Amphibians are among the most threatened taxonomic groups, with approximately one third of the known species threatened with extinction (Stuart et al. 2004; Bellard et al. 2016; IUCN 2020). Emerging fungal diseases are among the main drivers of amphibian declines and extinctions worldwide (Wake and Vredenburg 2008). In particular, the fungal pathogen Batrachochytrium dendrobatidis (Bd), causing chytridiomycosis in frogs, has negative impacts on anuran populations globally (Skerratt et al. 2007). The related and invasive B. salamandrivorans (Bsal) causes chytridiomycosis mainly in newts and salamanders (Urodela) and can be seen as a severe threat to European salamander diversity (Martel et al. 2013; Spitzen-van der Sluijs et al. 2016; Yap et al. 2017), as infections are fatal in many European species (Martel et al. 2014).

Bsal was first detected in the Netherlands and Belgium, where it led to serious declines in fire salamander populations in the wild (Martel et al. 2013; Spitzen-van der Sluijs et al. 2013; Stegen et al. 2017). Since then, Bsal has spread quickly throughout Europe and is continuously detected in new sites, especially in Germany but also in Spain (e.g., Martel et al. 2020; Schmeller et al. 2020; Thein et al. 2020; Lötters et al. 2020a). An important characteristic for the spread of Bsal may be its ability to build encysted spores that can potentially persist in environmental reservoirs (Stegen et al. 2017; Mosher et al. 2018). The spores are probably transported by natural and anthropogenic vectors and can stick to amphibian skin or the feet of waterfowl (Stegen et al. 2017).

Monitoring and epidemiological modelling of infected fire salamander populations suggest that Bsal can induce local extinction within one season even at low host density (Schmidt et al. 2017; Stegen et al. 2017). Therefore, few mitigation options are available once Bsal has emerged (Stegen et al. 2017; Canessa et al. 2018; Thomas et al. 2019) and a strict implementation of biosecurity protocols in the field and in captivity to reduce human-mediated dispersal of spores is vital (EFSA et al. 2018; Scheele et al. 2019). To develop in situ measures that interrupt pathogen spread (e.g. by fencing) or mitigate disease severeness (e.g. by habitat management), a comprehensive understanding of the host-pathogen-environment interaction triangle is necessary. In frog species affected by Bd, pathogenicity can vary depending on environmental factors, such as water temperature and salinity (Heard et al. 2015, 2018); thus, habitat management addressing those factors may help to provide refugia from severe chytridiomycosis (Heard et al. 2018; Scheele et al. 2019). In such cases, infected amphibian populations can be stabilised by connectivity to environmental refugia, e.g. when local extinctions due to Bsal outbreaks are mediated by recolonization from Bsal free refugia habitats that harbour salamander source populations. However, refugia that reduce pathogenicity have not yet been reported in urodeles across the range of the invasive Bsal strain in Europe (Beukema et al. 2020).

It is difficult to assess the role of host population connectivity in spreading infectious diseases and for metapopulation persistence. In his models, Hess (1996) stressed an increased extinction risk of well-connected host populations affected by disease. However, the demographic advantages of immigration and recolonization are likely to outweigh the risk of the spread of diseases in most systems (McCallum and Dobson 2002; Heard et al. 2015). In the Bsal-fire salamander system, little is known about the role of salamanders as Bsal vectors, and empirical, spatiotemporal data on disease and host metapopulation dynamics are missing (Schmidt et al. 2017). A concentric, continuous spread mediated by fire salamanders can be expected in homogenous habitats with constant host densities (Vredenburg et al. 2010; Schmidt et al. 2017), but those conditions are unrealistic in the landscape matrix of forest fragments, grasslands, intensively used croplands and settlements typically found in Central Europe. For instance, in the Netherlands, an uninfected subpopulation of fire salamanders has been discovered at a distance of only 800 m from the Bsal index site, with no obvious barriers to amphibian migration present (Martel et al. 2013; Spitzen-van der Sluijs et al. 2018). Although the study of single populations on local scales is important to understand how Bsal spreads, these findings might not represent the role of host population connectivity on Bsal distribution at the landscape level (Canessa et al. 2018).

Germany can be considered the current hot spot of the invasive range of Bsal in Europe, and most records lie within the Eifel region (Lötters et al. 2020a). In the northern Eifel, Bsal was first detected in 2015 (Dalbeck et al. 2018), but reports on fire salamander mass mortality in the region predate more than ten years (Lötters et al. 2020b). Despite the presence of Bsal in the region, fire salamanders are (still) widely distributed and inhabit largely forested areas as well as smaller woodland fragments.

