Ana səhifə

Bioivers I ty rese arch


Yüklə 1.2 Mb.
səhifə1/3
tarix24.06.2016
ölçüsü1.2 Mb.
  1   2   3

BIOIVERS I TY RESE ARCH

Integrating detection probabilities in species distribution models of amphibians breeding in Mediterranean temporary ponds
Carola Go´ mez-Rodrı´guez1*, Javier Bustamante1,2, Carmen Dı´az-Paniagua1

and Antoine Guisan3




1Department of Wetland Ecology, Estacio´n Biolo´gica de Don˜ana, CSIC, C/Ame´rico Vespucio s/n 41092-Sevilla, Spain, 2Remote Sensing and GIS Lab (LAST-EBD), Estacio´n Biolo´gica de Don˜ana, CSIC, C/Ame´rico Vespucio s/n 41092- Sevilla, Spain,

3Department of Ecology and Evolution,

University of Lausanne, CH-1015 Lausanne, Switzerland

*Correspondence: Carola Go´ mez-Rodr´ıguez, Estacio´ n Biolo´ gica de Don˜ ana, CSIC,

C/Ame´ rico Vespucio s/n 41092-Sevilla, Spain. E-mail: carola@ebd.csic.es

ABSTRACT
Aim The imperfect detection of species may lead to erroneous conclusions about species–environment relationships. Accuracy in species detection usually requires temporal replication at sampling sites, a time-consuming and costly monitoring scheme. Here, we applied a lower-cost alternative based on a double-sampling approach to incorporate the reliability of species detection into regression-based species distribution modelling.


Location Don˜ana National Park (south-western Spain).
Methods Using species-specific monthly detection probabilities, we estimated the detection reliability as the probability of having detected the species given the species-specific survey time. Such reliability estimates were used to account explicitly for data uncertainty by weighting each absence. We illustrated how this novel framework can be used to evaluate four competing hypotheses as to what constitutes primary environmental control of amphibian distribution: breeding habitat, aestivating habitat, spatial distribution of surrounding habitats and/or major ecosystems zonation. The study was conducted on six pond-breeding amphibian species during a 4-year period.
Results Non-detections should not be considered equivalent to real absences, as their reliability varied considerably. The occurrence of Hyla meridionalis and Triturus pygmaeus was related to a particular major ecosystem of the study area, where suitable habitat for these species seemed to be widely available. Characteristics of the breeding habitat (area and hydroperiod) were of high importance for the occurrence of Pelobates cultripes and Pleurodeles waltl. Terrestrial characteristics were the most important predictors of the occurrence of Discoglossus galganoi and Lissotriton boscai, along with spatial distribution of breeding habitats for the last species.
Main conclusions We did not find a single best supported hypothesis valid for all species, which stresses the importance of multiscale and multifactor approaches. More importantly, this study shows that estimating the reliability of non- detection records, an exercise that had been previously seen as a na¨ıve goal in species distribution modelling, is feasible and could be promoted in future studies, at least in comparable systems.
Keywords

Absence reliability, data uncertainty, false absence, species detectability, temporary ponds.








INTRODUCTIO N


Species distribution models (SDMs) attempt to quantify species–environment relationships, a central issue in ecology and conservation (Guisan & Zimmermann, 2000). A critical issue for the utility and validity of any model is the reliability of the data used to build it (Lobo, 2008). For a mobile organism, the recorded presence is usually the only reliable distribution information (Guisan & Thuiller, 2005). While the presence of a species is confirmed by simply detecting it at a site, it is usually not possible to confirm if an animal was truly absent or if the species was present but not detected during the survey (MacKenzie et al., 2006).

Few studies have tried to draw attention to the funda- mental issue of detection reliability in SDMs (but see Gu & Swihart, 2004; Lobo, 2008; Lobo et al., 2010). Although data quality is critical for model performance in general (Foody,

2011), the need for an increased awareness in this source of errors lies in the fact that the imperfect detection of species may lead to erroneous conclusions about species–environ- ment relationships (Gu & Swihart, 2004; Mazerolle et al.,

2005; MacKenzie, 2006). If lack of absence records is a main source of modelling error (Barry & Elith, 2006), a worse scenario is building a presence–absence model in which absence records do not represent unfavourable sites but are just a result of inaccurate sampling (methodological absences, sensu Lobo et al., 2010).

