Modelling the distribution of bats in relation to landscape structure in a temperate mountain environment


  • Christophe Jaberg,

    Corresponding author
    1. Swiss Co-ordination Centre for the Study and Protection of Bats, Neuchâtel (CCO-NE), Musée d’histoire naturelle, CH-2300 La Chaux-de-Fonds, Switzerland;
    Search for more papers by this author
  • Antoine Guisan

    1. Swiss Centre for Faunal Cartography (CSCF), Terreaux 14, CH-2000 Neuchâtel, Switzerland
    Search for more papers by this author
    • Present address: Institute of Ecology, University of Lausanne, CH-1015 Lausanne, Switzerland.

C. Jaberg, Centre de coordination ouest pour l’étude et la protection des chauves-souris c/o Musée d’histoire naturelle, Av. Léopold-Robert 63, CH-2300 La Chaux-de-Fonds, Switzerland (fax +41 32 913 3976; e-mail


  • 1Landscape modification is often considered the principal cause of population decline in many bat species. Thus, schemes for bat conservation rely heavily on knowledge about species–landscape relationships. So far, however, few studies have quantified the possible influence of landscape structure on large-scale spatial patterns in bat communities.
  • 2This study presents quantitative models that use landscape structure to predict (i) spatial patterns in overall community composition and (ii) individual species’ distributions through canonical correspondence analysis and generalized linear models, respectively. A geographical information system (GIS) was then used to draw up maps of (i) overall community patterns and (ii) distribution of potential species’ habitats. These models relied on field data from the Swiss Jura mountains.
  • 3Eight descriptors of landscape structure accounted for 30% of the variation in bat community composition. For some species, more than 60% of the variance in distribution could be explained by landscape structure. Elevation, forest or woodland cover, lakes and suburbs, were the most frequent predictors.
  • 4This study shows that community composition in bats is related to landscape structure through species-specific relationships to resources. Due to their nocturnal activities and the difficulties of remote identification, a comprehensive bat census is rarely possible, and we suggest that predictive modelling of the type described here provides an indispensable conservation tool.


Although legally protected in most European countries (for Switzerland see Moeschler 1991), bats are endangered throughout much of their distribution (Stebbings 1988). Only two of the 26 bat species in Switzerland have stable or possibly increasing populations, whereas seven species are in danger of extinction or are vulnerable (Duelli 1994). Possible causes of this decline involve the availability of suitable foraging habitats, use of toxic pesticides in agricultural landscapes, roost destruction and disturbance of subterranean hibernation sites (Stebbings 1988; Walsh & Harris 1996a,b).

Thorough knowledge of species distribution is of prime interest for biological conservation as distribution maps are widely used by conservation biologists to assess a species’ status, to draw up species red lists or to pinpoint areas of particular biological value (Cowley et al. 2000). However, identifying distribution in bats presents a considerable challenge. Intensive population surveys of bats are difficult to conduct because of their nocturnal behaviour, their wide home range and the problems related to species identification in flight (Walsh & Harris 1996b). As an alternative, species–habitat relationships could be modelled to predict species’ distributions, from chosen habitat descriptors (Guisan & Zimmermann 2000). Similar approaches have been applied recently to several large-scale conservation issues to predict presence/absence of species in remote areas (Manel et al. 1999), to assess impact of environmental changes or landscape management (Franklin 1995; Chamberlain et al. 1999; Cowley et al. 2000; Manel, Buckton & Ormerod 2000; Milsom et al. 2000; Paradis et al. 2000) and to improve comprehension of large-scale processes in the field of fundamental ecology (Manel, Buckton & Ormerod 2000; Collingham et al. 2000). More specifically, models can satisfy the crucial need to identify the ecological requirements of bat species (Moeschler 1991).

Key resources for bats occur in all landscape types but change during their life cycle (Law & Dickman 1998). Therefore, a combination of landscape descriptors might be required for predicting species’ distributions. Nevertheless, because some resources have a discontinuous distribution (e.g. roosts; Law, Anderson & Chidel 1999), species could be absent from suitable habitats, whereas unsuitable habitats could be occupied by vagrant individuals (Hanski & Gilpin 1997). Thus, species–habitat relationships should be investigated at a relatively coarse resolution in order to construct models of value in conservation biology.

