-Capel, F., Costa, R.M.S., Muñoz-Mas, R., Alcaraz-Hernández, J.D.,

Hernández-Mascarell, A.B., Patiño Rodríguez, J.H., Garófano-Gómez, V., 2009a.

Factores de degradación de las poblaciones de loina (chondrostoma arrigonis) y el

estado de su hábitat actual en la cuenca del río júcar (2006-2008), Gandia


124

WP3: MORPH

(València). Institut d'Investigació per a la Gestió Integrada de Zones Costaneres

(IGIC) . Universitat Politècnica de Valè.ncia.

Martínez-Capel, F., García de Jalón, D., 1999. Desarrollo de Curvas de preferencia de

microhábitat para Leuciscus pyrenaicus y Barbus bocagei por buceo en el río

Jarama (Cuenca del Tajo). Limnetica, 17: 4.

Martínez-Capel, F., García De Jalón, D., Werenitzky, D., Baeza, D., Rodilla-Alamá, M.,

2009b. Microhabitat use by three endemic Iberian cyprinids in Mediterranean rivers

(Tagus River Basin, Spain). Fisheries Management and Ecology, 16(1): 52–60.

Martínez-Capel, F.P., M. Hernández-Mascarell, AB. Munné, T., 2008. Nose velocity

calculation for spatial analysis of habitat and environmental flow assessments. 4th

ECRR International Conference on River Restoration, Venecia (Italia): ECRR

European Centre for River Restoration.

Mas-Martí, E., García-Berthou, E., Sabater, S., Tomanova, S., Muñoz, I., 2010.

Comparing fish assemblages and trophic ecology of permanent and intermittent

reaches in a Mediterranean stream. Hydrobiologia, 657(1): 167-180.

Mérigoux, S., Dolédec, S., 2004. Hydraulic requirements of stream communities: A

case study on invertebrates. Freshwater Biology, 49(5): 600-613.

Mérigoux, S., Lamouroux, N., Olivier, J.M., Dolédec, S., 2009. Invertebrate hydraulic

preferences and predicted impacts of changes in discharge in a large river.

Freshwater Biology, 54(6): 1343-1356.

Merritt, D.M., Scott, M.L., LeRoy Poff, N., Auble, G.T., Lytle, D.A., 2010. Theory,

methods and tools for determining environmental flows for riparian vegetation:

riparian vegetation-flow response guilds. Freshwater Biology, 55(1): 206-225.

Michalewicz, Z., Fogel, D.B., 2000. How to Solve It: Modern Heuristics. Springer-

Verlag, Berlin, Heilderberg: 467.

Mood, A.M., Graybill, F.A., C., B.D., 1974. Introduction to the Theory of Statistics New

York: McGraw-Hill.

Moore, D.R.J., Breton, R.L., MacDonald, D.B., 2003. A comparison of model

performance for six quantitative structure-activity relationship packages that predict

acute toxicity to fish. Environmental Toxicology and Chemistry, 22(8): 1799–1809.

Morante, T., García-Arberas, L., Antón, A., Rallo, A., 2012. Macroinvertebrate biomass

estimates in Cantabrian streams and relationship with brown trout (Salmo trutta)

populations. Limnetica, 31(1): 85-94.


125

WP3: MORPH

Morin, A., Harper, P.P., Peters, R.H., 1986. Microhabitat-preference curves of blackfly

larvae ( Diptera: Simuliidae): a comparison of three estimation methods. Canadian

Journal of Fisheries and Aquatic Sciences, 43(6): 1235-1241.

Mouton, A., Schneider, M., Goethals, P.L., De Pauw, N., 2007. Simulation of the

physical habitat suitability of bullhead (Cottus gobio L.) in the Zwalm River.

Communications in agricultural and applied biological sciences, 72(1): 75-77.

Mouton, A.M., Alcaraz-Hernández, J.D., De Baets, B., Goethals, P.L.M., Martínez-

Capel, F., 2011. Data-driven fuzzy habitat suitability models for brown trout in

Spanish Mediterranean rivers. Environmental Modelling and Software, 26(5): 615-

622.

Mouton, A.M., De Baets, B., Goethals, P.L.M., 2009a. Knowledge-based versus data-

driven fuzzy habitat suitability models for river management. Environmental

Modelling and Software.

Mouton, A.M., De Baets, B., Van Broekhoven, E., Goethals, P.L.M., 2009b.

Prevalence-adjusted optimisation of fuzzy models for species distribution. Ecological

Modelling, 220(15): 1776–1786.

Mouton, A.M., De Baets, B., Peter, A., Holzer, G., Müller, R., Goethals, P.L.M., 2008.

Entropy-based fuzzy set optimisation for reducing ecological model complexity.