Here, we analysed the influence of fire salamander habitat connectivity on (i) the presence of Bsal and (ii) the local fire salamander abundance over a five-year period in the northern Eifel region. As anthropogenic land use can interrupt functional connectivity between geographically close populations (Marsh and Trenham 2001; Nowakowski et al. 2013), we used circuit theory (McRae 2006) to estimate the pairwise landscape resistance between fire salamander occurrences. Furthermore, we analysed the influence of habitat connectivity on local salamander abundance derived from standardized larval monitoring to assess whether local populations still benefit from connectivity at the current stage of pathogen invasion. We hypothesized that Bsal will occur more frequently in well-connected fire salamander populations, as local populations are more likely to come into contact with infected conspecifics. However, we also hypothesized that fire salamander abundance is higher in large and unfragmented habitats. This study helps to inform conservation actions targeting salamander habitat connectivity at the core of the known invasive range of Bsal in Europe.


Study system

The study took place in the northern part of the Eifel, a low mountain range spanning over the German federal states North Rhine-Westphalia and Rhineland-Palatinate as well as eastern Belgium. In this landscape, extended forests and a high density of near natural creeks that serve as potential sites for the deposition of salamander larvae and their development until metamorphosis (hereafter “breeding habitat”) are characteristic features. A landscape segment of approximately 30 × 40 km was chosen as the study area as it covers the most intensively sampled parts of the northern Eifel since 2014 (Lötters et al. 2020a). Here, yearly precipitation and mean annual temperature range from 1280 mm to 6.5 °C in the most elevated (~ 693 m) areas to 664 mm and 10.24 °C in the lowest parts (115 m) of the study area (Fick and Hijmans 2017).

Salamander presence, abundance and Bsal screening

The data on Bsal distribution as well as fire salamander abundance and distribution used in this study come from three independent sampling activities, which were conducted with different levels of standardization.

  1. (i)

    Bsal screening: We used a subset of 39 Bsal screening sites reported by Lötters et al. (2020a) that lay within our study area (for details, see Lötters et al. 2020a). Sampled individuals were assigned to a central point per valley (pers. commun. Dalbeck), which resulted in a relatively coarse resolution of Bsal occurrences (minimum distance between screening area centroids: 562 m). Bsal screening survey effort was neither recorded nor standardized.

  2. (ii)

    Larvae removal sampling: To assess the local abundance of salamanders, standardized removal sampling of salamander larvae was conducted between 2015 and 2019. Larvae were searched for 135 min per year and site in a stream segment of 75 m and captured individuals were kept in a small plastic cup until the end of the sampling process to avoid double counts (for details, see Wagner et al. 2020a). Because not all sites in the study area were continuously sampled, we used a subset of 30 sites that were visited in at least three years between 2015 and 2019 (SI1).

  3. (iii)

    Larvae presence-only surveys: We surveyed streams across the whole study area for the presence of fire salamander larvae to assess the overall (remaining) distribution of the species. These streams predominantly included first- and second-order streams in or close to forest areas, where predatory fish are unlikely to occur (Thiesmeier 2004). A torch (flashlight) was used to break water surface reflections to visually detect larvae, and potential shelters like stones and leaf litter in the water were systematically checked. Searches during rainfall were avoided to maximise the detectability of larvae. Each creek or creek segment survey was conducted once or twice in 2018 and/or 2019, with a maximum sampling effort of 30 min per person and year (total number of surveys = 246). Survey localities were recorded using a Garmin GPSMAP 64 S device. During all sampling activities, a strict hygiene protocol, including the disinfection of all material that came into contact with amphibians or their habitat, was followed (van Rooij et al. 2017).

Habitat suitability models

As sampling effort of salamander larvae surveys was not constant and the surveyed localities were complemented by opportunistic larvae records spanning from 2014 to 2019, including 30 removal sampling sites and larvae localities provided by the “Biologische Station im Kreis Düren”, fire salamander absence could not be confirmed. Therefore, a presence-only method, the maximum entropy species distribution modelling (MaxEnt), was used to map suitable salamander habitats (Phillips et al. 2006). In comparative studies, MaxEnt is ranked among the best performing presence-only methods (e.g., Giovanelli et al. 2010) since it can handle comparatively small sets of species occurrence data (Wisz et al. 2008; Aguirre-Gutiérrez et al. 2013). Presence data were contrasted to the environmental background sampled at 10,000 randomly distributed points, and raw MaxEnt outputs were interpreted as a relative measurement of habitat suitability (Merow et al. 2013).