Previous attempts to model species–resources relationships in bats are inadequate for the large-scale prediction of species’ distributions. For example, ecomorphological models are available (Aldridge & Rautenbach 1987; Baagøe 1987; Norberg & Rayner 1987) but resource utilization cannot be predicted from species’ ecomorphology alone (Brigham, Aldridge & Mackey 1992; Saunders & Barclay 1992; Arlettaz 1999). Available models of bat abundance (Walsh & Harris 1996b) are usually not species-specific or, if they are (Vaughan, Jones & Harris 1997; Law, Anderson & Chidel 1999), they do not consider the utilization of roosts or hibernation sites although it could affect species’ distributions (Sedgeley 2001). In other faunal groups, potential distribution is usually assessed at the individual species level through univariate methods (i.e. one response variable at a time, e.g. generalized linear models; Chamberlain et al. 1999; Milsom et al. 2000). A holistic view of how communities vary with the environment is thus neglected. Multivariate approaches that predict the distribution of all the species of the community simultaneously is an interesting alternative (ter Braak 1986) that is rarely used with bats. In this study, both multivariate and univariate approaches have been used in a complementary way.

By using a database, set up by the Swiss Co-ordination Centre for the Study and Protection of Bats (CCS) between 1984 and 1998, we aimed to (1) test the hypothesis that large-scale spatial patterns in bats are related to landscape structure; (2) identify landscape factors that could be used in predictive models of (i) overall community composition and (ii) individual species’ distributions; and (3) implement the statistical models into a geographical information system (GIS) to produce maps of (i) synthetic landscape gradients affecting species’ assemblages and (ii) potential habitat distribution of individual species.

The species-specific approach focused on four species of particular conservation interest for Switzerland. Vespertilio murinus L. 1758 and Eptesicus nilssoni (Keyserling & Blasius 1839) reach the south-west limit of their world-wide breeding area in the Jura mountains. Their populations are isolated and therefore potentially endangered in Switzerland (Duelli 1994). For reasons that remain unclear, Pipistrellus nathusii (Keyserling & Blasius 1839), although abundant, has a very patchy distribution in the study area. Myotis myotis (Borkhausen 1797) is a highly endangered species in Switzerland (Duelli 1994), as in many European countries (Stebbings 1988), mainly because of extensive building renovations that have caused the destruction of roosts, and also because of agricultural intensification that leads to modification of foraging habitat. With the construction of two seasonal distribution models for Myotis myotis, we tested the hypothesis that this species exhibits seasonal shifts of habitat use (C. Jaberg, personal observation). Identifying these shifts could be of prime importance for establishing more efficient conservation planning.

Study area

The data came from the Canton of Neuchâtel, western Switzerland (47°00′ N, 6°46′ E; Fig. 1), which covers 786 km2 and stretches from the Jura mountains (1200–1400 m a.s.l.) in the north-west to lake Neuchâtel (430 m a.s.l.) in the south-east. Because of its limestone nature, the study area is poor in surface rivers but has a wide underground karstic system. Forests cover about one-third of the area. Because of the accentuated relief, the climate, vegetation and human impact vary between sites. In the upper Jura valleys, rainfall averages 2000 mm a year and annual temperature 5 °C. Vegetation comprises some boreoarctic features. Summits are usually forested with Picea abies (L.) within Aceri–Fagion, Abieti–Fagion and Asperulo–Fagion communities. Outside the two main cities (La Chaux-de-Fonds and Le Locle), these valleys are sparsely populated and used for cattle-rearing. The lake shores and neighbouring Jura foothills have a mild climate (rainfall averaging 980 mm and annual temperature 9 °C). Climax vegetation communities are mesophilous to thermophilous (Carpinion, Cephalanthero–Fagion and Quercion pubescentis) and vineyard or cereal cultures are widespread. Human activities in this densely populated area have contributed to the disappearance of the typical lakeside vegetation (Phragmition and Fraxinion) in many places.

Figure 1.

Location of the study area in Switzerland, sampling design (2·5 × 2·5-km grid squares) and number of observations.

Materials and methods


The database used for this study comprised about 3500 bat records (Table 1 and Fig. 1) obtained using several complementary methods adequate for sampling all species potentially occurring in the study area and for taking into account all activities (foraging, roosting, rearing young, mating, hibernating). (i) Systematic investigations of public buildings (churches, schools, historical buildings) were used to sample all roosting or breeding anthropophilous species. (ii) Mist-netting of individuals flying above water was used intensively because it permits the sampling of the majority of species, although this method is inadequate for aerial species flying above mist-net height. (iii) Systematic surveys of hibernation sites, using either mist-nets at the cave entrance or visual counting of hibernating bats, were used to record cave-dwelling species during hibernation. (4) Identification by a bat specialist of individuals found by the public was used widely. However, forest-dwelling species were rarely sampled by this method because the chances of discovering an individual in such dense habitats are very low. (5) Acoustic and/or visual identification of foraging or commuting bats with bat detectors was used to record aerial species that usually have a distinctive echolocation signal (Ahlén 1990). These data were recorded in the field with an accuracy varying between ±25 and ±100 m, and stored in the ORACLE database of the Swiss Centre for Faunal Cartography (CSCF) in Neuchâtel.