Sànchez-Marrè, M., Béjar, J., Comas, J., Rizzoli, A., Guariso, G. (Eds.), iEMSs

Fourth Biennial Meeting: “Summit on Environmental Modelling and Software”.

International Modelling and Software Society/CD-ROM Publication, Barcelona,

Spain.

Mouton, A.M., Dedecker, A.P., Lek, S., Goethals, P.L.M., 2010. Selecting variables for

habitat suitability of Asellus (Crustacea, Isopoda) by applying input variable

contribution methods to artificial neural network models. Environmental Modeling

and Assessment, 15(1): 65-79.

Mouton, A.M., Jowett, I., Goethals, P.L.M., De Baets, B., 2009c. Prevalence-adjusted

optimisation of fuzzy habitat suitability models for aquatic invertebrate and fish

species in New Zealand. Ecological Informatics, 4(4): 215-225.

Mouton, A.M., Schneider, M., Peter, A., Holzer, G., Müller, R., Goethals, P.L.M., De

Pauw, N., 2008. Optimisation of a fuzzy physical habitat model for spawning

European grayling (Thymallus thymallus L.) in the Aare river (Thun, Switzerland).



126

WP3: MORPH

Moyle, P.B., 2002. Inland Fishes of California. Berkeley, California: University of

California Press.

Moyle, P.B., Baltz, D.M., 1985. Microhabitat Use by an Assemblage of California

Stream Fishes: Developing Criteria for Instream Flow Determinations. Transactions

of the American Fisheries Society, 114(5): 695-704.

Muñoz-Mas, R., Martínez-Capel, F., Schneider, M., Mouton, A.M., 2012. Assessment

of brown trout habitat suitability in the Jucar River Basin (SPAIN): Comparison of

data-driven approaches with fuzzy-logic models and univariate suitability curves.

Science of the Total Environment, 440: 123-131.

Naiman, R.J., Bechtold, J.S., Drake, D.C., Latterell, J.J., O’Keefe, T.C., Balian, E.V.,

2006. Origins, patterns, and importance of heterogeneity in riparian systems. In:

G.M. Lovett, M.G. Turner, C.G. Jones and K.C. Weathers (Editors), Ecosystem

function in heterogeneous landscapes. New York (USA): Springer Science and

Business Media, Inc, pp. 279-310.

Nakamura, F., Yajima, T., Kikuchi, S., 1997. Structure and composition of riparian

forests with special reference to geomorphic site conditions along the Tokachi River,

northern Japan. Plant Ecology, 133(2): 209-219.

Nilsson, C., 1986. Change in riparian plant community composition along two rivers in

northern Sweden. Canadian Journal of Botany, 64(3): 589-592.

Nilsson, C., Ekblad, A., Dynesius, M., Backe, S., Gardfjell, M., Carlberg, B., Hellqviist,

S., Jansson, R., 1994. A Comparison of species richness and traits of riparian plants

between a main river channel and its tributaries. Journal of Ecology, 82(2): 281-295.

Oksanen, J., Blanchet, F.G., Kindt, R., Legendre, P., O'Hara, R.B., Simpson, G.L.,

Solymos, P., M. Henry H., S., Wagner, H., 2011. Vegan: Community Ecology

Package. R package version 1.17-6.

Orth, D.J., Maughan, O.E., 1982. Evaluation of the incremental methodology for

recommending instream flows for fishes. Transactions of the American Fisheries

Society, 111(4): 413–445.

Orth, D.J., Maughan, O.E., 1983. Microhabitat preferences of benthic fauna in a

woodland stream. Hydrobiologia, 106(2): 157-168.

Parzen, E., 1962. On estimation of a probability density function and mode. The Annals

of Mathematical Statistics, 33(3): 1065-1076.


127

WP3: MORPH

Pastuchová, Z., Lehotský, M., Grešková, A., 2008. Influence of morphohydraulic

habitat structure on invertebrate communities (Ephemeroptera, Plecoptera and

Trichoptera). Biologia, 63(5): 720-729.

Payne, T.R.A., Mark A., 2009. Application of the use-to-availability electivity ratio for

developing habitat suitability criteria in PHABSIM instream flow studies. Proceedings

of the 7th International Symposium on Ecohydraulics, Concepción, Chile, January

12-16, 2009.

Perea, S., Garzón, P., González, J.L., Almada, V.C., Pereira, A., Doadrio, I., 2011.

New distribution data on Spanish autochthonous species of freshwater fish. Nuevos

datos de distribución de especies autóctonas de peces de agua dulce de españa,

67(1): 91-102.