The optimal modelling approach to overcome this source of error is to evaluate species–habitat relationships while explicitly accounting for the probability of detecting the species when present (MacKenzie et al., 2006). When species detection is imperfect, these site-occupancy models are better for predicting species occurrence than more traditional regression analyses (Ke´ ry et al., 2010; Rota et al., 2011). However, this modelling technique is not being widely used by species distribution modellers yet (but see Urban & Swihart, 2009; Adams et al.,

2010; or Martin et al., 2010 as some recent examples).

A main statistical limitation of occupancy models is that they require temporal replication at all sampled sites, a condition that may not be always easy to fulfil. An interme- diate approach can be to conduct a double-sampling design, consisting in estimating detection probabilities from the data collected at few sites, where repeated surveys were conducted, and then applying this information to the sites surveyed only once (MacKenzie et al., 2006). Following this approach, here we propose that one can integrate information from species detectability at a site, as a surrogate for the reliability of the absence record, into traditional presence–absence models. This surrogacy relies on the premise that the higher the probability of having detected the species when present at a site, the higher the reliability of the absence record. So, with a low-cost approach, it would be easy to select those absence records that really represent sites not occupied by the species and, thus, that are supposed to be unsuitable habitats.

Amphibians are inconspicuous organisms (Mazerolle et al.,

2007) for which the probability of detecting a species with a

single visit may be low, species-specific and variable over time (Go´ mez-Rodrı´guez et al., 2010d). Moreover, the reliability of absence data is expected to be highly limited in species breeding in temporary ponds as a result of interannual turnover in assemblage composition (Jakob et al., 2003; Go´ mez-Rodrı´guez et al., 2010c). Thus, any yearly survey will probably yield many ‘false absences’ since data from several breeding seasons would be needed to characterize the species assemblage associated with a given pond. Previous studies have quantified the relationships between habitat characteristics and amphibian richness, species occurrence or species relative abundance in temporary ponds (e.g. Beja & Alcazar, 2003; Richter-Boix et al., 2007). However, to our knowledge, no study has explicitly accounted for the reliability of absence data.

Most ecological models about amphibian habitat selection focus on four main aspects that have been identified as critical for amphibian ecology:

1. Abiotic characteristics of the breeding habitat, such as pond area or hydroperiod (i.e. annual duration of aquatic phase in temporary waters). Amphibian species are supposed to be sorted along the hydroperiod gradient according to whether they are able to metamorphose in short-duration ponds or tolerant to the presence of major predators in ponds of longer duration (Wellborn et al., 1996). The relationship between species occurrence and pond size is twofold. First, metapop- ulation theory predicts that the probability of occurrence would increase with pond size because it assumes a functional relationship between the area of a patch and its extinction probability (Hanski, 1998). Second, patch area and habitat heterogeneity are highly interconnected (Rosenzweig, 1995), and the latter provides more niches and diverse ways of exploiting the environmental resources (Tews et al., 2004).

2. Biotic interactions in the breeding habitat, such as compe- tition or predation (e.g. Wells, 2007).

3. Characteristics of the aestivating habitat (i.e. landscape composition) since terrestrial habitats provide refuges for amphibian species during the dry season and also constitute the matrix that interconnects ponds (Gibbons, 2003).

4. Spatial structure of the habitat patches (i.e. distance to nearest site, density of surrounding ponds, etc.), which determines the dispersal or regular movements of individuals among ponds (Smith & Green, 2005).

Here, we develop a novel approach to show how detection probabilities can be incorporated in SDMs. Using data from a double-sampling design, we demonstrate a method to account for the reliability of non-detection records, which can be used as weight in SDMs. We illustrate this approach with amphib- ians breeding in a system of temporary ponds in Don˜ana National Park (DNP), in south-western Spain. For each pond where a species was not detected, we used the single-visit probability of detection computed from a different survey (Go´ mez-Rodrı´guez et al., 2010d) to estimate the reliability of each non-detection record, taking into account the history of pond surveys (number and date of sampling visits). We test four competing hypotheses of which environmental factors are






correlated with amphibian distribution in DNP: (1) charac- teristics of their breeding habitat, (2) characteristics of their aestivating habitat, (3) spatial distribution of breeding habitat patches, and (4) general local characteristics of the major ecosystems in DNP.