Table 1.  Structure of the bat database used for modelling community composition and species’ distributions in the Swiss Jura mountains: number of 2·5 × 2·5-km squares in which a given species was recorded; number of bat recordings sampled through five different methods – building = systematic investigation of public buildings, mist-net = mist-net captures of foraging individuals, hibernation = mist-net capture of individuals visiting hibernation sites or visual observation of hibernating bats, citizens = bat data communicated by citizens, acoustic = acoustic recordings of foraging or commuting individuals (see text for details)
SpeciesNumber of squaresNumber of bat recordings per method
Barbastella barbastellus 3  0   0  6  1 07
Eptesicus nilssoni21 10  75 17 58 0160
Eptesicus serotinus10  7  32  3  1 043
Miniopterus schreibersi 2  0   0 11  0 011
Myotis bechsteini 6  0  15  3  0 018
Myotis brandti 3  0   2  0  4 06
Myotis daubentoni28  21558247 15 01822
Myotis myotis25 24 141123  8 1297
Myotis mystacinus28  1  36 41 23 0101
Myotis nattereri 3  0   3  6  0 09
Nyctalus leisleri12  4   2  0 10 016
Nyctalus noctula16  2  28 11 28 978
Pipistrellus kuhli 3  1   1  0  1 03
Pipistrellus nathusii21 15  11  2 54 183
Pipistrellus pipistrellus54 35 186 19142 0382
Plecotus auritus43 22  55132 40 0249
Plecotus austriacus 5  0   1  0  5 06
Rhinolophus ferrumequinum 2  0   0  7  0 07
Rhinolophus hipposideros 1  2   1  0  0 03
Vespertilio murinus10 24  17  0145 0186
Total 1492164628535113487

These various methods were inadequate for quantifying species’ abundances accurately. Thus we developed a semi-quantitative index, the species’ regularity, defined as the number of years during which the species was recorded at least once in the observation unit. Species’ regularity constituted the dependent variable in each model.


The descriptors of landscape structure were derived from 25 × 25-m and 100 × 100-m resolution data provided by the Federal Office for Statistics (Neuchâtel; land cover) and the Federal Office of Topography (Bern; topography, contours of study area and hydrography) (Table 2 and Fig. 2). The longitude and latitude of the grid squares’ centroid were used to examine spatial trends. A mean value of elevation was provided by aggregating the 25-m resolution digital elevation model (DEM) of the area (Fig. 2a). The distance between the grid square centroid and the nearest lake was calculated to measure the importance of lake habitats to the distribution of bat species. Finally, a series of land-cover classes was used to test the importance of various landscape elements (Fig. 2b–f). Cover measures were given as a percentage (number of 100 × 100-m pixels attributed to a given habitat divided by the total number of pixels in a grid square, i.e. 625 pixels). We did not use climate descriptors because this information was already provided by the mean elevation.

Table 2.  Influence of the 17 landscape structure descriptors on community species composition according to the CCA. % explained variance = percentages of variance explained by each descriptor; P-level = significance as obtained from the Monte Carlo test of the first canonical axis (1000 random permutations)
Descriptor% explained varianceP-level
  1. Significance levels: NS > 0·1; *≤ 0·1; **≤ 0·05; ***≤ 0·01; ****≤ 0·001.

 1. Longitude 5·4****
 2. Latitude 3·2**
 3. Elevation12·1****
 4. Distance to the nearest lake10·1****
 5. Dense forest 3·3**
 6. Open forest 2·9*
 7. Forest patches 4·1**
 8. Hedges 4·1**
 9. Isolated trees 6·0****
10. Vineyard 6·0***
11. Orchards 1·6NS
12. Intensive agricultural land 3·2*
13. Extensive agricultural land 2·0NS
14. Pastures 2·7NS
15. Lake 7·9****
16. Industrial building 2·4NS
17. Residential houses 6·6***
Figure 2.

Spatial variation of six major landscape descriptors, selected in the CCA and GLM: (a) elevation (m a.s.l.); (b) cover of pastures; (c) cover of isolated trees; (d) cover of hedges; (e) cover of residential buildings; (f) cover of dense forest. Cover values are as a percentage (see text).