Poff, N.L., Allan, J.D., Bain, M.B., Karr, J.R., Prestegaard, K.L., Richter, B.D., Sparks,

R.E., Stromberg, J.C., 1997. The natural flow regime: A paradigm for river

conservation and restoration. BioScience, 47(11): 769-784.

Quinn, J.M., Hickey, C.W., 1994. Hydraulic parameters and benthic invertebrate

distributions in two gravel-bed New Zealand rivers. Freshwater Biology, 32(3): 489-

500.

R Development Core Team, 2010. R: A language and environment for statistical

computing. Vienna, Austria.: R Foundation for Statistical Computing.

Raleigh, R.F.Z., Laurence D.; Nelson, Patrick C., 1984. Habitat Suitability Index Models

and Instream Flow Suitability Curves: Brown Trout: U.S. Fish and Wildlife Service.

Reyjol, Y., Lim, P., Belaud, A., Lek, S., 2001. Modelling of microhabitat used by fish in

natural and regulated flows in the river Garonne (France). Ecological Modelling,

146(1–3): 131–142.

Rincon, P.A., Lobon-Cervia, J., 1993. Microhabitat use by stream-resident brown trout:

bioenergetic consequences. Transactions - American Fisheries Society, 122(4):

575-587.

Rowan, T.H., 1990. Functional stability analysis of numerical algorithms. Doctoral

Dissertation, Austin, Texas: University of Texas.

Rubec, P.J., J.C.W. Bexley, H. Norris, M.S. Coyne, M.E. Monaco, S.G. Smith, Ault

J.S., 1999. Suitability modeling to delineate habitat essential to sustainable curves:

white sucker. Fish and Wildlife Service. U.S. Department of the Interior. Washington,

DC.: 56.


128

WP3: MORPH

Rust, S.W., Filgner, M.A., 1984. A modification of the kruskal-wallis statistic for the

generalized behrens-fisher problem. Communications in Statistics - Theory and

Methods 13(16): 2013-2027.

Salo, R., Hakkinen, I., Makinen, Y., Miemela, P., Puhakka, M., Coley, P., 1986. River

dynamics and the diversity of Amazon lowland forest. Nature, 332: 254-258.

Savage, R.E., 1931. The relation between the feeding of the herring off the east coast

of England and the plankton of the surrounding waters. Fishery Investigation,

Ministry of Agriculture, Food and Fisheries, Series 2., 12: 88.

Schindler, D.E., Hilborn, R., Chasco, B., Boatright, C.P., Quinn, T.P., Rogers, L.A.,

Webster, M.S., 2010. Population diversity and the portfolio effect in an exploited

species. Nature, 465(7298): 609-612.

Scruton, D.A., Ollerhead, L.M.N., Clarke, K.D., Pennell, C., Alfredsen, K., Harby, A.,

Kelley, D., 2003. The behavioural response of juvenile Atlantic salmon (Salmo salar)

and brook trout (Salvelinus fontinalis) to experimental hydropeaking on a

Newfoundland (Canada) River. River Research and Applications, 19(5-6): 577-587.

Unknown author, 1981. Standards for the Development of Habitat Suitability Index

Models. U.S. Fish and Wildlife Service. Washington, DC.

Shafroth, P., Auble, G., Stromberg, J., Patten, D., 1998. Establishment of woody

riparian vegetation in relation to annual patterns of streamflow, Bill Williams River,

Arizona. Wetlands, 18(4): 577-590.

Shannon, C.E., Weaver, W., 1963. Mathematical theory of communication. University

of Illinois Press. Urbana, Illinois: University of Illinois Press.

Shannon, J.P., Blinn, D.W., Benenati, P.L., Wilson, K.P., 1996. Organic drift in a

regulated desert river. Canadian Journal of Fisheries and Aquatic Sciences, 53(6):

1360-1369.

Simberloff, D., Dayan, T., 1991. The guild concept and the structure of ecological

communities. Annual Review of Ecology and Systematics, 22(1): 115-143.

Slater, F.M., Curry, P., Chadwell, C., 1987. A practical approach to the evaluation of

the conservation status of vegetation in river corridors in Wales. Biological

Conservation, 40(1): 53-68.

Specht, D.F., 1989. Probabilistic neutral networks (a one-pass learning method) and

potential applications. Western Electronic Show and Convention (WESCON), San

Francisco, California: Western Periodicals Conference, pp. 780-785.


129

WP3: MORPH

Specht, D.F., 1990. Probabilistic neural networks. Neural Networks, 3(1): 109–118.

Statzner, B., Gore, J.A., Resh, V.H., 1988. Hydraulic Stream Ecology: Observed

Patterns and Potential Applications. Journal of the North American Benthological

Society, 7(4): 53.

Statzner, B., Müller, R., 1989. Standard hemispheres as indicators of flow

characteristics in lotic benthos research. Freshwater Biology, 21(3): 445-459.

