Habitat Availability Is Not Limiting the Distribution of the Bohemian–Bavarian Lynx Lynx Lynx Population

Home , Bavarian Forest, Bohemian Forest, Brdy, Fichtel Mountains, Franconian Forest, Franconian Jura

Habitat availability is not limiting the distribution of the Bohemian–Bavarian lynx Lynx lynx population

N ORA M AGG,JÖRG M ÜLLER,CHRISTOPH H EIBL,KLAUS H ACKLÄNDER S YBILLE W ÖLFL,MANFRED W ÖLFL,LUDÊK B UFKA J AROSLAV Č ERVENÝ and M ARCO H EURICH

Abstract A population of Eurasian lynx Lynx lynx was es- c.  (–) resident lynx. We assessed connectivity via tablished by reintroductions in the Bohemian Forest least-cost paths and found that all suitable patches could Ecosystem in the s and s. The most recent informa- be reached by the lynx. A comparison with the current dis- tion on the population status indicates that the distribution tribution of lynx, however, confirms that a significant pro- has stagnated since the late s, for unknown reasons. We portion of suitable habitat is not occupied, which indicates assessed the availability of suitable habitat along the that the distribution is limited by factors other than habitat, Austrian–German–Czech border, and hypothesized that with illegal killing being the most likely cause. Our study the Bohemian–Bavarian lynx population is not in equilib- provides crucial information for the development of a con- rium with habitat suitability. Based on global positioning servation strategy and regional planning for the Bohemian– system data from  radio-collared lynx, we used a max- Bavarian lynx population. imum entropy approach to model suitable habitat. Keywords Habitat connectivity, home range, large carni- Variables reflecting anthropogenic influence contributed vore conservation, least-cost paths, Lynx lynx, radio track- most to the model and were negatively associated with the ing, species distribution modelling occurrence of lynx. We evaluated the model prediction using independent records of lynx from monitoring in To view supplementary material for this article, please visit Bavaria, Germany. Using our habitat approach we estimated http://dx.doi.org/./S the area of potential habitat, based on a mean annual home   range of  km for males and  km for females. Our re-  sults indicated there were , km of suitable habitat, distributed among  patches, for a potential population of Introduction fter decades of persecution and extermination of large Acarnivores in Europe, populations have started to re- ‡ † NORA MAGG* (Corresponding author) and JÖRG MÜLLER Bavarian Forest cover (Chapron et al., ). Management policy has im- National Park, Freyunger Str. 2, 94481 Grafenau, Germany E-mail [email protected] proved considerably and large carnivores are protected by

CHRISTOPH HEIBL Department of Ecology and Ecosystem Management, law in most European countries (Molinari-Jobin et al., Technische Universität München, Freising, Germany ). They are recolonizing their former ranges both natur-  KLAUS HACKLÄNDER Department of Integrative Biology and Biodiversity ally and through reintroduction (Linnell et al., ); how- Research, Institute of Wildlife Biology and Game Management, University of ever, they are confronted with a human-dominated Natural Resources and Life Sciences, Vienna, Austria landscape, where their habitats are diminished and frag- SYBILLE WÖLFL Lynx Project Bavaria, Lam, Germany mented as a result of direct destruction and the development MANFRED WÖLFL Bavarian Environmental Agency, Hof/Saale, Germany of roads and railways (Fischer & Lindenmayer, ). With LUDÊK BUFKA§ Šumava National Park, Kašperské Hory, Czech Republic the reduction in habitat, wildlife populations become smal- JAROSLAV ČERVENÝ Faculty of Forestry and Wood Sciences, Czech University of ler and more isolated, both of which increase the risk of local Life Sciences Prague, Czech Republic extinction. Large carnivores are particularly vulnerable to MARCO HEURICH** Bavarian Forest National Park, Grafenau, Germany local extinction in fragmented environments because they *Also at: Department of Integrative Biology and Biodiversity Research, Institute require large contiguous spaces, and their populations are of Wildlife Biology and Game Management, University of Natural Resources low in density (Ripple et al., ). and Life Sciences, Vienna, Austria †Also at: Department of Ecology and Ecosystem Management, Technische For the Eurasian lynx Lynx lynx, stagnation and declines Universität München, Freising, Germany of reintroduced and formerly increasing populations have §Also at: Faculty of Forestry and Wood Sciences, Czech University of Life – Sciences Prague, Czech Republic been reported (e.g. the Vosges Palatinian population in **Also at: Chair of Wildlife Ecology and Management, University of Freiburg, France; the Dinaric population in Slovenia, Croatia, Bosnia Faculty of Environment and Natural Resources, Freiburg, Germany – ‡ and Herzegovina; the Bohemian Bavarian population along Current address: Forest Research Institute of Baden-Württemberg, Freiburg, – – Germany the Austrian German Czech border; Kaczensky et al.,  Received  October . Revision requested  October . ). The latter population originated from lynx captured Accepted  March . First published online  August . in the Carpathian Mountains and reintroduced to the

Oryx, 2016, 50(4), 742–752 © 2015 Fauna & Flora International doi:10.1017/S0030605315000411 Downloaded from https://www.cambridge.org/core. IP address: 170.106.35.234, on 29 Sep 2021 at 23:20:00, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0030605315000411 Distribution of Eurasian lynx population 743