Species’ environmental requirements can change according to life stages or seasonality, leading to spatially or temporally different habitat uses (Mateo Sánchez et al. 2014). Modelling frameworks can thus be split into multiple models that use different scales and predictors (Bani et al. 2015; Frans et al. 2018; Goudarzi et al. 2021). Therefore, fire salamander HSMs were split into (i) a spatially restricted breeding habitat and (ii) a whole life cycle habitat model, following the framework of Bani et al. (2015). The breeding habitat model covered a 30 m buffer along the digitised streams. Most streams surveyed were already available as digitised linear features (vector data) in the digital landscape model “DLM50” (status: 2019, see Geobasis NRW 2019a for details), but another 50 first-order creeks that were not included were manually digitised using satellite imagery and water flow accumulation. Following the recommendation of Phillips et al. (2009), the extent of the breeding model was restricted to streams with a maximum distance of 50 m to forest fragments with a minimum size of 1 ha, as larvae presence surveys were based on the same criteria. Background sampling of the whole life cycle models was not spatially restricted since most of the study area (82%) lies within a reasonable postnatal dispersal distance of 500 m to potential breeding habitats (Klewen 1985; Schulte et al. 2007; Merow et al. 2013).

Processing of fire salamander occurrence data

In presence-only habitat suitability models, it is vital to account for sampling bias (Dudík et al. 2005; Kramer-Schadt et al. 2013; Yackulic et al. 2013). Among others, spatial autocorrelation can effectively be reduced by filtering occurrence data with a minimum distance or a grid for systematic sampling (Veloz 2009; Syfert et al. 2013; Fourcade et al. 2014). Therefore, we thinned fire salamander larvae localities (n = 250) in the R environment (R Core Team 2019), using the package spThin (Aiello-Lammens et al. 2015). The algorithm retains the maximum number of observations by randomly deleting points within a user-defined buffer. We used a buffer of 500 m to delete double observations at the same stream from different years. Additionally, Bsal occurrences were considered separated on a similar scale, but postnatal salamander dispersal can still be expected over such distances (Klewen 1985; Schulte et al. 2007; Hendrix et al. 2017). This resulted in a spatially balanced set of 154 points.

Candidate predictors—breeding habitat

Choosing biologically meaningful predictors is vital in HSMs (e.g., Phillips et al. 2006; Guisan et al. 2017; Fourcade et al. 2018). As the breeding model extent was already restricted based on land cover criteria, we exclusively included topographic and geological predictors describing microclimate, stream size and structure. Fire salamanders prefer to deposit their larvae in small creeks close to the spring, which can be measured using stream order, depth or width (Thiesmeier 2004; Manenti et al. 2009; Bani et al. 2015). We used water flow accumulation instead, which additionally accounts for the risk of downstream drift after heavy rainfall (Thiesmeier 2004; Bani et al. 2015). We used bedrock type as an approximation of the structural characteristics of the stream. Due to the unavailability of fine scaled empirical climate data, we used the topographic positioning index (TPI) and elevation as proxy for the local microclimate, which is a limiting factor in urodele core habitats (Wagner et al. 2020b; Goudarzi et al. 2021). Elevation, water flow accumulation and TPI were derived from a digital elevation model (status: 2019, 1 m resolution, see Geobasis NRW 2019b for details). We categorised bedrock using geological maps “IS GK 100 DS” (status: 2016, available as vector data at a scale of 1:100,000, see Geobasis NRW 2019c for details) based on its structural characteristics: loose geological layers, such as sand, gravel and silt, and solid bedrocks. Environmental data were processed to raster format with a cell size of 5 × 5 m using ArcMapv.10.7 (ESRI 2019).

Candidate predictors—whole life cycle habitat

In the whole life cycle model, we included the mean of the modelled breeding habitat suitability in a radius of 500 m as a measure of potentially available breeding habitats for adult salamanders. As topography was already included in the breeding model, whole life cycle models were built using land cover predictors only. We categorised land cover based on the “LBM-DE2018” landscape model (available as vector data, processed to 5 × 5 m raster, for details see BKG 2019) into grassland, settlement, lentic waterbodies (predominantly storage reservoirs) and coniferous, mixed and deciduous forest. To reduce model complexity, we did not include cropland because it barely occurs within potential salamander corridors in the chosen study area. Roads were available as linear elements (vector data, Geobasis NRW 2019a). Therefore, we added a buffer of 5 m around them, categorised them as settlements and processed them to a 5 × 5 m raster. Borders of land cover types were smoothed by calculating the mean proportional cover in a radius of 500 m to account for the spatial ecology of adult salamanders that deposited the recorded larvae (Schulte et al. 2007; Manenti et al. 2009; Bani et al. 2015). However, as potential barriers to fire salamander migration and dispersal are relevant on smaller scales, we calculated settlement and waterbody cover in a radius of 30 m.