For the statistical analysis, both species and environmental predictors were aggregated per square unit of 2·5 × 2·5 km. Species data were aggregated directly in the database. The aggregation of environmental descriptors was performed within the ARCINFO® (Environmental Science Research Institute Inc., Redlands, CA) GIS. For practical reasons (data availability), the aggregation value of the grid squares located at the border of the study area was calculated from a limited number of inner grid cells, and those outside the study area (e.g. in France) were not taken into account in the calculation. Because complete absence of bats is unlikely in an area of 2·5 × 2·5 km, squares without any bat records were assumed to have been insufficiently prospected and were removed from the analysis (Fig. 1).

Overall community composition model

Relationships between overall community composition and landscape structure were modelled through canonical correspondence analysis (CCA; ter Braak 1986), which simultaneously relates the species’ distributions to synthetic gradients, the canonical axes, which are linear combinations of the measured environmental predictors (ter Braak & Verdonschot 1995). CCA was performed on the regularity matrix of 20 species recorded in 56 grid squares (i.e. 36·6% of all squares, n = 153). Grid squares in which less than two species were recorded were omitted. Of the 20 species, Rhinolophus hipposideros (Bechstein 1800) was omitted because it was recorded in a unique site and thus would have created distortion in the analysis. Furthermore, Pipistrellus kuhli (Natterer in Kuhl 1819), Myotis brandti (Eversmann 1845), Myotis nattereri (Kuhl 1818), Rhinolophus ferrumequinum (Schreber 1774), Barbastella barbastellus (Schreber 1774) and Miniopterus schreibersi (Natterer in Kuhl 1819) had fewer than five occurrences and were down-weighted as ‘passive species’ in the analysis, in order to identify only the patterns drawn from the more common species. Each descriptor was introduced singly in 17 separate preliminary models in order to quantify and test their individual influence on community composition. Factors accounting for a significant part of species’ variation and which were not cross-correlated were selected manually to construct the final model. Canonical axes of the final model were then recalculated in the GIS to map the predicted community composition over the whole study area. CCA analyses were run within the canoco program version 3.2 (ter Braak 1988) using log-transformed data.

Individual species’ distribution models

For individual species, we used generalized linear models (GLM; McCullagh & Nelder 1989; Nicholls 1989) as employed for other faunal groups (Scott et al. 2000). Species’ regularity was recorded within 83 grid squares, i.e. all squares with at least one bat record (54·2% of all squares). GLM, combined with a GIS, were used to generate potential habitat distribution maps of species (Guisan, Theurillat & Kienast 1998; Guisan, Weiss & Weiss 1999). We did not fit a simple presence–absence model (e.g. GLM with binomial distribution and logistic link) because, in our database, the long sampling period (14 years) offered the opportunity to extract information about the regularity of species. Instead, GLM were fitted by specifying a Poisson distribution and a logarithmic link function. Our species’ regularity index effectively proved to be best modelled as counts (of years), rather than specifying a Gaussian distribution in the GLM or by reclassifying the response according to a semi-quantitative scale and modelling it with a particular case of GLM (Guisan & Harrell 2000).

Exploratory GLM were first run to assess the importance of the various predictors, selected by an Akaike information criterion (AIC)-based stepwise procedure. Final models included only those predictors whose deviance reduction was significant against χ2 at the 0·05 confidence level. GLM were fitted using the S statistical language in the S-Plus 2000 software (Mathsoft Inc., Cambridge, MA).

The seasonal models of Myotis myotis were fitted on two separate data subsets: (i) a summer subset, corresponding to data sampled during the breeding period (May to August), and (ii) a winter subset, corresponding to data sampled during the dispersal, mating and hibernation periods (September to April).

Model evaluation

We evaluated each GLM using a leave-one-out jack-knife procedure (Efron & Tibshirani 1993; Manel et al. 1999; Guisan & Zimmermann 2000). Eighty-three GLM were run for each species, each time excluding a single observation. All the submodel parameters and summary statistics, regression coefficients and their associated deviance reduction, as well as model D2 and adjusted D2, were stored in the same output matrix. The stability of the models could be assessed by looking at the jack-knife distribution of these measures. At each run, the fitted model was used to predict the response at the plot excluded from the analysis, producing 83 predicted values of species’ regularity. These predicted values were compared to observed (real) species’ regularity using a simple Pearson correlation coefficient (Paradis et al. 2000). The mean and the sum of absolute differences between these values were provided as complementary information to the correlation coefficient.