Bohemian Forest in the s and s (Wölfl et al., ). suitable habitat and of lynx, and estimated a potential popu- The population resides in the Bohemian Forest Ecosystem, lation, considering its entire transnational range. in which the Bavarian Forest National Park and the We developed a habitat suitability model with a max- Šumava National Park are embedded. Initially data indicated imum entropy approach for the Bohemian–Bavarian lynx an increase in numbers and distribution but at the end of population, based on GPS telemetry data for  individuals the s population increase stagnated, and in the Czech in the region; considered spatial requirements of lynx, using part of the Bohemian Forest the number of individuals de- estimates of annual home range; assessed the size and con- creased (Wölfl et al., ). The latest status report on the nectivity of suitable habitat patches; estimated a potential population also presumes stagnation in population size population size for the area along the border of Germany, and range in Germany and Austria. It is estimated the the Czech Republic and Austria; and compared our results Bohemian–Bavarian population comprises – indivi- with regional occurrence of lynx according to the latest sta- duals (Kaczensky et al., ). tus report (Kaczensky et al., ). We used the Bohemian The reason for the stagnation of this population is of con- Forest Ecosystem, with its characteristic Central European cern for lynx conservation in Central Europe because of the lynx population and low mountain range habitat, as a population’s central geographical location and thus its model area. Our objective was to provide wildlife managers potential to act as a link between other small and isolated and administrators involved in the monitoring and manage- populations. Isolation of a small population may result in ment of populations of large carnivores in Europe with a reduced genetic variability, lower reproductive success and state-of-the-art tool to compile information about local increased risk of extinction (Schmidt, ). Human- habitat suitability and carrying capacity for the conservation induced mortality may have an even greater impact than of a population that, as we hypothesize, is not realizing its low reproductive success (Linnell et al., ), particularly potential distribution. mortality caused by road traffic or illegal killing (Andrén et al., ; Molinari-Jobin et al., ; Müller et al., ). Habitat availability and connectivity may influence popula- Study area tion dynamics and thus may be the factor limiting the dis- The Bohemian Forest Ecosystem, Europe’s largest region of tribution of the Bohemian–Bavarian lynx. strictly protected forest, comprises a forested mountain Information on the spatial distribution, size and connect- range along the German–Czech border and includes the ivity of suitable habitat patches is a prerequisite when com-  Bavarian Forest National Park ( km ) on the German paring actual and potential lynx distribution to maximize side of the border and the Šumava National Park ( the success of population conservation. Models based on a  km ) on the Czech side (Fig. a). Human population dens- geographical information system provide an effective means  ities are comparably low: ,  inhabitants per km in the of gathering information about spatial distribution and ex-  National Parks and c.  per km in nearby regions. Mean panse of potential habitat. Previous analyses of lynx habitat annual temperature is .°C in montane and .°C in sub- have highlighted the importance of forests, and the negative alpine elevation zones. Snow cover persists for – association of lynx presence with habitat fragmentation, days, depending on altitude. GPS data from radio-collared human settlements and areas of intensive land use  lynx were collected in an area of , km within the (Niedziałkowska et al., ; Breitenmoser-Würsten et al., Bohemian Forest. This area, defined by the % minimum ; Basille et al., ). In previous models of lynx habitat convex polygon (MCP) of all lynx localizations, served as in Germany habitat suitability was assessed based on the the model training area (Fig. a). Sixty-six percent of this availability of forest cover (Schadt et al., a), and using area is covered by forest and % is used for agriculture, pri- VHF-telemetry data for  lynx in the Swiss Jura Mountains marily grassland. Model prediction covered a more exten- (Schadt et al., b). Thus there was a risk of predictive un-  sive area (, km ) along the borders between certainty because the model was extrapolated beyond the Germany, the Czech Republic and Austria (Fig. a), %of domain of the data (Pearson et al., ). At the time of which is covered with forest and % used for agriculture. our study, novel techniques for species distribution modelling and global positioning system (GPS) telemetry data for – lynx of the Bohemian Bavarian population were available, Methods which facilitated a qualified assessment of habitat in this area by staying within the domain of the data (Guisan & Habitat data Zimmermann, ). Müller et al. () analysed monitoring data for lynx in Bavaria and found occurrence of lynx We configured landscape variables based on CORINE was predicted most clearly by proximity to national park (Coordination of Information on the Environment) land areas. Whereas these authors sought predictors to explain cover  with a resolution of  m (European lynx occurrence, we focused on the spatial distribution of Environment Agency, ). The multiple land-use features

FIG. 1 (a) The study area along the borders between Germany, the Czech Republic and Austria, for which MaxEnt predictions were obtained for distribution of Eurasian lynx Lynx lynx. (b) MaxEnt prediction of habitat suitability for Eurasian lynx in the study area. Patches with habitat suitability index . . and a coherent  area .  km are considered suitable (Table ). Least-cost paths between these patches are indicated. White circles denote records of lynx of Status and Conservation of the Alpine Lynx Population categories C and C,in Bavaria, Germany.

of CORINE were pooled into  categories (Supplementary proportion per grid cell to maintain data quality when Table S). We generated one variable that specified the coarsening spatial resolution (Seebach et al., ). We Euclidean distance to human settlements, and another for included altitude as a variable because analyses of faecal altitude (Hijmans et al., ; Table ). All variables were pellet count indicated a negative correlation between alti- continuous because we calculated land use as the tude and the density of roe deer Capreolus capreolus