Modelling process and predictor selection

To account for multicollinearity, predictors with correlation coefficients r ≥ 0.8 and VIF ≥ 5 need to be excluded (Pradhan 2016, see SI2). Therefore, we removed the proportion of sandy and loose bedrock in the breeding model. For the whole life cycle model, forest types were strongly correlated. We excluded coniferous forest as mixed and deciduous forests are the preferred terrestrial habitats of fire salamanders in central Europe (Thiesmeyer 2004). For model evaluation, MaxEnt v. 3.4.0. was called from the R environment using the ENMeval package v. 1.0 (Muscarella et al. 2014). Occurrence data were separated into a training and a test dataset for cross-validation using randomised 10-folded splitting. For background sampling, the same random points were used for the training and testing data following Merow et al. (2013). In both modelling steps, we used and compared various transformations of the predictors, so-called feature types, including linear[L], quadratic[Q], hinge[H] and product[P] features. Moreover, regulation multiplier (RM) values, which are used to control model overfitting, gradually increased (see SI3 for all candidate models). Candidate models were ranked by the Akaike Information Criterion corrected for small sample sizes (AICc) (Warren and Seifert 2011; Katz and Zellmer 2018).

Because breeding habitat suitability was included as a predictor in the whole life cycle models, only one of two candidate breeding habitat models with ΔAICc ≤ 2 was chosen based on the biological plausibility of the underlying response curves (SI4). In contrast, we selected two different scenarios of habitat suitability throughout the life cycle to be used in the subsequent resistance modelling. These models should represent the best fitting (low AICc) but also biologically distinct scenarios. The latter was based on niche overlap in geographic space (Schoener’s D), which measures the deviation between each raster cell of the habitat suitability predictions. Raw MaxEnt outputs of the two selected whole life cycle models were projected on the whole study area and used without further transformations or definition of absolute barriers as conductance surfaces in resistance models.

Resistance modelling

Habitat connectivity was assessed by calculating the pairwise resistances between 167 fire salamander occurrences in Circuitscape v. 4.0 (McRae et al. 2013). Compared to other approaches, such as the least cost path algorithm, circuit theory reflects diffuse dispersal, similar to random walk theory in ecology (McRae and Beier 2007). Due to the high uncertainty about possible pathways, such an approach seems more suitable to model disease transmission by hosts (Nobert et al. 2016). To reduce computing time, suitability maps were upscaled to 20 × 20 m resolution by averaging. This resulted in a resistance surface with 5,519,281 grid cells in total. We added a buffer of random resistance spanning over 20% of the maximum width of the study area to account for edge effects (Koen et al. 2010, 2014). We used random resistances as the buffer exceeded the political border of Germany, beyond which environmental data was not available in the same format and resolution. Salamander occurrences defined as nodes included the centroids of Bsal screening areas (tested nodes = 39) and those larval localities used to build the HSMs (nodes not tested for Bsal, n = 128) that had no screening centroid within a radius of 500 m.

Habitat connectivity and Bsal presence

As our independent variables (pairwise resistances) referred to pairs of sites while our dependent variables (Bsal presence, larvae abundance) referred to single sites, we derived three node-based measures of connectivity for each site (cf. Koen et al. 2016) to predict the presence of Bsal: (1) the Euclidean distance and (2) the resistance to a single neighbouring (geographically closest) Bsal outbreak site (DBsal and RBsal), (3) the summed up resistance to (a) all (n = 166) and the neighbouring (b) 32, (c) 16 and (d) 8 tested and untested salamander occurrences (RSum, Fig. 1). We included untested localities in the latter variable to measure the sites general connectivity within the salamander metapopulation, as the resource intensive Bsal screening was only conducted in 39 out of 167 available salamander localities (see Fig. 2). Additionally, we tested for potential anthropogenic vectors by including the predicted number of recreational visitors per year and ha (Schägner et al. 2016) and the local density of trails in a radius of 500 m (provided by Beukema et al. 2021).

Fig. 1
figure 1

Node-based parameters of salamander habitat connectivity derived from pairwise resistances. For RBsal, the resistance to the geographically closest known Bsal outbreak site was extracted. For RSum, the pairwise resistances to all (n = 166) and to the geographically closest 8, 16 and 32 neighbours were summed up as a measure of (local) connectivity within the salamander metapopulation. All resistances were modelled twice based on two different habitat suitability models used as conductance surface (L1 and LQH3, see Fig. 2)

Fig. 2
figure 2

Current density based on the LQH3 (left) and L1 (right) whole life cycle habitat suitability models used as conductance surfaces in circuit theory-based connectivity analyses. Circles (39) indicate Bsal screening area centroids (see Lötters et al. 2020a). Triangles (128) indicate larvae localities without Bsal screening centroids within a radius of 500 m. Squares include larvae removal sampling sites (30). Note that 15 out of 30 larvae removal sampling sites were screened for Bsal as well