Finally, in order to evaluate the success of predictions reduced to presence/absence, we calculated various measures of predictive success based on the presence–absence (0/1) contingency table (or error matrix) obtained by cross-tabulating real observations against predictions, both reclassified into 0 or 1 (0 = 0, > 0 = 1). This resulted in four possible situations: (i) true positive (number of squares with correctly predicted presences); (ii) false positive (number of squares with falsely predicted presence, i.e. absence observed); (iii) false negative (number of squares with falsely predicted absence, i.e. presence observed); (iv) true negative (number of squares with correctly predicted absences). From the ratios between these four cases, various performance measures could be calculated: (i) true positive rate; (ii) true negative rate; (iii) sensitivity; (iv) specificity; (v) kappa (for a more thorough presentation and discussion of these measures see Fielding & Bell 1997). They all range between 0 and 1, 0 indicating no agreement and 1 a full agreement. Kappa additionally tests against chance performance.



Elevation appeared to affect the distribution of the four species (Table 3) and the overall community composition (Table 2). Its influence seemed relatively strong, as it explained up to 44·5% of the deviance in individual species’ distributions (Vespertilio murinus; Table 3). Natural or semi-natural vegetation cover (dense forest, hedges, isolated trees, forest patches and pastures) also appeared to affect the distribution of each species (Table 3) and the community composition (Table 2), although more weakly (isolated trees: 23% of deviance in Eptesicus nilssoni’s distribution). Thus, elevation and vegetation structure appeared to be adequate for predicting the distribution of most species.

Table 3.  Predictive distribution models (GLM) of four chosen species. % explained deviance = percentages of deviance explained successively by each predictor; P-level chi = significance level associated with the chi-test of deviance reduction (for each predictor); P-level perm = empirical significance as obtained from 1000 random permutations of the response variable; D2= total proportion of deviance explained by the model; adj. D2 = same as D2 but corrected for the number of freedom used to build the model; cor = correlation between observed and fitted values. Significance levels given for D2, adj. D2 and cor were also obtained from the permutation tests
ModelLandscape descriptors (predictors)% explained devianceP-level chiP-level perm
  1. Significance levels: NS > 0·1; *≤ 0·1; **≤ 0·05; ***≤ 0·01; ****≤ 0·001.

Vespertilio murinusElevation37·54********
 Elevation2 7·01****NS
 Lake 9·32*****
 D2 = 0·68****/adj. D2 = 0·66****/cor = 0·79****
Pipistrellus nathusiiLatitude 8·12******
 Elevation2 3·42**NS
 Distance to lake 3·56***NS
 Pastures 5·79*****
 D2 = 0·58****/adj. D2 = 0·55****/cor = 0·70****
Eptesicus nilssoniElevation 7·99******
 Isolated trees12·11*******
 Isolated trees210·9******
 Hedges 6·06*****
 Dense forest2 8·24******
 Residential houses 2·47**NS
 D2 = 0·48****/adj. D2 = 0·44****/cor = 0·76****
Myotis myotis summerElevation16·45********
 Elevation2 9·67******
 Isolated trees2 4·72**NS
 Pastures 7·33*****
 D2 = 0·38***/adj. D2 = 0·36***/cor = 0·48****
Myotis myotis winterDistance to lake11·76********
 Dense forest2 3·08***
 Elevation 2·06**NS
 Hedges 5·12****
 Forest patches2 3·58*****
 D2 = 0·26**/adj. D2 = 0·21**/cor = 0·52****

Geographic location, lake and residential buildings seemed to affect the overall community composition (Table 2) but not each species’ distribution (Table 3). However, their impact, when significant, appeared to be relatively strong (e.g. latitude: 24% of deviance in Pipistrellus nathusii’s distribution). They appeared thus to be good predictors of distribution but only for particular species.

Anthropogenic landscape structures (agricultural land, vineyard, orchards, industrial buildings), except residential buildings, were poor predictors of individual distribution (Table 3) or overall community composition (Table 2).

GLM for predicting summer and winter distribution of M. myotis were different (Table 3), suggesting a seasonal shift of habitat use by this species. The species’ distributions seemed related to low elevation during summer and to higher elevation and structured vegetation cover (hedges) outside the breeding period (Fig. 3d–e).

Figure 3.

Potential habitat distribution maps of four bat species. The maps were generated by introducing the regression coefficients of the GLM into a GIS. (a) Vespertilio murinus; (b) Pipistrellus nathusii; (c) Eptesicus nilssoni; (d) Myotis myotis during summer; (e) Myotis myotis during winter.