TABLE 1 Predictor variables used to describe habitats in the sampled without replacement  locations for each individ- Bohemian Forest Ecosystem (Fig. a), with their definition, unit, ual to avoid uneven contribution. However, for three indivi- range, and contribution to the MaxEnt model based on GPS data duals only ,  and  locations were available, resulting  from radio-collared lynx Lynx lynx and environmental variables in a total of , locations used. These provided informa- derived from CORINE land-cover data, given as percentage contribution in regularized log-likelihood distribution compared to a tion about daytime resting sites as well as crepuscular and uniform distribution. Estimates were determined at each iteration nocturnal habitat use. We assume that the locations used re- − of the training algorithm (convergence threshold  ×  ), starting present a random sample of lynx occurrence because, in a with a uniform distribution. The increase in regularized gain was test, reception of GPS collars in various habitats was high added to or subtracted from the contribution of the corresponding (%) and independent of habitat structure (Heurich variable. Product, linear, quadratic and hinge features were used et al., unpubl. data).   with regularization values of . for the first three features and With repeated sampling of one individual it may be neces- . for the hinge feature (Phillips & Dudík, ; Elith et al., ). sary to consider the individuality of the lynx. As MaxEnt does % not include random effects, we fitted a generalized linear Variable Definition Unit Range contribution mixed model beforehand to examine individuality. Using Pasture Pastures % 0–100 23.7 the same data set as used subsequently for MaxEnt we tested Acre Non-irrigated %0–87 18.7 peculiar individual effects by sampling background points arable land individually for each lynx in its spatial range. Randomly – Altitude Altitude m 466 15.0 sampled background points facilitate presence-only model- 1,330 ling, and rather than representing species’ absence they Human Human settle- %0–88 11.7 ments, industry, characterize the environment. Tested individual effects were ,   artificial surfaces negligible (random effect standard deviation . %offixed Distanc_hum Distance to m0– 9.1 effect intercept; Supplementary Table S). human settle- 9,444 ments, industry, artificial surfaces Woodshrub Transitional %0–100 5.8 Distribution modelling woodland shrub & wetlands To model the potential distribution of the lynx population    Forestconif Coniferous forest % 0–100 4.4 we used MaxEnt v. . (Phillips et al., ). MaxEnt at- Naturalagri Agricultural areas %0–84 4.4 tempts to calculate the probability distribution of a species’ with significant presence based on the constraints derived from environ- natural vegetation mental data at the presence locations. Following the prin- or complex culti- ciple of maximum entropy, the method remains as close vation patterns as possible to uniform distribution. We used , back- Forestmix Mixed forest % 0–100 2.9 – ground points to achieve the best possible results (Phillips Forestleaves Broad-leaved %094 2.6   forest & Dudík, ; Barbet-Massin et al., ). Natgrassl Natural grassland % 0–100 1.8 As the choice of resolution may affect model predictions Water Inland water %0–48 0 (Guisan et al., ), we aimed to detect the most suitable bodies grid cell size by calculating MaxEnt models under default settings for grids with edge lengths of , , , ,, , and , m. Following Guisan et al. () we did not (Heurich et al., ), which is the main prey of the regional reduce the number of records by removing duplicate records lynx (Heurich et al., ). in the same grid cell, but left sample size consistent between the models of various grid cell sizes. Each model was fitted and evaluated five times by k-fold cross-validation, and thus  Lynx presence data each model had random samples of % lynx telemetry data for training and % for testing (Hirzel et al., ). Model The habitat model was based on GPS data from  radio- performance was assessed by comparing the means of the collared lynx (six males and four females). The animals area under the receiver operating characteristic curves – were caught in baited box traps and fitted with GPS GSM (AUC). With presence-only (PO) modelling, AUCPO de- collars (VECTRONIC Aerospace, Berlin, Germany; see picts the probability that a randomly chosen presence site Podolski et al. () for details). The lynx were collared dur- will be ranked above a randomly chosen background site, ing –; the operating time of individual collars was whereby a random ranking (reflecting uniform distribution)  –   . months. To limit autocorrelation we standardized achieves an AUCPO of . and a nearly perfect ranking ap-  sampling frequency to one location per day and randomly proaches an AUCPO of (but with the use of background

 locations it never achieves this; Wiley et al., ). AUCPO most consistent approach. We assumed all values represent- values vary with the spatial extent used to select background ing suitable habitat were above the point at which the dens- points as well as with the number of background points; be- ity of presence locations exceeded that of the background cause we only modified grid cell sizes but not the extent of locations that characterized the environment. the area or the number of background points, value com- For the potential population estimate we considered only parison was possible. suitable habitat patches larger than the mean home range of a In the model prediction we aimed for an output grid female lynx, and calculated the percentages of suitable habitat comprising a map of suitability scores. We used the logistic cells within their spatial ranges (for MCP,MCP,MCP). output to obtain a relative habitat suitability index of –, To assess the connectivity between suitable habitat with  indicating unsuitable habitat and  indicating highly patches we calculated least-cost paths based on a resistance suitable habitat. grid at a resolution of  ×  m. The resistance grid re- To select only raster cells for which we were able to make sulted from an overlay of the reciprocal resampled habitat reliable predictions (i.e. cells characterized by a parameter suitability map (low habitat suitability index implies high re- composition similar to that of cells inside the model training sistance and vice versa; LaRue & Nielsen, ) and a raster- area) we conducted a principal component analysis for all ized road layer (ESRI, Redlands, USA) with resistance values raster cells within the model prediction area, including the assigned according to road status (Hebblewhite et al., ). training area. Positions of raster cells were mapped in a scat- We assigned values within the range of the inverse habitat terplot and a range of predictability was set on the basis of suitability, with the highest cost assigned to highways the distribution of raster cells characterizing the model (Autobahn, resistance value = ), medium cost assigned to training area. We predicted the model only on raster cells main roads (Bundesstraße, resistance value = .), and located within this range. low cost assigned to municipal roads (Landstraße, resistance value = .). Least-cost paths start at the centre of the patch of departure and end when the contour of the target patch is Model evaluation reached. We extracted the path lengths (.  km) between the patch contours to define the distances that lynx must To assess model accuracy by AUC we used monitoring data travel when moving from one patch to another. For the as- recorded for lynx in eastern Bavaria during –,a sessment of patch connectivity we also considered  VHF data set independent of our GPS telemetry locations used telemetry positions of a subadult dispersing lynx, recorded for training the model. We recorded observations of lynx, in . The comparison of suitable patches with lynx occur- using the categories of the Status and Conservation of the rence in the region according to the latest status report Alpine Lynx Population project (SCALP; Molinari-Jobin (Kaczensky et al., ) resulted from a spatial overlay. et al., ): category C denotes proof of lynx presence MaxEnt modelling, generalized linear mixed modelling, (dead or captured lynx, photographs, genetic samples); C and home range computations were carried out in Rv... accounts for confirmed observations (e.g. lynx kill or paw (R Development Core Team, ) using the packages dismo prints confirmed by an expert); C accounts for uncon- (Hijmans et al., ), lme (Bates et al., ), and firmed observations. We used C and C observations for adehabitatHR (Calenge, ). We used ArcMap v. . model evaluation, resulting in a data set of  observations. (ESRI, Redlands, USA) to calculate least-cost paths and cre- ate maps. Habitat quantification and connectivity