METH ODS Study area


The study was conducted in the aeolian sands ecoregion of

DNP in south-western Spain (Fig. 1, see Siljestro¨ m et al.,

1994). DNP is considered to be one of the largest and most important wetlands in southern Europe. Within this region, Montes et al. (1998) differentiated eight ecosections based on differences in their geomorphologic, stratigraphic and hydro- dynamic characteristics (Fig. 1). Many temporary ponds of natural origin are located amid relatively small topographic depressions and flood during the rainy season. The area also includes two natural large permanent ponds and small artificial permanent water bodies (maintained for watering cattle and

locally named zacallones). Pond size is largely variable, from rain puddles (several square metres) to large temporary ponds (> 1 ha). Hydroperiod varies among ponds and years, from one to 10 months (Go´ mez-Rodrı´guez et al., 2009; D´ıaz- Paniagua et al., 2010). Many pond basins are completely or partially enclosed by a fringe of dense vegetation mainly composed of Erica scoparia L., E. ciliaris L., Calluna vulgaris (L.) Hull and Ulex minor Roth. A detailed description of DNP temporary ponds can be found in the studies by Go´ mez- Rodr´ıguez et al. (2009) and D´ıaz-Paniagua et al. (2010).

Field sampling
Amphibian sampling
We sampled 221 amphibian breeding sites (169 natural ponds and 52 zacallones) located in seven different ecosections (Fig. 1). Amphibian data were collected during the breeding season in a 4-year survey (from 2003 to 2006) (Table 1). Some ponds did not flood in 2005 and could therefore not be sampled. We could not monitor all ponds every year. A total of


(a)

(b)

(c)



Figure 1 (a) Location of Don˜ana National Park, (DNP), in south-western Spain. (b) Location of the study ponds. The map also shows the different ecosections within the aeolian sands in DNP. Note that ecosection number eight (terrestrial human-transformed areas) consists on isolated and small locations, not visible in this graphical representation. (c) Presence–absence data of each species in the study area. Absences are weighted according to their reliability, as obtained from the probability of detection after all the sampling visits conducted in that pond. Absence records with null reliability are not shown.

Ecosection code: 1: Ecotone marshes-stabilized sands; 2: Dry stabilized sands at higher elevation; 3: Wet stabilized sands at higher elevation; 4: Stabilized sands at low elevation; 5: Semi-stabilized dunes; 6: Mobile dunes; 7: Former beaches.






Table 1 Number of ponds surveyed during each breeding season (sampling period indicated in brackets) and mean number of sampling visits (total number of ponds surveyed over the entire study period = 221).
2003 2004 2006 (February–May) (January–June) (March–May)

et al., 1996) and pond size (Werner et al., 2007). Hydroperiod was categorized in four wide groups because a ranked ordination of ponds hydroperiod is similar both in wet and dry years (Go´ mez-Rodrı´guez et al., 2009). Since most ponds were visited only once, hydroperiod categories were based on characteristics related to flooding duration such as the presence of particular plant species (i.e. four main groups of aquatic plants can be differentiated according to their water



Total number of

ponds surveyed Number of visits per pond



(mean ± SD)

94 95 129


1.5 ± 1.1 1.7 ± 1.4 1.1 ± 0.4

dependence: floodplain species, wetland species, anchored species and free-floating species, see Dı´az-Paniagua et al.,

2010), as well as basin topography (i.e. pond depth, basin slope, etc.) and past recordings of hydroperiod in those ponds

Number of ponds surveyed

Only that year 29 61 50

All years 16 16 16

In 2003 and 2004 2 2

In 2004 and 2006 16 16

In 2003 and 2006 47 47
The number of ponds is detailed as number of ponds visited only in that season, in all seasons or in two of the three seasons.