To reduce multicollinearity among independent variables, only eight descriptors were included in the final CCA model. They produced two synthetic gradients of landscape structure that explained a significant part of variation in community composition (Table 4): (i) an elevation/distance from lake gradient extending from the lake shore to the summits of the Jura mountains, and (ii) a human-influence gradient extending from a high urbanization level to semi-natural forest cover (Table 5). They segregated the community into four assemblages of species sharing similar ecological requirements (Fig. 4) and were used to predict their distribution (Fig. 5): (i) the ‘low elevation species’ (or lake species), including Pipistrellus kuhli, Plecotus austriacus (Fischer 1829), Vespertilio murinus, Pipistrellus nathusii and Eptesicus serotinus (Schreber 1774) (see high sample scores on canonical axis 1; Fig. 5); (ii) a unique ‘high elevation species’, which also relies on urbanized areas (Eptesicus nilssoni alone) (low scores on axis 1); (iii) the ‘urban species’, which are independent of elevation, comprising Nyctalus leisleri (Kuhl 1818), Nyctalus noctula (Schreber 1774) and Pipistrellus pipistrellus (Schreber 1774) (high scores on axis 2); and (iv) the ‘forest-dwelling species’, including Plecotus auritus (L. 1758), all Myotis spp., Miniopterus schreibersi, Barbastella barbastellus and Rhinolophus ferrumequinum (low scores on axis 2).

Table 4.  Summary of final CCA model. Eight landscape descriptors were used to explain simultaneously variation of the 19 species × 56 sites matrix
  1. Significance levels: NS > 0·1; *≤ 0·1; **≤ 0·05; ***≤ 0·01; ****≤ 0·001.

  2. –, test not computed.

Eigenvalues 0·198 0·101 0·044 0·0361·453
Species–environment correlation 0·832 0·703 0·585 0·557 
Cumulative percentage variance
of species data13·620·623·626·1 
of species–environment relation46·870·881·389·8 
Sum of all unconstrained eigenvalues    1·453
Sum of all canonical eigenvalues    0·423
Monte Carlo tests of significance (1000 permutations)
F-ratio 7·42 4·13 1·872·41
Table 5.  CCA model: inter-set correlation coefficients of landscape descriptors with canonical axis. For each axis, the two higher correlation coefficients are in bold. Axis 1 is an elevation gradient (or lake gradient) whereas axis 2 is a forested–urbanized gradient
Landscape descriptorAxis 1Axis 2
Dense forest−0·301−0·253
Forest patches−0·374−0·218
Isolated trees−0·509−0·190
Residential houses0·4240·396
Figure 4.

CCA biplot. Axes 1 and 2 accounted significantly for 13·6% and 7·0% of variance, respectively. Buildings = cover of residential houses; vineyard = cover of vineyard; lake = cover of lake; forest = cover of dense forest; f. patches = cover of forest patches; trees = cover of isolated trees; hedges = cover of hedges; elevation = elevation; bar =Barabastella barbastellus; nil =Eptesicus nilssoni; ser = Eptesicus serotinus; bec =Myotis bechsteini; bra = Myotis brandti; dau = Myotis daubentoni; myo = Myotis myotis; mys = Myotis mystacinus; natt = Myotis nattereri; sch =Miniopterus schreibersi; lei =Nyctalus leisleri; noc = Nyctalus noctula; aur =Plecotus auritus; aus = Plecotus austriacus; kuh =Pipistrellus kuhli; nat = Pipistrellus nathusii; pip = Pipistrellus pipistrellus; fer =Rhinolophus ferrumequinum; mur =Vespertilio murinus.

Figure 5.

Spatial patterns in landscape influence on bat community patterns: predicted sample scores of first (a) and second (b) canonical axes. Maps were generated by introducing the canonical coefficients of CCA axes into a GIS. Dark squares are sites with high scores and light squares have low scores on CCA axes. See Fig. 4 for correspondence between scores and species composition or landscape attributes of sites.


Correlations between observed species’ regularity and predicted values (jack-knife) were fair to good, except for the seasonal models of Myotis myotis (Table 6 and Fig. 3d–e). Overall prediction success of the models ranged from 0·71 to 0·87. Absence was better predicted than presence as, in all five cases, specificity and true negative rate were always higher than sensitivity and true positive rate (Table 7 and Fig. 3). This resulted in kappa values between 0·18 and 0·63, hence encompassing agreement qualification ranging from poor to fair to good on the subjective scale of Monserud & Leemans (1992).