To estimate the potential population size we needed information Results on the spatial requirements of lynx and on the amount of suitable habitat. We assessed spatial requirements by calculating Distribution modelling %MCPs(MCP)forhomerangesizeinaccordancewith previous studies (Herfindal et al., ; Breitenmoser-Würsten We tested the MaxEnt model at various grid cell sizes and se- et al., ), and chose %MCPs(MCP)aswellas% lected the , × , m grid cell size for further analysis be- MCPs (MCP) to obtain a range of population estimates. We cause it yielded the best model performance (Supplementary selected time periods of  days and standardized the sampling Table S). We limited the range of model predictability on frequency to one location per day. the basis of the distribution of the model training area cells To quantify available habitat it is necessary to convert in the principal component analysis scatterplot (Fig. ). continuous habitat suitability index values into binary re- Twenty-three percent of cells were thus excluded, including sults. There are various approaches that can be used to select cells with high percentages of human settlements (human) a threshold (Liu et al., ). We used a new approach based or arable land (acre), which differed most from cells inside on our model evaluation, which we considered to be the the model training area.

FIG. 3 Habitat suitability of  locations along the borders between Germany, the Czech Republic and Austria (Fig. a) where presence of lynx was confirmed, and , random background locations.

FIG. 2 Scatterplot resulting from principal component analysis of the model prediction area, including the training area (Fig. a). index values of c. .–.; in contrast, background locations  The lines are vectors of the environmental variables (Table ). with the highest densities of observations had habitat suitability values of .–.. With the aim of quantifying available habitat we selected the intercept of the density curves, at Contributions of the specific environmental variables to   the MaxEnt model are in Table . Response curves a habitat suitability index of . , as our threshold and de-  fined habitat suitability index values $ . as suitable habi- (Supplementary Fig. S ) provide more information on the  relationship between environmental variables and the habitat. According to this threshold % of the Status and tat suitability index. Anthropogenic open-space land-use Conservation of the Alpine Lynx Population evaluation types contributed most to the model and were negatively data comprise suitable habitat. correlated with habitat suitability index. The occurrence of human settlements, industry or artificial surfaces (variable human; Supplementary Fig. S) was linked to a comparably Habitat quantification, connectivity and occupancy low habitat suitability index, and the response curve of variable distanc_hum demonstrated an increase in habitat suit- Mean home range sizes for male and female lynx were  ±    ±   ability index with increasing distance to human settlements. SD km and SD km , respectively. Home Transitional woodland shrub and wetlands (variable ranges of males were significantly larger than those of fe-     woodshrub) and the three forest types broad-leaved, conifer- males (W = ,P= . ; Table ). For the lynx named  ous, and mixed provided comparatively suitable habitat, but Patrik we only considered a -day period because he for the variables forestmix (mixed forests) and forestconif shifted his home range after that period. (coniferous forest) the increase in habitat suitability index Percentages of suitable habitat cells within the home     was marginal. High percentages of agricultural areas with ranges (MCP , Table ) were % for females and %   significant natural vegetation or complex cultivation pat- for males. We therefore assumed a habitat area of km         terns (variable naturalagri) as well as high altitudes (variable ( × . ) for an average female and km ( × . ) altitude) resulted in a lower habitat suitability index. By pro- for an average male. Consequently, we calculated a density     jection of the trained MaxEnt model onto the model predic- of . resident lynx per km of suitable habitat, resulting  tion area we created a habitat suitability map (Fig. b). in a potential population size of c. individuals (distributed among  patches; Fig. b, Table ). The percentage of suitable habitat cells within MCP (Table ) was %  Model evaluation and threshold selection for an average female lynx and % for a male. The percentage of suitable habitat cells within MCP (Table )was Threshold-independent extrinsic model evaluation resulted % for an average female and % for a male, resulting   –  in an AUCPO of . . However, habitat suitability index va- in an estimated population of individuals (Table ). lues are also of interest. Lynx were observed in areas with The results of the least-cost path analyses (Fig. b; higher habitat suitability index values (median = .) Supplementary Table S) depict the connectivity between than background locations (median = .; Wilcoxon these patches. The shortest least-cost path between these test: W = ,,,P, .). The density plot (Fig. ) patches was  km; the longest was  km. Comparison of displays similar information, namely that highest densities suitable habitat patches with reported lynx occurrence of observations were found in areas with habitat suitability (Kaczensky et al., ) indicated permanent occupancy in

TABLE 2 Home range estimations (based on minimum convex polygon, MCP) for radio-collared adult lynx in the Bohemian Forest Ecosystem (Fig. a) during –, with name and sex of individuals, number of kittens, MCP, MCP, MCP, time period of data collection, and number of locations recorded.

Name Sex No. of kittens MCP90 (km2) MCP95 (km2) MCP100 (km2) Time period (days) No. of locations Ctirad Male 322 462 688 365 351 Emanuel Male 191 225 293 100 85 Kika Male 216 257 369 365 347 Kubicka Female 2 130 145 187 365 334 Matilda Female 2 96 101 175 365 348 Milan Male 509 532 693 365 335 Nimo Male 507 585 622 365 288 Nora Female 1 142 166 186 365 217 Patrik Male 354 389 625 300 267 Tessa Female 2 64 77 127 365 354 Mean female 108 122 187 Mean male (excluding Emanuel*) 382 445 599

*Excluded because GPS locations were recorded only on  days

TABLE 3 Estimated area of suitable lynx habitat patches in the study area (Fig. a), with potential number of lynx, and whether occupancy is permanent or sporadic.