Strahler, A.N., 1952. Hypsometric (area-altitude) analysis of erosional topography.

Geological Society of America Bulletin, 63: 25.

Stromberg, J.C., 1993. Frémont cottonwood-goodding willow riparian forests: A review

of their ecology, threats, and recovery potential. Arizona-Nevada Academy of

Science, 27(1): 97-110.

Stromberg, J.C., Patten, D.T., 1996. Instream flow and cottonwood growth in the

eastern Sierra Nevada of California, USA. Regulated Rivers: Research &

Management, 12(1): 1-12.

Tabacchi, E., Planty-Tabacchi, A.-M., Salinas, M.J., Décamps, H., 1996. Landscape

structure and diversity in riparian plant communities: a longitudinal comparative

study. Regulated Rivers: Research & Management, 12(4-5): 367-390.

Terrell, J.W., 1984. Proceedings of a workshop on fish habitat suitability index models.

Fish and Wildlife Service, U.S. Department of the Interior Biological Report, 85(6):

393.

Thomas, J.A., Bovee, K.D., 1993. Application and testing of a procedure to evaluate

transferability of habitat suitability criteria. Regulated Rivers: Research &

Management, 8(3): 285–294.

Titus, R.G., 1990. Territorial behavior and its role in population regulation of young

brown trout (Salmo trutta): new perspectives. Annales Zoologici Fennici, 27(2): 119-

130.

Turgeon, K., Rodríguez, M.A., 2005. Predicting microhabitat selection in juvenile

Atlantic salmon Salmo salar by the use of logistic regression and classification trees.

Freshwater Biology, 50(4): 539-551.

U.S. Fish and Wildlife Service - Unknown author, 1981. Standards for the Development

of Habitat Suitability Index Models. Washington, DC.


130

WP3: MORPH

Vadas R.L, Jr., Orth, D.J., 2001. Formulation of habitat suitability models for stream

fish guilds: Do the standard methods work? Transactions of the American Fisheries

Society, 130(2): 217-235.

Van Broekhoven, E., Adriaenssens, V., De Baets, B., 2007. Interpretability-preserving

genetic optimization of linguistic terms in fuzzy models for fuzzy ordered

classification: An ecological case study. International Journal of Approximate

Reasoning, 44(1): 65-90.

Van Broekhoven, E., Adriaenssens, V., De Baets, B., Verdonschot, P.F.M., 2006.

Fuzzy rule-based macroinvertebrate habitat suitability models for running waters.


Van Coller, A., Rogers, K., Heritage, G., 1997. Linking riparian vegetation types and

fluvial geomorphology along the Sabie River within the Kruger National Park, South

Africa. African Journal of Ecology, 35(3): 194-212.

Vismara, R., Azzellino, A., Bosi, R., Crosa, G., Gentili, G., 2001. Habitat suitability

curves for brown trout (Salmo trutta fario L.) in the River Adda, Northern Italy:

Comparing univariate and multivariate approaches. River Research and

Applications, 17(1): 37-50.

Voos, K.A., 1981. Simulated Use of the Exponential Polynomial/maximum Likelihood

Technique in Developing Suitability of Use Functions for Fish Habitat. Logan: Utah

State University. Department of Civil and Environmental Engineering.

Wallace, J.B., Webster, J.R., 1996. The role of macroinvertebrates in stream

ecosystem function, p. 115-139.

Wańkowski, J.W.J., Thorpe, J.E., 1979. Spatial distribution and feeding in atlantic

salmon, Salmo salar L. juveniles. Journal of Fish Biology, 14(3): 239-247.

Waters, B.F., 1976. A Methodology for Evaluating the Effects of Different Streamflows

on Salmonid Habitat - Proceedings of the Symposium and Specialty Conference on

Instream Flow Needs. Bathesda, MD: American Fisheries Society, p. 13.

Welch, B.L., 1951. On the comparison of several mean values: An alternative

approach. Biometrika, 38(3/4): 6.

Zadeh, L.A., 1965. Fuzzy sets. Information and Control, 8(3): 338-353.

Zhong, M., Hecker, J., Maidhoff, I., Shibly, P., Georgiopoulos, M., Anagnostopoulos,

G., Mollaghasemi, M., 2005. Probabilistic neural network: comparisons of the cross-


131

WP3: MORPH

validation approach and a fast heuristic to choose the smoothing parameters, 2005

Artificial Neural Networks in Engineering. St. Louis, MI.

Deliverable 3.5 final - CSIC · Figure 38. A) Data-driven fuzzy sets obtained from the Shannon-Waver entropy-based optimization for the standardized density in Ephemeroptera, Trichoptera,

Documents