Table 6.  Results of the leave-one-out jack-knife for evaluating the stability of the five GLM. Adj. D2= adjusted D2 for the model with all observations; adj. D2 jack = mean adj. D2 of the 83 models fitted by the jack-knife procedure, with indication of the standard deviation around the mean; cor = correlation (in per cent) between observed species’ regularity and, respectively, (i) fitted values (o/f) and (ii) regularity values predicted from the jack-knife (o/p); mean AD = mean absolute difference between observed and (i) fitted (o/f) and (ii) predicted jack-knife values (o/p); sum AD = sum of all absolute differences between observed and (i) fitted (o/f) and (ii) predicted jack-knife values (o/p)
ModelAdj. D2 (%)Adj. D2 jack (%) ± SDCor (%) (o/f)Cor (%) (o/p)Mean AD (o/f)Sum AD (o/f)Mean AD (o/p)Sum AD (o/p)
Vespertilio murinus66·466·5 ± 1·778·863·10·4737·30·5847·8
Pipistrellus nathusii54·654·6 ± 1·470·354·10·5747·10·6755·3
Eptesicus nilssoni43·643·7 ± 1·775·558·00·5949·30·7259·9
Myotis myotis summer36·536·5 ± 1·648·124·40·4436·90·5041·8
Myotis myotis winter20·820·8 ± 1·852·421·40·8771·91·0183·7
Table 7.  Prediction success and performance measures of the five GLM, based on the presence–absence contingency table obtained by cross-tabulating real observations against predictions, both re-classified into 0 or 1 (see text for details). n= overall number of observations (= 83)
Performance measuresVespertilio murinusPipistrellus nathusiiEptesicus nilssoniMyotis myotis
True positive (a) 816 7 6 7
False positive (b)10 8 6 413
False negative (c) 2 414 711
True negative (d)6355566652
True positive rate a/(a + b) 0·44 0·67 0·54 0·60 0·35
True negative rate d/(d + c) 0·97 0·93 0·80 0·90 0·83
Sensitivity a/(a + c) 0·80 0·80 0·33 0·46 0·39
Specificity d/(d + b) 0·86 0·87 0·90 0·94 0·80
Overall correctly classified prediction rate (a + d)/N 0·86 0·86 0·76 0·87 0·71
Kappa 0·49 0·63 0·27 0·45 0·18



Because resources were insufficient to carry out an exhaustive and systematic survey of bat populations (Walsh & Harris 1996a,b), we based our modelling approach on data that were available from inventories whose completeness was not systematically assessed (Moreno & Halffter 2000). The relationships between bat distribution and landscape descriptors could thus be open to question. In order to assess any potential bias, we investigated (i) if community composition was influenced by the sampling methods and (ii) if sampling methods were correlated with landscape structure, using CCA and redundancy analysis (RDA) (ter Braak & Prentice 1988), respectively. The relationships between species and landscape described above should be evaluated with caution in six cases. Miniopterus schreibersi, Barbastella barbastellus, Rhinolophus ferrumequinum and, to a lesser extent, Myotis nattereri and Myotis bechsteini (Leisler in Kuhl 1818) were more frequently recorded in surveys of hibernation sites than with other methods (Table 1). As cave surveys were carried out exclusively within highly forested areas, it cannot be excluded that the modelled relationships in forested habitats are spurious or, at least, invalid for other areas. In addition, the strong relationship of Nyctalus leisleri with urban areas could be influenced by the high rate of sightings recorded by citizens. The risk of biased habitat-use patterns for other species is lower because (i) these species were recorded using all methods more or less equally or (ii) the preferred method (with regard to a given species) was uniformly used for all landscape types.

Models enabled us to quantify the influence of some of the main landscape characteristics affecting species’ distributions. Although modest for some species, the proportions of variance explained (20–60%) were significant as stated by permutation tests. Although stability can be assessed by checking for over dispersion (Breslow 1996), permuting the response variable a thousand times and refitting the model after each permutation allows one to identify those predictors whose contribution (usually small) to the reduction in deviance might not differ from a random effect (Guisan & Zimmermann 2000). This is important for the interpretation of models, as some predictors significantly retained by the chi-test of deviance reduction were not considered significant after computing a permutation test (Table 3). The fractions of variation left unexplained might be due to (i) non-rigorous sampling increasing the ‘noise’ in the data; (ii) other ecologically important unmapped landscape descriptors; (iii) unmapped environmental or anthropogenic perturbation effects (Borcard, Legendre & Drapeau 1992) (e.g. roost destruction); (iv) environmental factors that are undetectable at the spatial scale involved in this study (e.g. local habitat quality) but which affect community patterns at the landscape scale (Ormerod & Watkinson 2000b); (v) interspecific competition (Baagøe 1986); (vi) intraspecific variations in ecological requirements related to variation in resource availability or to demographic category. Similar variation in the proportions of variance explained resulted from other (similar) distribution modelling studies (Borcard, Legendre & Drapeau 1992; Guisan, Weiss & Weiss 1999; Guisan & Hofer 2001).