To assess the parameter importance for Bsal presence, we first built Generalized Linear Models (GLMs) using the glmulti package (Calcagno and de Mazancourt 2010). Models were fitted using maximum likelihood estimation since our primary goal was to derive importance across multiple parameter combinations. As the variables RSum and RBsal (Fig. 1) were highly correlated (see SI5), each node-based resistance parameter was combined with trail and visitor density in a separate call (n = 11). In the next step, the averaged effect across all candidate models per call (n = 64) was calculated using the coef.glmulti function (Burnham and Anderson 2002; Johnson and Omland 2004). We did not include random effects to account for potential non-independence of neighbouring sites since no spatial autocorrelation in the models’ residuals was detected using Moran’s I test (see SI5). As we could not define Bsal absence for most of the known salamander localities, the above-described modelling steps were then applied to a second subset of Bsal-tested sites only. However, no interactions could be modelled for this subset due to the reduced sampling size of 39 sites (potential overparameterization). All predictors were normalized with an orderNorm transformation using the bestNormalize package (Peterson and Peterson 2020) prior to the analysis.

Habitat connectivity and salamander abundance

The rounded mean number of larvae per monitoring site (n = 30) was used as a measure of local salamander abundance. Because of overdispersion in the count data, we built univariate negative binomial regression models (nbGLMs) to assess the effect of RBsal and RSum on local salamander abundance using the MASS package. RBsal was set to “0” for monitoring sites that were also screened for Bsal (15 out of 30) and found to be positive (n = 4). Models were ranked by AICc to identify the best predictor. Again, no spatial autocorrelation was detected in the model residuals.


Salamander and Bsal presence

Fire salamanders were still widely distributed in forest areas of the northern Eifel (Fig. 2), as we found larvae in 169 of our 246 larvae presence-only surveys in 2018 and 2019. No or few larvae were found in elevated areas dominated by coniferous forest, such as the upper “Weiße Wehe” valley (northern central region, Fig. 2) and the “Erkensruhr” valley (southwestern region). Additionally, no larvae were found in most streams sampled in the “Sauerbach” valley (southern central region). Here, Bsal was detected on newts in a pond system at the stream source as well as on adult salamanders downstream. In total, 17 Bsal-positive sites that met the criteria of recent fire salamander records within a 500 m radius were used as nodes in the following resistance models.

Breeding habitat suitability

The two top-ranked breeding habitat models with ΔAICc ≤ 2 (i.e., L0.5 [linear features, regulation multiplier = 0.5] and P0.5 [product features]) barely differed in their predictive ability or level of overparameterization according to AUC (0.794 and 0.799), average difference in AUC of models built with test data (0.021, Var: 0.01 and 0.024, Var: 0.01, respectively) or omission rate (both 0.007, Var: <0.001; see SI3 for all candidate models). Therefore, we relied on the response curves of the models (SI4) for final model selection. We selected the L0.5 (linear feature only, regulation multiplier  = 0.5) breeding habitat model since no response to the proportion of solid bedrock was visible in the P0.5 model. In the selected L0.5 model, breeding habitat suitability decreased with increasing flow accumulation, elevation and TPI, while an increasing amount of solid bedrock led to increased predicted suitability values. TPI was by far the most important variable in terms of percent contribution (49.1%) and permutation importance (45.3%). Nevertheless, solid bedrock, water flow accumulation and elevation contributed considerably to the predicted suitability (Table 1). In particular, permuting water flow accumulation values worsened the predictive performance, indicating the high importance of this environmental parameter for salamander breeding habitats.

Table 1 Percent contribution to model variance and permutation importance of variables used in the selected breeding habitat suitability model for fire salamander

Whole life cycle habitat suitability

For the whole salamander life cycle, the top-ranked model (LQH3: linear, quadratic and hinge features, regulation multiplier =  3) was more complex than the selected breeding model (see SI3 for all candidate models). The predictive ability was good (AUC  = 0.846, average or MTP  = 0.013 (Var: 0.001)), and the AUC values of models built with test localities slightly dropped to 0.832 (Var: 0.004) on average, indicating a moderate level of overparameterization. The mean breeding habitat suitability and deciduous and mixed forest cover were by far the most important variables, as they contributed to more than 80% of the model variance (Table 2). As a second, biologically distinct (ΔSchoeners D =  0.735, see SI3) scenario, we selected a model using an RM-value of 1 and linear features only (AUC =  0.808, average or MTP =  0.006 (Var: <0.001), average ΔAUCtest = 0.017 (Var. 0.007)).