Patch region Area of suitable Potential no. of Occupancy according to Kaczensky et al. (2013)(No.of10×10km (Fig. 3) habitat (km2) lynx1 (range2) grid cells with permanent presence overlaying the patch) Bohemian Forest 3,904 46 (31–52) Permanent presence (40) Erz Mountains 1,726 20 (13–23) Sporadic occurrence Forest Quarter 1,300 15 (10–17) Permanent presence (3) Krkonoše Mountains 954 11 (7–12) Sporadic occurrence Upper Palatinate Forest 937 11 (7–12) Permanent presence (5) Franconian Forest 690 8 (5–9) No occurrence detected Slavkovský les 686 8 (5–9) Sporadic occurrence Brdy 514 6 (4–6) Permanent presence (2) Fichtel Mountains 481 5 (3–6) Sporadic occurrence Patch CZ South 385 4 (3–5) Sporadic occurrence Northern Franconian Jura 326 3 (2–4) No occurrence detected Labské pískovce 289 3 (2–3) Sporadic occurrence Patch CZ North 223 2 (1–2) No occurrence detected Total 12,415 142 (93–160)

  Based on . per  km suitable habitat   Based on .–. per  km suitable habitat

four patches, sporadic occurrence in six patches, and no oc- (Kaczensky et al., ) revealed permanent lynx presence currence in three patches (Table ). in four of these patches (Bohemian Forest, Upper Palatinate Forest, Brdy, Forest Quarter; Table ). Of these, there is a permanent patch-wide distribution of lynx only Discussion in the Bohemian Forest. This represents the source of the population and includes the Bavarian Forest National In our study area we estimated a habitat carrying capacity Park and the Šumava National Park, which have been for a lynx population of c.  individuals (based on popu- shown to be of significant importance for lynx conservation  lation estimates of  (MCP)to (MCP) indivi- in the region (Müller et al., ). Nine patches (, km in duals), distributed across  patches (Fig. b, Table ). This total) are not permanently occupied by lynx. estimate refers to territorial lynx and does not include the We analysed least-cost paths to assess connectivity be- population of subadults, which can be –% of the resident tween suitable habitat patches and considered available in- population (Jędrzejewski et al., ; Zimmermann & formation on regional lynx distribution. In Bavaria there is Breitenmoser, ). A spatial comparison of potential evidence (Status and Conservation of the Alpine Lynx and actual lynx distribution with the latest status report Population C and C categories) that lynx are able to

reach the Fichtel Mountains and Franconian Forest in the does not require thresholding habitat suitability data north-west of our model prediction area (Fig. b). (Boyce & McDonald, ) would be informative. Such an Connectivity to Brdy is not immediately apparent but approach has been applied in large carnivore conservation VHF positions of a subadult lynx reveal that Brdy is access- (Hebblewhite et al., ), but a reliable estimate of the ible (Supplementary Fig. S). Reported lynx distribution size of the existing population as well as a detailed delinea- (Kaczensky et al., ) confirms accessibility of the remain- tion of its range are needed as references. ing patches except for Northern Franconian Jura and Patch In accordance with results of previous studies (Schadt CZ North (Table ). Moreover, data on lynx dispersal in et al., b; Niedziałkowska et al., ) our MaxEnt Switzerland (– km; Zimmermann et al., ) and model of lynx distribution indicated that lynx avoid areas Scandinavia (– km; Samelius et al., ) indicate the close to human settlements or with intense anthropogenic ability of lynx to access suitable patches in our model predic- disturbance. Similarly, Basille et al. () reported that tion area. The roads located between suitable patches in our lynx avoid agricultural areas, but in contrast to our study model prediction area are mostly main roads and municipal they did not find that lynx avoided areas close to artificial roads, which pose risks but not barriers for lynx. Highway areas. Previous studies have emphasized the importance of A, which separates the Franconian Forest from the other forest for lynx (Schadt et al., b; Niedziałkowska et al., patches, may be more difficult to cross but a category C ob- ; Basille et al., ; Müller et al., ). Our model servation indicated that crossing is possible. Thus, we as- was not primarily determined by variables linked to forest, sume all patches are reachable for lynx, but highlight that possibly because of the comparatively high availability of fragmentation of patches needs to be considered in land- forest (% of the model training area). Our results also in- scape planning to prevent degradation of connectivity. dicated that in the study area lynx preferred areas of lower The habitat map (raster file available on request) may be altitude, which is in accordance with the results of Basille used for this purpose by administrators and wildlife man- et al. () and matches the increasing density of roe agers, and also as a basis for planning lynx reintroductions. deer with decreasing altitude (Heurich et al., ). Prey Besides increasing numbers of lynx and thus raising the density could not be included in the model because no probability of connecting with isolated populations, reintro- transnational data were available, but we expect prey dens- ductions help to overcome deleterious effects of any poten- ities in the model prediction area to be equivalent to or high- tial inbreeding depression (Johnson et al., ). The genetic er than within the forest-dominated model training area. status of the Bohemian–Bavarian population is unknown as We base this assumption on the roe deer density index in genetic analyses are lacking. Nevertheless, camera-trap data Bavaria (Hothorn et al., ) and on previous records of (Supplementary Table S; for methods see Weingarth et al., high roe deer densities in less-forested areas (Melis et al., ) indicate substantial reproduction in the Bohemian ). An advantage of the use of coarse transnational Forest patch, and lynx dispersal is supported by sporadic oc- data was the ability to consider the entire Bohemian– currence in most patches. Despite this, lynx are apparently Bavarian population, unlike the study of Müller et al. unable to establish permanent subpopulations there, and in (), which was restricted to Bavaria. However, we are line with other authors (Wölfl et al., ; Červený et al., aware that we disregarded fine-scale habitat characteristics ; Molinari-Jobin et al., ; Müller et al., ) we as- of forests, such as essential cover for den sites, daily resting sume that illegal killing is the most likely factor limiting the sites (Podgórski et al., ), and cover from hunting distribution of the population. (Dickson & Beier, ). Considering the intentions of Our home-range-based population estimate is higher our study, however, we assume these characteristics to be than that of Schadt et al. (a,b), which extrapolated negligible as forest should provide them sufficiently. data for Swiss lynx to Germany and thus had predictive un- We used MaxEnt to assess habitat suitability, a  certainty. We estimated there was up to , km of suit- presence-only method that is designed to make predictions able habitat in the prediction area. Schadt et al. (a,b) based on incomplete data (Elith et al., ). The predictive estimated a smaller total area of suitable habitat along the accuracy of MaxEnt is high, which makes it consistently Austrian–German–Czech border but they did not consider competitive among the highest-performing methods the entire range of the Bohemian–Bavarian population but (Guisan et al., ; Phillips & Dudík, ). MaxEnt was rather the different spatial requirements of lynx. Schadt et al. not developed primarily for use with telemetry data (b) assumed an overlap of one male per female lynx (Baldwin, ), which are characterized by repeated sam-  within  km and reported an approximate estimate of pling of one individual. However, as reception of GPS data  individuals. Schadt et al. (a) modelled  potential on lynx locations did not vary with habitat the detection territories, considering the spatial requirements of Swiss probability of telemetry data is consistent across a study lynx. Our analyses resulted in a higher potential population area and the data represent a random sample of presence lo- estimate of  (–). For further comparisons of cations, thereby meeting the basic assumptions of MaxEnt population estimation, a habitat-based approach that (Yackulic et al., ).