140 ponds were visited in only one of the 3 years whereas 16 ponds were visited every year (Table 1). In 2006, a year with scarce autumn rainfall, we visited all ponds monitored in the previous years and surveyed the flooded ones (n = 129). Most ponds were visited once per year except 19 ponds, which were sampled monthly during the whole amphibian breeding seasons to compute the species monthly detection probabilities (see Go´ mez-Rodrı´guez et al., 2010d). The number of visits to a given pond ranged from 1 to 12 (mean = 2.01 ± 2.42 SD).

We used dipnetting techniques (Heyer et al., 1994) to sample the amphibian larvae. We identified in situ the individuals captured in each sampling unit (three consecutive sweeps on a stretch of c. 1.5 m length) and then released them back. Sampling effort was proportional to pond size, except when not logistically achievable because of the large size of the water body, in which case we tried to sample all different pond microhabitats. Larval sampling was complemented with visual surveys in and around the pond to detect eggs, larvae and metamorphic individuals. Since this study analyses the habitat requirements for amphibian breeding, we only included data from breeding attempts, not just the occurrence of a species. So we excluded the contingent detection of adults or calling activity because the sampling protocol was not optimized for detection of this life stage (Heyer et al., 1994).

Predictors and underlying hypotheses


We selected habitat variables to test the competing hypotheses regarding amphibian habitat selection, based on available ecological information. The habitat variables can be grouped into the following sets (see Table 2):

Breeding habitat. We recorded two major structuring drivers of amphibian communities: pond hydroperiod (Wellborn

(C. D´ıaz-Paniagua, unpublished data). Pond size was extracted

from a 5-m resolution pond map layer obtained at a large flooding event (Go´ mez-Rodrı´guez et al., 2010b).

Aestivating habitat. We have differentiated three terrestrial habitat types according to a gradient of moist–arid environ- ment (forest habitat > scrub habitat > dune habitat). Aridity may be a barrier to interpond movements for amphibians and also a source for hydric stress during aestivation, when individual survival may be severely compromised because of dehydration (Pinder et al., 1992). We used the ecosystem cartography of DNP (Montes et al., 1998) to assess the percentage of each terrestrial habitat category in a 200-m buffer area from the edge of each pond.

Spatial distribution of breeding habitats. These are variables related to pond accessibility from nearby water bodies. As a measure of pond accessibility, we classified in three categories the proportion of pond shore surrounded by adjacent hygro- phyte vegetation (Table 2) using aerial photography (Junta de Andalucı´a, 2003). As a measure of pond connectivity, we measured the edge-to-edge distance to the nearest pond and to the marshes, using the 5-m resolution pond map layer with ArcView GIS 3.2. We also counted the number of nearby water bodies (excluding the marshes) surrounding each study pond in a 200-m buffer area from the edge of the pond. This distance has been reported for routinary movements between ponds in other amphibian species (Marsh et al., 1999). To account for interannual variability in pond connectivity attributable to meteorological conditions, we categorized surrounding ponds according to their size, which is generally related to the hydroperiod/permanence of temporary ponds in the study area (see Fortuna et al., 2006). So we discriminated ponds flooding in very wet years (all ponds, including those of small size) from those ponds that flood even during dry years

(ponds larger than 4000 m2).

Ecosection type. We recorded the ecosection in which the pond was located, as extracted from the ecosystem cartography in Montes et al. (1998). This is an indirect predictor, with no direct biological relevance for a species, but it informs us whether habitat selection is affected by spatial autocorrelation or conditioned by local attributes related to ecosystem type (i.e. dry zones, ecotones, etc.) that may have not been considered in the remaining sets of variables.

In total, 11 variables were used in 12 competing models (see

Table 2).


Table 2 Set of candidate models evaluated within each hypothesis. The variable and its form (continuous/categorical/ordered categorical)

are specified.