Table 2 Percent contribution to model variance and permutation importance of predictors used in the whole life cycle model

Compared to the top-ranked model, grassland cover replaced breeding habitat suitability as the most influential predictor. Waterbody and settlement cover, potentially interrupting salamander functional connectivity, were of minor importance in both models. The underlying response curves (SI4) of the two selected whole life cycle models were similar except for deciduous and mixed forest cover. In the LQH3 model, the underlying curve was hump-shaped, and the predicted suitability value peaked at approximately 50% cover of deciduous and mixed forest in a radius of 500 m. Regardless, areas completely covered by this forest type were more suitable than areas with different land cover. In the L1 model, the predicted suitability values monotonically increased with increasing forest cover. In both selected whole life cycle models, increasing grassland, settlement and waterbody cover led to a reduction, while mean breeding habitat suitability led to an increase in predicted habitat suitability values.

Habitat connectivity and Bsal occurrence

The resistance to the neighbouring Bsal outbreak site (RBsal) was the most important predictor for Bsal presence and sites that had higher RBsal were less likely to be Bsal positive in both datasets, including all (Table 3) and tested sites only (Table 4). Likewise, the general connectivity within the salamander metapopulation measured by RSum was higher in Bsal-positive sites. Euclidean distance to the neighbouring Bsal outbreak site (DBsal) as well as visitor and trail density were of little importance compared to RBsal and RSum, irrespective of the scaling (number of neighbours) or resistance surfaces (L or LQH) used (see SI6). However, the interaction of trail density and the summed-up resistance to the closest 16 neighbouring salamander occurrences (RSum16 LQH) was of considerable importance (ranked 4th ). Interactions between visitor density and node-based connectivity measures were less important [DBsal (km): Visitors was ranked 14th ]. Overall parameter importance was moderate, especially when Bsal-tested sites were used exclusively (maximum importance = 0.59, see Table 4) and no interactions could be modelled due to the reduced sampling size.

Table 3 Model-averaged parameter importance for the presence of Bsal in 167 tested and untested fire salamander occurrences: Bsal positive (n = 17) and untested/Bsal negative sites (n = 150)
Table 4 Model-averaged parameter importance for the presence of Bsal in 39 tested fire salamander occurrences: Bsal positive (n = 17) and Bsal negative (n = 22)

Habitat connectivity and salamander abundance

The mean number of larvae strongly differed between the monitoring sites (0.3 to 138.3, see SI1) and was best explained by the resistance to the neighbouring 32 salamander occurrences (RSum32 LQH, Table 5). Larvae abundance increased with local habitat connectivity, as fewer larvae were found in sites with high summed-up resistances (Fig. 3). Likewise, the resistance to the neighbouring Bsal site was lower in sites with many larvae, even though models using RBsal as predictor showed worse predictive ability according to AICc and the models’ standard errors.

Table 5 Univariate negative binomial GLMs used to predict the rounded mean number of fire salamander larvae abundance using differently scaled measures of salamander habitat connectivity (RSum) and connectivity to neighbouring Bsal outbreak sites (RBsal)
Fig. 3
figure 3

Projected effect of the resistance to the geographically closest know Bsal outbreak site (RBsal, left) and the sum of resistances to 32 neighbouring salamander occurrences (RSum, right) on the rounded mean salamander larvae abundance between 2015 and 2019 in 30 monitoring sites using univariate nbGLMs.


Emerging infectious diseases can spread through animal populations in multiple ways, but direct transmission by host migration and dispersal is often assumed to be the most relevant way of infecting individuals (White et al. 2000; Altizer et al. 2011; Schmidt et al. 2017). Among all European tailed amphibians, the European fire salamander (Salamandra salamandra) has been the species impacted most by the spread of Bsal. Adult salamanders are highly susceptible, and infections are lethal in most cases. Population crashes through Bsal infection have repeatedly been documented, for instance, in the Netherlands (Spitzen-van der Sluijs et al. 2013) or the Ruhr district in Germany (Schulz et al. 2020). Most studies dealing with Bsal infection dynamics assume salamander individual-based disease transmission (Schmidt et al. 2017; Canessa et al. 2018), and migrating and dispersing infected salamanders likely represent an important natural vector. However, there is poor evidence of the extent of their contribution to the distributional pattern of Bsal observed at the landscape level. In this study, Bsal occurrence was positively associated with fire salamander habitat connectivity, which indicates that Bsal spread via salamander movement seems to be an important factor in large, unfragmented habitats (cf. Beukema et al. 2021). However, salamander larvae, and therefore also adults, were more abundant in well-connected areas.

Habitat suitability models

Due to Bsal-induced salamander absences, the equilibrium state assumption of the HSMs used as conductance surfaces might be violated (cf. Sandvoß et al. 2020). However, salamander larvae were found in most surveyed habitats in the northern Eifel. Breeding habitat models were built along digitised streams in forest areas only, since breeding in small, stagnant water bodies is of minor importance for Eifel populations compared to lowland populations, and it has thus far not been reported for Eifel populations (Dalbeck and Thiesmeier 2011; Hendrix et al. 2017). Puddles and pools close to streams might be used occasionally, but larvae were almost exclusively found in flowing waters. Small-scale data on the structural and biotic characteristics of streams, such as watercourse heterogeneity, stream depth and width, number of pools, substrate, macrozoobenthos communities or periphyton, were not available for the whole study area, which could improve breeding habitat model precision (Manenti et al. 2009; Wagner et al. 2020b). Instead, water flow accumulation, solid bedrock, elevation and topographic positioning were found to be suitable alternatives to predict the streams’ topographic and structural suitability.