The model was complex because lynx presence could not Chapron et al. () indicates that factors other than habi- be associated with a particular variable, but evaluation re- tat limit the distribution, with illegal killing being the most vealed good model performance according to the AUCPO. likely factor. We thus recommend focusing future research Qualification of the AUC as a measure of model accuracy on dispersing subadult lynx, with the aim of increasing ac- has been criticized (Yackulic et al., ); nevertheless, our ceptance of lynx by people. Secondly, the potential carrying approach of thresholding based on model evaluation cate- capacity of habitat in the study area is not sufficient for a gorized % of the Status and Conservation of the Alpine long-term viable population, which according to expert Lynx Population data correctly. Our approach therefore opinion should comprise c.  individuals (Linnell et al., equals a cut-point probability at %, compared with the ). Thus, the population requires connectivity to other cut-point probabilities of Hebblewhite et al. (), who populations (i.e. Czech Carpathian, German Harz, and chose % with Amur tiger Panthera tigris altaica tracks, Austrian Alpine populations). Thirdly, patches of suitable and Hebblewhite et al. (), with % for Far Eastern leo- habitat are fragmented; their distribution needs to be con- pards Panthera pardus orientalis. sidered in landscape and road development to prevent fur- As the scale of predictor variables affects the model pre- ther fragmentation, and when planning and locating sites diction, we evaluated model performance for various grid for lynx reintroduction. cell sizes. Our results indicated that the habitat model per-  formed best using a grid cell of  km and that, in line with Guisan et al. (), a lower resolution does not necessarily Acknowledgements imply poorer performance. In contrast, the model per- formed best at a grid cell size five times coarser than the This study is part of the Eurasian lynx project of the smallest resolution tested. From an ecological point of Bavarian Forest National Park, Department of Research view, a comparably coarse grid cell size makes sense when and Documentation. Financial support was provided by modelling habitat of a species with large spatial require- EU Programme Interreg IV (EFRE Ziel ). We thank ments, such as the lynx, and facilitates the necessarily wide- O. Vojtěch, J. Mokrý, H. Burghart, M. Gahbauer, and ranging conservation planning (Chapron et al., ). other staff of Šumava National Park and Bavarian Forest Overly coarse grain is not effective, however; model per- National Park for their help with field work, and  formance decreased when a grid cell size of .  km was F. Gutkowski for analyses of GPS collar performance. We used. A study on the sensitivity of MaxEnt to scale differ- also thank Karen A. Brune for improving the writing of ences indicated that the maximum grain size should be the manuscript, and Miha Krofel and three anonymous re- c. . km (Song et al., ). viewers for their helpful comments. We assume our model predictions are reasonable as we did not extrapolate our model beyond the ranges of variable values used to train it. Others have cautioned against going References beyond the calibration range (Pearson et al., ; Hirzel & Le Lay, ), yet this is rarely considered in model applica- ANDRÉN, H., LINNELL, J.D.C., LIBERG, O., ANDERSEN, R., DANELL, A., KARLSSON, J. et al. () Survival rates and causes of mortality in tions. We benefited from lynx data originating from a part Eurasian lynx (Lynx lynx) in multi-use landscapes. Biological of the model prediction area, and by defining a range of pre- Conservation, , –. dictability we were able to predict % of cells within the BALDWIN, R.A. () Use of maximum entropy modeling in wildlife study area. research. Entropy, , –. The applied least-cost path method estimates the dis- BARBET-MASSIN, M., JIGUET, F., ALBERT, C.H. & THUILLER,W. () Selecting pseudo-absences for species distribution models: tances that lynx have to overcome between suitable habitat how, where and how many? Methods in Ecology and Evolution, , patches, but assumes that an individual has a planned des- –. tination when leaving a patch. We acknowledge that a more BASILLE, M., HERFINDAL, I., SANTIN-JANIN, H., LINNELL, J.D.C., complex stochastic movement simulator without this as- ODDEN, J., ANDERSEN, R. et al. () What shapes Eurasian lynx sumption could map dispersal paths more precisely distribution in human dominated landscapes: selecting prey or  – (Palmer et al., ). As we intended to estimate the dis- avoiding people? Ecography, , . BATES, D., MAECHLER,M.&BOLKER,B.() lme: Linear tances rather than determine the exact paths, we believe mixed-effects models using S classes. R package version the least-cost path method is sufficient. .-. Http://cran.r-project.org/package=lme [accessed  From our results we draw some conclusions regarding May ]. the conservation of a characteristic Central European lynx BOYCE, M.S. & MCDONALD, L.L. () Relating populations to population with the potential to connect small and isolated habitats using resource selection functions. Trends in Ecology & Evolution, , –. populations. Firstly, despite connectivity, the current distri- BREITENMOSER-WÜRSTEN, C., ZIMMERMANN, F., MOLINARI-JOBIN, – bution of the Bohemian Bavarian lynx population is not in A., MOLINARI, P., CAPT, S., VANDEL, J.-M. et al. () Spatial and equilibrium with available habitat, which in line with social stability of a Eurasian lynx Lynx lynx population: an