Hypothesis

Model

Habitat variables

Observations

Ecosection

Breeding habitat

1. Ecosection

2. Pond size

Ecosection

Pond size + (Pond size)2

Factor


Continuous




3. Hydroperiod

Hydroperiod

Ordered factor:

ephemeral pond (flooded 1–2

months in a wet year), intermediate temporary pond (3–6 months in a wet year), long-duration temporary pond (7–11 months in a wet year) zacallon

4. Global breeding habitat Hydroperiod + Pond size (see previous observations) Aestivating habitat 5. Suitable Forest + Scrub Continuous

6. Unsuitable Dunes Continuous

7. Global aestivating habitat Forest + Scrub + Dunes (see previous observations)

Spatial distribution of breeding habitats

8. Accesibility Surrounding vegetation (Surr. veg.) Ordered factor:

no hygrophyte vegetation, intermediate hygrophyte vegetation (surrounding 25–75% of the pond shore),

hygrophyte vegetation surrounding more than 75%

of the pond

9. Closest source Distance to marshes + Distance to nearest pond

Continuous



10. Sources (wet year) Number of ponds Continuous

11. Sources (dry year) Number of large ponds Continuous



12. Global breeding habitats distribution

Surr. veg. + Distance to marshes

+ Distance to nearest pond

+ Number of ponds



+ Number of large ponds

(see previous observations)




Statistical analyses
Estimation of absence data reliability
A non-detection record (equivalent to a recorded absence) represents a lack of evidence that the species bred in that given pond during the entire study period. Recorded presences were assumed to be completely reliable (P* = 1). Detection reliabil- ity was computed from species-specific single-visit detection probability values, calculated and reported in a previous study for each month of survey (Go´ mez-Rodrı´guez et al., 2010d). The single-visit detection probability (Pi) of a species is the probability that it will be detected at a pond in one sampling visit, given that the species breeds in the pond. To compute such probability values, Go´ mez-Rodrı´guez et al. (2010d) sampled every month 19 temporary ponds (also included in this study) and two permanent ponds in DNP during the amphibian breeding season in three different years: February – May 2003, January – May 2004 and March – May 2006. They estimated the single-visit probability of detection (Pi) for each species and sampling month using single-species occupancy models (MacKenzie et al. 2002, MacKenzie et al., 2006) with the program Presence (MacKenzie et al. 2002).

In this study, we estimate the reliability of each non- detection record as the probability of having detected the species given the timing of survey for that particular case (i.e. taking into account the dates in which the pond was surveyed). So we computed the probability of detecting a species after k visits, i.e. reliability, (P*) by applying the formula provided by MacKenzie et al. (2006), where Pi is the single-visit probability of detection, which depends on the species and month of survey:


k

P* ¼ 1 Y ð1 Pi Þ

i¼1


Species distributions models
We developed a set of twelve a priori candidate models (Table 2) for each species, based on scientific, available field and expert knowledge, to assess which variables best explained the species’ probability of breeding attempt in a given pond during the study period. Models were classified under the aforementioned four main hypotheses, depending on the type of habitat characteristics included (Table 2). Within each

Table 3 Models with highest Akaike’s weight (Akaike x) and their corresponding AUC (± standard error) are shown for each species*. Results are shown for the two approaches for SDMs building. The number of valid cases used to compute the AUC is specified in the SDMs accounting for the absence data reliability.






Model


Akaike x


AUC ± SE (valid cases)




Model


Akaike x


AUC ± SE

Pelobates cultripes

4. Global breeding habitat


0.989

0.824 ± 0.048 (117)




1. Ecosection


0.810

0.775 ± 0.027

















4. Global breeding habitat

0.185

0.803 ± 0.039

Discoglossus galganoi

7. Global aestivating

0.326

0.863 ± 0.189 (37)




1. Ecosection

0.919

0.754 ± 0.026



5. Suitable


0.249

0.823 ± 0.127 (37)

















6. Unsuitable

0.16

0.661 ± 0.068 (37)
















11. Sources (dry year)

0.066

0.681 ± 0.098 (37)
















10. Sources (wet year)

0.046

0.516 ± 0.090 (37)
















2. Pond size

0.039

0.490 ± 0.110 (37)













Hyla meridionalis

1. Ecosection

1

0.814 ± 0.035 (165)




1. Ecosection

1.000

0.806 ± 0.029

Pleurodeles waltl

2. Pond size

0.29

0.627 ± 0.074 (110)