In their terrestrial habitat, fire salamanders are mostly found in deciduous and mixed forests, while plantations of coniferous trees, such as spruce (Picea abies), are largely avoided (Joly 1968; Thiesmeier 2004). Overall, this expectation was fulfilled, but in the top-ranked whole life cycle model (LQH3), areas dominated by deciduous and mixed forest were predicted to be (moderately) less suitable than areas with intermediate deciduous and mixed forest cover. This might be due to the typical position of streams at the border of two different land cover types, and consequently, many larvae were located at forest edges. To account for this artefact, a second, biologically distinct whole life cycle model (L1) was used, predicting a monotonous relation to deciduous and mixed forest cover, which seems biologically more accurate (Joly 1968; Thiesmeier 2004). Nevertheless, validation with spatially independent test data is needed (Veloz 2009) to project these models to landscapes where salamanders inhabit other habitat types, such as railway embankments (Thiesmeier 2004).

Habitat connectivity and Bsal occurrence

In our connectivity models, we assumed salamanders to migrate preferentially along areas of high projected habitat suitability, as suggested by Bani et al. (2015) and Goudarzi et al. (2021). However, microclimatic factors influencing the rate of water loss can be relevant for salamander migration within suitable habitats (Peterman et al. 2014; Nowakowski et al. 2015). Salamanders might use suboptimal habitats for compensatory movement (Peterman et al. 2014) and, compared to occurrence, effective dispersal can be driven by divergent landscape characteristics (Cushman et al. 2013). Notwithstanding, we assumed that our nodes-based resistances derived from HSMs are best suited to explain the spread of Bsal, as these measure the continuous availability of salamanders for Bsal (cf. Fig. 3). This assumption is also supported by the minor importance of the geographical distance to neighbouring Bsal sites (DBsal) in our models, highlighting that resistance to salamander movement rather than pure distance to Bsal occurrence drives the distribution of Bsal. Moreover, unfragmented habitats of high quality supported higher salamander abundances, which should accelerate the spread of Bsal as well. In contrast to other studies (e.g. Beukema et al. 2021), we did not determine Bsal absence. The detectability of chytrids depends on seasonality, the local demographic structure of hosts, and human activities (Adams et al. 2010; Mosher et al. 2018; Bozzuto and Canessa 2019; Beukema et al. 2021), but most importantly, it depends on the number of animals tested (DiGiacomo and Koepsell 1986). The latter was biased in Bsal-positive sites in our study region (see Suppl. 3 in Lötters et al. 2020a) and by omitting untested sites (128 out of 167), large parts of the known salamander distribution would be ignored. Therefore, we modelled both tested and untested sites, but still found Bsal presence to be best explained by the resistance to neighbouring Bsal outbreak sites (RBsal), followed by the sites’ connectivity within the local metapopulation (RSum) in both datasets.

Human activities can obscure natural patterns of pathogen spread (Greer and Collins 2008; Adams et al. 2010; Padgett-Flohr and Hopkins 2010; Cohen et al. 2016). For instance, fragmentation and road density may facilitate Bsal spread, e.g., via touristic activities in the Eifel (Allan et al. 2003; Jousimo et al. 2014; Paquette et al. 2014; Beukema et al. 2021). Despite that, trail and visitor density were poor predictors for Bsal presence in our study, indicating that anthropogenic vectors play a minor role for the spread of Bsal at least on local scales. Our findings thereby emphasize the potential of chytrids to afflict amphibians especially in pristine, unfragmented habitats (Lötters et al. 2009; Becker and Zamudio 2011).