assessment of  years of observation in the Jura Mountains. Wildlife JĘDRZEJEWSKI, W., JĘDRZEJEWSKA, B., OKARMA, H., SCHMIDT, K., Biology, , –. BUNEVICH, A.N. & MIŁKOWSKI,L.() Population dynamics CALENGE,C.() The package “adehabitat” for the R software: a (–), demography, and home ranges of the lynx in tool for the analysis of space and habitat use by animals. Ecological Białowieża Primeval Forest (Poland and Belarus). Ecography, , Modelling, , –. –. ČERVENÝ, J., KOUBEK,P.&BUFKA,L.() Eurasian lynx (Lynx JOHNSON, W.E., ONORATO, D.P., ROELKE, M.E., LAND, E.D., lynx) and its chance for survival in Central Europe: the case of the CUNNINGHAM, M., BELDEN, R.C. et al. () Genetic restoration of Czech Republic. Acta Zoologica Lituanica, , –. the Florida panther. Science, , –. CHAPRON, G., KACZENSKY, P., LINNELL, J.D.C., VON ARX, M., HUBER, KACZENSKY, P., CHAPRON, G., VON ARX, M., HUBER, D., ANDRÉN,H. D., ANDRÉN, H. et al. () Recovery of large carnivores in Europe’s &LINNELL, J.D.C. () Status, Management and Distribution of modern human-dominated landscapes. Science, , –. Large Carnivores—Bear, Lynx, Wolf & Wolverine—in Europe. DICKSON, B.G. & BEIER,P.() Home-range and habitat selection Report to the European Commission. by adult cougars in southern California. Journal of Wildlife LARUE, M.A. & NIELSEN, C.K. () Modelling potential dispersal Management, , –. corridors for cougars in midwestern North America using least-cost ELITH, J., PHILLIPS, S.J., HASTIE, T., DUDÍK, M., CHEE, Y.E. & YATES, path methods. Ecological Modelling, , –. C.J. () A statistical explanation of MaxEnt for ecologists. LINNELL, J.D.C., SALVATORI,V.&BOITANI,L.() Guidelines for Diversity and Distributions, , –. Population Level Management Plans for Large Carnivores in Europe. EUROPEAN ENVIRONMENT AGENCY () Corine land cover— A Large Carnivore Initiative for Europe report prepared for the Germany. Http://www.corine.dfd.dlr.de/intro_en.html [accessed  European Commission. May ]. LINNELL, J.D.C., SWENSON, J.E. & ANDERSEN,R.() Predators and FISCHER,J.&LINDENMAYER, D.B. () Landscape modification and people: conservation of large carnivores is possible at high human habitat fragmentation: a synthesis. Global Ecology and Biogeography, densities if management policy is favourable. Animal Conservation, , –. , –. GUISAN, A., GRAHAM, C.H., ELITH,J.&HUETTMANN,F.() LIU, C., BERRY, P.M., DAWSON,T.P.&PEARSON, R.G. () Selecting Sensitivity of predictive species distribution models to change in thresholds of occurrence in the prediction of species distributions. grain size. Diversity and Distributions, , –. Ecography, , –. GUISAN,A.&ZIMMERMANN, N.E. () Predictive habitat MELIS, C., JĘDRZEJEWSKA, B., APOLLONIO, M., BARTOŃ, K.A., distribution models in ecology. Ecological Modelling, , –. JĘDRZEJEWSKI, W., LINNELL, J.D.C. et al. () Predation has a HEBBLEWHITE, M., MIQUELLE, D.G., MURZIN, A.A., ARAMILEV, V.V. greater impact in less productive environments: variation in roe &PIKUNOV, D.G. () Predicting potential habitat and population deer, Capreolus capreolus, population density across Europe. Global size for reintroduction of the Far Eastern leopards in the Russian Far Ecology and Biogeography, , –. East. Biological Conservation, , –. MOLINARI-JOBIN, A., KÉRY, M., MARBOUTIN, E., MOLINARI,P., HEBBLEWHITE, M., ZIMMERMANN, F., LI, Z., MIQUELLE, D.G., KOREN, I., FUXJÄGER, C. et al. () Monitoring in the presence of ZHANG, M., SUN, H. et al. () Is there a future for Amur tigers in species misidentification: the case of the Eurasian lynx in the Alps. a restored tiger conservation landscape in Northeast China? Animal Animal Conservation, , –. Conservation, , –. MOLINARI-JOBIN, A., MARBOUTIN, E., WÖLFL, S., WÖLFL, M., HERFINDAL, I., LINNELL, J.D.C., ODDEN, J., BIRKELAND NILSEN,E.& MOLINARI, P., FASEL, M. et al. () Recovery of the Alpine lynx ANDERSEN,R.() Prey density, environmental productivity and Lynx lynx metapopulation. Oryx, , –. home-range size in the Eurasian lynx (Lynx lynx). Journal of MÜLLER, J., WÖLFL, M., WÖLFL, S., MÜLLER, D.W.H., HOTHORN,T. Zoology, , –. &HEURICH,M.() Protected areas shape the spatial distribution HEURICH, M., BRAND, T.T.G., KAANDORP, M.Y., ŠUSTR, P., MÜLLER, of a European lynx population more than  years after J. & REINEKING,B.() Country, cover or protection: what shapes reintroduction. Biological Conservation, , –. the distribution of red deer and roe deer in the Bohemian Forest NIEDZIAŁKOWSKA,M.,JĘDRZEJEWSKI, W., MYSŁAJEK, R.W., Ecosystem? PLoS ONE, (), e. NOWAK,S.,JĘDRZEJEWSKA,B.&SCHMIDT,K.() HEURICH, M., MÖST, L., SCHAUBERGER, G., REULEN, H., ŠUSTR,P.& Environmental correlates of Eurasian lynx occurrence in Poland— HOTHORN,T.() Survival and causes of death of European roe large scale census and GIS mapping. Biological Conservation, , deer before and after Eurasian lynx reintroduction in the Bavarian –. Forest National Park. European Journal of Wildlife Research, , PALMER, S.C.F., COULON,A.&TRAVIS, J.M.J. () Introducing a –. ‘stochastic movement simulator’ for estimating habitat connectivity. HIJMANS, R.J., CAMERON, S.E., PARRA, J.L., JONES, P.G. & JARVIS,A. Methods in Ecology and Evolution, , –. () Very high resolution interpolated climate surfaces for global PEARSON, R.G., THUILLER, W., ARAÚJO, M.B., MARTINEZ-MEYER, E., land areas. International Journal of Climatology, , –. BROTONS, L., MCCLEAN, C. et al. () Model-based HIJMANS, R.J., PHILLIPS, S., LEATHWICK,J.&ELITH,J.() dismo: uncertainty in species range prediction. Journal of Biogeography, , Species distribution modeling. R package version .–. Http:// –. cran.r-project.org/package=dismo [accessed  May ]. PHILLIPS, S.J. & DUDÍK,M.() Modeling of species distributions HIRZEL, A.H. & LE LAY,G.() Habitat suitability modelling and with Maxent: new extensions and a comprehensive evaluation. niche theory. Journal of Applied Ecology, , –. Ecography, , –. HIRZEL, A.H., LE LAY, G., HELFER, V., RANDIN,C.&GUISAN,A. PHILLIPS, S., DUDÍK,M.&SCHAPIRE,R.() Maxent software, v. .. () Evaluating the ability of habitat suitability models to predict Https://www.cs.princeton.edu/~schapire/maxent/ [accessed  May species presences. Ecological Modelling, , –. ]. HOTHORN, T., BRANDL,R.&MÜLLER,J.() Large-scale PODGÓRSKI, T., SCHMIDT, K., KOWALCZYK,R.&GULCZYŃSKA,A. model-based assessment of deer–vehicle collision risk. PLoS ONE,  () Microhabitat selection by Eurasian lynx and its implications (), e. for species conservation. Acta Theriologica, , –.