4. Global breeding

0.809

0.720 ± 0.039



3. Hydroperiod


0.233

0.585 ± 0.069 (110)




3. Hydroperiod


0.122

0.688 ± 0.034





4. Global breeding habitat

0.198

0.609 ± 0.067 (110)
















10. Sources (wet year)

0.103

0.592 ± 0.067 (110)
















11. Sources (dry year)

0.093

0.556 ± 0.049 (110)













Triturus pygmaeus

1. Ecosection

0.762

0.695 ± 0.032 (202)




1. Ecosection

0.785

0.726 ± 0.030

Lissotriton boscai



12. Global breeding habitats distribution

7. Global aestivating habitat



0.096
0.483

0.748 ± 0.033 (202)
0.727 ± 0.056 (97)




12. Global breeding habitats distribution

8. Accessibility



0.098
0.383

0.736 ± 0.033
0.735 ± 0.039




12. Global breeding

0.376

0.798 ± 0.043 (97)




5. Suitable

0.314

0.759 ± 0.036




habitats distribution










7. Global

0.196

0.766 ± 0.035



SDMs accounting for absence data reliability Traditional SDMs

habitat


habitat


aestivating habitat
SDM, species distribution models.

*Only the minimum number of models necessary to achieve a global Akaike¢s weight above 0.85 is shown.




hypothesis, we developed single-predictor models to test the relevance of specific habitat characteristics, and a global one, including all predictors, to compare the alternative main hypotheses. A previous exploration with generalized additive models (GAMs) (Hastie & Tibshirani, 1990) evidenced that only the species–pond size relationship might be curvilinear and therefore a quadratic term was only considered for this variable. We did not construct a complete model (i.e. including all variables together) since it would have included far more variables than reasonable given the sample size. We did not consider all possible combinations of variables, as this approach typically inflates the number of models beyond the number that can be reliably analysed (Burnham & Anderson,

2002).


Candidate models were built using generalized linear models (GLMs) (McCullagh & Nelder, 1989) with binomial errors and a logit link (function ‘glm’ in ‘Stats’ package of R software, R Development Core Team, 2010). The response variable was the presence–absence (breeding evidence vs. no breeding evidence) of the species in a given pond during the entire study period. We explicitly accounted for the reliability of the data by weighting each case by its reliability (P*).

To identify the best model within the set of candidate models, we followed a model selection approach based on Akaike’s information criterion (AIC) and multimodel infer- ence of parameters (Burnham & Anderson, 2002; see Vicente et al., 2010 for an example in distribution modelling) so that estimates of model parameters were based on the entire set of models rather than on the one selected as best. We ranked models according to their AIC values to obtain Akaike’s model weights (x), ranging between 0 (low model impor- tance) and 1 (high model importance) and quantifying the uncertainty that each model is the target best model (see Burnham & Anderson, 2002). Afterwards, we computed the relative importance of each parameter by summing the Akaike’s weights across all the models in the set where the variable occurred (Burnham & Anderson, 2002). For each parameter, we also computed its model-averaged estimate and its unconditional standard error, which incorporated model selection uncertainty into estimates of precision (Burnham & Anderson, 2002). Model-averaged estimates are less biased compared to the estimator from just the selected best model and are especially useful if no model is clearly best (Burnham

& Anderson, 2002).




Models were evaluated based on the Area Under Curve, AUC (function ‘somers2’, library ‘Hmisc’), the percentage of explained deviance (D2) and the adjusted D2, which takes into account the number of parameters in the model (Guisan & Zimmermann, 2000). We computed the AUC only from cases with high reliability. We set P* ‡ 0.80 to identify cases with high reliability except in the case of Lissotriton boscai (Lataste,

1879), for which we set P* ‡ 0.50 since only two ponds showed P* ‡ 0.80 (note that P* was always higher than 0.5 in all positive cases). We assessed the standard error of each evaluation statistic using a parametric bootstrap (1000 sam- ples) in which the species prevalence in each sample was kept constant and equal to the one in the real data set. Bootstrap is recommended to assess the stability of a model when the data set is too small to be split into separate data sets for model building and evaluation (Guisan & Zimmermann, 2000), as in this study.