To our knowledge, such a modelling approach, and particularly the direct implementation of the statistical models into a GIS, has never been applied to bats. The current work demonstrates that this approach is relevant even for highly mobile organisms, although model stability and agreement tend to decrease for more widespread species such as Myotis myotis. The work also meets the need to extend macroecological studies to new groups of organisms in order to improve knowledge of large-scale ecological processes (Ormerod & Watkinson 2000a).

Identifying the true causes of distribution at large spatial scale presents a considerable challenge (Walsh & Harris 1996b; Ormerod & Watkinson 2000a). The aim of the present study was not to explain the spatial patterns observed but to select landscape correlates that could be used as predictors. However, several authors (Walsh & Harris 1996a; Vaughan, Jones & Harris 1997) have similarly found that the landscape descriptors selected in our models directly influence bat spatial patterns. These landscape structures are thus valuable to predict bat distribution or to identify areas of important biological value. Lakes and associated wet habitats constitute good foraging areas for several species (e.g. Vespertilio murinus; Jaberg, Leuthold & Blant 1998). Lake shores should thus be managed in ways that integrate the conservation of natural or semi-natural habitats in order to maintain the main bat food resources, i.e. high-density insects swarms. Forests and other woodland habitats positively affect foraging activity (Walsh & Harris 1996a; Vaughan, Jones & Harris 1997) and roost availability (Altringham 1996; Sedgeley 2001) for many bat species (e.g. Myotis spp.). Several landscape elements are used as flight paths and sonar guidelines for short-range echolocating species (Limpens & Kapteyn 1991) or as shelters against predators (Rydell, Entwistle & Racey 1996). Hedges, isolated trees or small forest patches should be more systematically protected or recreated within agricultural landscapes in order to improve the availability of foraging areas. Trees cavities used as roosts and important hibernacula also need special protection within forested areas. Presence of urbanized areas might also influence species’ distributions, as some species regularly forage near artificial lights (Nyctalus, Pipistrellus and Eptesicus spp.) although others do not (e.g. Myotis spp.) (Rydell & Racey 1995). Lofts of residential houses constitute roosts for numerous species (Altringham 1996) and precautionary measures should be taken during building restoration (Blant 1992). Some landscape descriptors also reflect unsuitable foraging habitats that are avoided by bats (e.g. intensive agricultural land; Walsh & Harris 1996a) or some unrecorded environmental correlate that might influence foraging more directly (e.g. mild climate and thermophilous vegetation that coincide with vineyard distribution in the case of Eptesicus serotinus; Gerber, Haffner & Ziswiler 1996; C. Jaberg, personal observation).

Performance measures of models provide recommendation for further data collection. Because predicted absence was globally more likely to be true than presence (see also Manel et al. 1999), priority areas to (re)sample should be those where presence was predicted, i.e. where a species’ occurrence is more uncertain. Lower priority should be given to the sampling of areas with predicted absences, because the given species is very unlikely to occur in these areas. Also, by using such models, the logistical and financial problems intrinsic to surveys of remote areas could be reduced (Manel et al. 1999).

Future developments might include improvement of the models by adding other descriptors of landscape, such as the presence of caves, slope aspect, geomorphology or distance from landscape features other than lakes (e.g. distance from caves or rivers). Adding more accurate data to the models, e.g. on landscape management, food availability or climate (for larger spatial extents), could greatly improve their predictive power. Shifts in species-specific habitat use, observed in this study with Myotis myotis, suggest that bat distribution models could be improved by splitting species data between observation periods or demographic groups. It is well known that individuals of the same species can have different ecological requirements depending on their reproductive status or the time of the year (Anthony & Kunz 1977; Arlettaz 1995). Finally, as suggested by Manel et al. (1999), the models should be applied to new sites in order to test the hypothesis that a cause-and-effect relationship links landscape structure and bat distribution.


We would like to thank Jean-Daniel Blant, Michel Blant, Pascal Moeschler and all who helped in the field work. Financial support for the field work was provided by the Service de la faune du Canton de Neuchâtel and the Federal Office of the Environment, Forests and Landscape. Bat capture and handling were carried out with the permission of the Service de la faune du Canton de Neuchâtel. We are grateful to the Federal Office of Statistics and the Federal Office of Topography who provided some landscape data. We are indebted to Jean-Marc Weber who made helpful comments on an earlier draft of the manuscript, and to Gillian Kerby and Steve Ormerod who helped to improve the manuscript. Julie Warrillow and Benoit and Cristina Molineaux revised the English.