As Bsal occurrence is not restricted to the range of fire salamanders in Central Europe, many other pathways for Bsal dispersal must exist (Stegen et al. 2017; Canessa et al. 2018; Dalbeck et al. 2018; Beukema et al. 2021). Nearly all native amphibian species can come into contact with fire salamanders (Dalbeck et al. 2007; Dalbeck and Thiesmeier 2011), and in the northern Eifel, Bsal was found on the skin of smooth, palmate, alpine and crested newts (Lissotriton vulgaris, L. helveticus, Ichthyosaura alpestris and Triturus cristatus) in proximity and distant from fire salamander habitats (Lötters et al. 2020a). Infected newts suffer less from infection and may migrate longer distances compared to fire salamanders (Kovar et al. 2009), which could make them important Bsal vectors as well (Canessa et al. 2018; Spitzen-van der Sluijs et al. 2018; Beninde et al. 2021). As the modelled salamander habitat suitability partly entails suitable newt habitats, salamander connectivity likely overlaps with newt connectivity. Moreover, anurans, waterfowl or crayfish can be reservoirs and vectors for chytrid spores (McMahon et al. 2013; Stegen et al. 2017). Environmental transmission, e.g. by spores floating in streams, may have contributed to the pattern found in this study (Stegen et al. 2017; Mosher et al. 2018), but former studies indicate that biotic factors primarily drive the distribution of pathogens (Cohen et al. 2016; Spitzen-van der Sluijs et al. 2018).

We acknowledge that due to our multistep modelling framework, prediction biases may be multiplied in the final GLM’s (compound error), especially since we combined results of survey activities that yield differently resolved data on Bsal and salamander larvae occurrence. However, larvae localities included in our habitat and connectivity models represented one of several possible sets of points, as larvae could typically be found over a few hundred meters of the streams. This also counteracts potential bias due to the upscaling of resistance data, as bigger pixels still included streams with larval presence. Additionally, as we modelled occurrence, our models should be less sensitive to uncertainties and biases in the predictors of the first modelling step compared to quantitative analysis like modelling Bsal prevalence. Even though we selected the two most distinct (Schoender’D) habitat suitability models (SI3) as resistances surfaces, predictors showed consistent qualitative responses (SI4). Therefore, major compound errors leading to misidentified key parameters for Bsal occurrence seem unlikely. Still, the final estimates of regression parameters likely have higher standard errors than indicated in our tables due to our stepwise framework.

Habitat connectivity and salamander abundance

Despite Bsal presence being associated with well-connected salamander habitats, these habitats generally supported higher salamander abundance. Large populations that typically occupy extensive, unfragmented habitat patches are more resilient and less likely to become extinct because local extinctions can be buffered by immigration (Hanski and Simberloff 1997; Marsh and Trenham 2001; Jousimo et al. 2014; Heard et al. 2015). Nevertheless, drastic declines in salamander abundance have been observed at sites in the study area and elsewhere (Spitzen-van der Sluijs et al. 2013; Dalbeck et al. 2018), and slightly fewer larvae were found at Bsal outbreak sites in the Eifel (Wagner et al. 2020a). Fire salamander populations may be able to persist at low densities despite Bsal presence in large, unfragmented forest areas where recolonisation is facilitated (cf. Schmidt et al. 2017). Additionally, Bsal primarily affects reproductive adults (Stegen et al. 2017), and fire salamanders in Germany mostly start to reproduce in their 5th year at the earliest (Thiesmeier 2004). Therefore, facilitated population recovery due to increased connectivity can be time-lagged.

Implication for mitigation efforts

As eradication attempts of Bsal in the northern Eifel or comparable regions are logistically unfeasible and likely to fail (Bosch et al. 2015), in situ conservation efforts need to focus on population and habitat management that support the resilience of afflicted salamanders and newts (Heard et al. 2018; Scheele et al. 2019). In situ conservation strategies, such as fencing, include measures that counteract salamander connectivity (Thomas et al. 2019), so it is highly important to gauge their demographic consequences. Here, it has been demonstrated that isolation attempts seeking to interrupt Bsal transmission mediated by amphibians could potentially counteract demographic advantages of connectivity (McCallum and Dobson 2002; Heard et al. 2015). However, long-term demographic consequences are difficult to predict, and overall advantages of salamander habitat connectivity can be expected only if Bsal-induced mortality appears in a patchy pattern that is outweighed by increased recruitment and recolonisation from neighbouring areas (cf. Heard et al. 2015; Heard et al. 2018).

Climatic or other environmental gradients, such as water salinity or temperatures that exceeds the pathogen’s physiological niche or optimum, can contribute to natural refugia from chytrids that harbour source populations (Puschendorf et al. 2009; Doddington et al. 2013; Heard et al. 2015; Urban et al. 2015; Nowakowski et al. 2016; Hettyey et al. 2019), but no such refugia habitats from Bsal have been detected thus far (Beukema et al. 2020). Future research must identify mismatches between salamander and Bsal niches that can be exploited by feasible in situ conservation measures to amplify the resilience of salamander metapopulations in regions with extended and well-connected habitats, such as the northern Eifel (Campbell Grant et al. 2018; Scheele et al. 2019). This requires a comprehensive and spatiotemporally explicit research approach that assesses local salamander population dynamics and Bsal occupancy depending on reservoir host abundance, landscape connectivity, physiochemical parameters and microbiota communities (Campbell Grant et al. 2018; Bozzuto and Canessa 2019).