PODOLSKI, I., BELOTTI, E., BUFKA, L., REULEN,H.&HEURICH,M. WILEY, E.O., MCNYSET, K.M., PETERSON, A.T., ROBINS, C.R. & () Seasonal and daily activity patterns of free-living Eurasian lynx STEWART, A.M. () Niche modeling perspective on geographic Lynx lynx in relation to availability of kills. Wildlife Biology, , –. range predictions in the marine environment using a RDEVELOPMENT CORE TEAM () R: A Language and Environment machine-learning algorithm. Oceanography, , –. for Statistical Computing. R Foundation for Statistical Computing, WÖLFL, M., BUFKA, L., ČERVENÝ, J., KOUBEK, P., HEURICH, M., Vienna, Austria. HABEL, H. et al. () Distribution and status of lynx in the border RIPPLE, W.J., ESTES, J.A., BESCHTA, R.L., WILMERS, C.C., RITCHIE,E. region between Czech Republic, Germany and Austria. Acta G., HEBBLEWHITE, M. et al. () Status and ecological effects of Theriologica, , –. the world’s largest carnivores. Science, , –. YACKULIC, C.B., CHANDLER, R., ZIPKIN, E.F., ROYLE, J.A., NICHOLS, SAMELIUS, G., ANDRÉN, H., LIBERG, O., LINNELL, J.D.C., ODDEN, J., J.D., CAMPBELL GRANT, E.H. & VERAN,S.() Presence-only AHLQVIST,P.etal.() Spatial and temporal variation in natal modelling using MAXENT: when can we trust the inferences? dispersal by Eurasian lynx in Scandinavia. Journal of Zoology, , Methods in Ecology and Evolution, , –. –. ZIMMERMANN,F.&BREITENMOSER,U.() Potential distribution SCHADT, S., KNAUER, F., KACZENSKY, P., REVILLA, E., WIEGAND,T.& and population size of the Eurasian lynx Lynx lynx in the Jura TREPL,L.(a) Rule-based assessment of suitable habitat and Mountains and possible corridors to adjacent ranges. Wildlife patch connectivity for the Eurasian lynx. Ecological Applications, , Biology, , –. –. ZIMMERMANN, F., BREITENMOSER-WÜRSTEN,C.&BREITENMOSER, SCHADT, S., REVILLA, E., WIEGAND, T., KNAUER, F., KACZENSKY,P., U. () Natal dispersal of Eurasian lynx (Lynx lynx)in BREITENMOSER, U. et al. (b) Assessing the suitability of central Switzerland. Journal of Zoology, , –. European landscapes for the reintroduction of Eurasian lynx. Journal of Applied Ecology, , –. SCHMIDT,K.() Factors shaping the Eurasian lynx (Lynx lynx) population in the north-eastern Poland. Nature Conservation, , Biographical sketches –. SEEBACH, L.M., STROBL, P., SAN MIGUEL-AYANZ,J.& N ORA M AGG develops guidelines for conservation of faunistic species BASTRUP-BIRK,A.() Identifying strengths and limitations of in managed forests. J ÖRG M ÜLLER is a conservation biologist inter- pan-European forest cover maps through spatial comparison. ested in the effects of varying management intensities on forest bio- International Journal of Geographical Information Science, , – diversity. C HRISTOPH H EIBL works on data analyses in ecology and . evolutionary biology. K LAUS H ACKLÄNDER focuses on wildlife ecol- SONG, W., KIM, E., LEE, D., LEE,M.&JEON, S.-W. () The ogy and develops management strategies for sustainable use, conserva- sensitivity of species distribution modeling to scale differences. tion or control of wildlife species. S YBILLE W ÖLFL works on lynx Ecological Modelling, , –. conservation and management. M ANFRED W ÖLFL works on wildlife WEINGARTH, K., HEIBL, C., KNAUER, F., ZIMMERMANN, F., BUFKA,L. management, with a focus on human dimensions and conflict mitiga- &HEURICH,M.() First estimation of Eurasian lynx (Lynx lynx) tion. L UDÊK B UFKA is a zoologist specializing in research and conser- abundance and density using digital cameras and capture–recapture vation of large carnivores. J AROSLAV Č ERVENÝ focuses on wildlife techniques in a German national park. Animal Biodiversity and biology and the ecology of mammals. M ARCO H EURICH is interested Conservation, ., –. in the ecology and conservation of natural forest ecosystems.