1.0
0.8
0.6
0.4
0.2
0.0
–0.2
Mean ± SD

For the purpose of comparison, we also built traditional SDMs following the same procedure. In these models, the reliability of absence data was not accounted for, and thus, the cases were not weighted. Similarly, AUC was computed from all the cases.

RE SUL T S


We detected eight species in the study area: Bufo calamita Laurenti, 1768; Pelobates cultripes (Cuvier, 1829); Discoglossus galganoi Capula, Nascetti, Lanza, Bullini & Crespo, 1985; Pelophylax perezi (Seoane, 1885); Hyla meridionalis Boettger,

1874; Pleurodeles waltl Michahelles, 1830; Triturus pygmaeus (Wolterstorff, 1905) and L. boscai. However, we could not estimate the probability of detection nor build models for B. calamita and P. perezi, since a single-visit probability of detection was lacking for most sampling visits (see Go´ mez-Rodrı´guez et al., 2010d), and thereby, these species were excluded from the study. Triturus pygmaeus and H. meridionalis were the species that bred in a larger propor- tion of ponds (55% and 46%, respectively), whereas D. gal- ganoi was only detected in 13% of the ponds (Fig. 1). In any given pond, the reliability of non-detection records differed widely among species (Kruskal–Wallis test: H = 462.19, d.f. = 5; P < 0.001; Fig. 2). Triturus pygmaeus and H. merid- ionalis showed the highest mean probability of detection in ponds where we did not detect the species, thereby evidencing high levels of reliability in their absence data (Fig. 2). On the contrary, D. galganoi and L. boscai were the species with less reliable absence data, their reliability being null in 73 ponds and in 124 ponds, respectively.

We observed differences in model ranking and model weights among species, evidencing that there was not a ‘best supported hypothesis’ valid for all of them (Fig. 3, see Table S1 in the Supporting Information for details). Only P. cultripes and H. meridionalis showed a model clearly ranked as best (Akaike x > 0.98), whereas the rest of the species showed similar support for competing models, although those models were within the same main hypothesis except in the case of

Figure 2 Mean value and standard deviation of absence data reliability in the ponds where the non-detection of the species was recorded. Absence reliability represents the probability of having detected the species, given presence, after the sampling visits conducted during the entire study period.


L. boscai (Fig. 3). The hypotheses that obtained a higher mean support were the ‘ecosection’, the ‘global breeding habitat’ and the ‘global aestivating habitat’, whereas the ones with lower mean support were the ones related to the ‘spatial distribution of breeding habitats’, except the one including all the variables (Fig. 3).

Models ranked as best showed a useful to good value of AUC (> 0.80) in the case of anurans and lower values for urodele species (Table 3). Similarly, SDMs accounting for data reliability showed higher AUC values than traditional SDMs in the case of anurans. In the case of urodeles, AUC values were similar between both types of models except in the case of P. waltl, for which traditional regression models seemed to work better. We also observed differences in model ranking and model weights between the two regression methods, the most remarkable difference being the larger support obtained for the ecosection hypothesis with tradi- tional SDMs (Table 3).

In models accounting for the reliability of non-detection records, the ecosection hypothesis obtained a great support from species breeding in a large proportion of ponds (H. meridionalis and T. pygmaeus) (Fig. 3 and Table 4) as they mainly occurred in a single ecosection (Ecosection 3: ‘Wet stabilized sands at higher elevation’) (Fig. 4). Models within the ‘breeding habitat hypothesis’ explained best the distribu- tion of species with long larval development (P. waltl and P. cultripes) (Fig. 3). Their probability of occurrence increased with pond size (see Supporting Information: Tables S2 and S3 for details) and hydroperiod in temporary ponds (Fig. 4). Models related to the ‘aestivating habitat hypothesis’ best explained the distribution of D. galganoi, while this hypothesis as well as the ‘spatial distribution of breeding habitat patches’




(a) (b)

0.5

1



































































Mean ± SE


0.4

0.8



0.3

0.6



0.2

0.4



0.1

0.2




0.0 0


Breeding habitat hypothesis
  1   2   3


Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur ©atelim.com 2016
rəhbərliyinə müraciət