Toward community predictions: Multi‐scale modelling of mountain breeding birds' habitat suitability, landscape preferences, and environmental drivers

Abstract Across a large mountain area of the western Swiss Alps, we used occurrence data (presence‐only points) of bird species to find suitable modelling solutions and build reliable distribution maps to deal with biodiversity and conservation necessities of bird species at finer scales. We have performed a multi‐scale method of modelling, which uses distance, climatic, and focal variables at different scales (neighboring window sizes), to estimate the efficient scale of each environmental predictor and enhance our knowledge on how birds interact with their complex environment. To identify the best radius for each focal variable and the most efficient impact scale of each predictor, we have fitted univariate models per species. In the last step, the final set of variables were subsequently employed to build ensemble of small models (ESMs) at a fine spatial resolution of 100 m and generate species distribution maps as tools of conservation. We could build useful habitat suitability models for the three groups of species in the national red list. Our results indicate that, in general, the most important variables were in the group of bioclimatic variables including “Bio11” (Mean Temperature of Coldest Quarter), and “Bio 4” (Temperature Seasonality), then in the focal variables including “Forest”, “Orchard”, and “Agriculture area” as potential foraging, feeding and nesting sites. Our distribution maps are useful for identifying the most threatened species and their habitat and also for improving conservation effort to locate bird hotspots. It is a powerful strategy to improve the ecological understanding of the distribution of bird species in a dynamic heterogeneous environment.

In this context, landscape compositions and arrangements of natural areas, such as mountains, and knowledge of the interaction between species and their environment play a key role in population persistence and biodiversity conservation (Austin, 2002;Guisan & Thuiller, 2005;Scridel et al., 2018;Zambrano et al., 2019).
This could be achieved with modelling tools, such as species distribution models (SDMs), that facilitate conservation decisions in spatially complex landscapes (Guisan & Thuiller, 2005;Guisan et al., 2013;Habibzadeh & Ludwig, 2019). SDMs are statistical methods for assessing species' ecological requirements and predicting geographical distribution in space and time on the basis of the correlations of environmental variables and species occurrences (Austin, 2007;Hirzel, Lay, Helfer, Randin, & Guisan, 2006;Naimi & Araújo, 2016). When using SDMs, it is critical and problematic to identify the appropriate scale within which environmental variables affect the distribution of species and most SDMs fall short to incorporate environmental and land use variables at different spatial scales (Fournier, Barbet-Massin, Rome, & Courchamp, 2017;Guisan & Thuiller, 2005;Mackey & Lindenmayer, 2001;Vicente et al., 2014).
Most SDM studies rarely support the role of the best scale (the optimal range of spatial scales affecting distribution patterns of mobile organisms) of effects of focal predictors (that sum up information on the surrounding landscape within the focal cell; see Bellamy, Scott, & Altringham, 2013;Fournier et al., 2017;Guisan & Thuiller, 2005) on highly mobile organisms like bird species in complex landscapes (Chase et al., 2018;Guisan & Thuiller, 2005). SDMs usually consider a single predictor at a single spatial scale that is restricted to the extent of the sampling and the grain size and does not sufficiently address scale dependencies (Chase et al., 2018;Fournier et al., 2017;Hamer & Hill, 2000;Vicente et al., 2014). But the specific value of a single cell is not sufficient where we need several landscape elements, such as distance to resources, land structures, and land cover variables to incorporate spatial context and neighborhood and address the complexity of the ecological niche of the mobile species (Fournier et al., 2017;Guisan & Thuiller, 2005;Jaberg & Guisan, 2001;Rainho & Palmeirim, 2011;Scherrer, Christe, & Guisan, 2019).
Compared to plants, bird species pose more modelling problems at high spatial resolutions since it demands to take into account the effect of different spatial scales on a highly mobile group such as birds (Fournier et al., 2017;Guisan & Thuiller, 2005;Scherrer et al., 2019). In addition, research into feeding and nesting requirements for many bird species at multiple spatial scales still remains inadequate, and most of the recent studies tend to use a fixed spatial scale (Fournier et al., 2017;Jaberg & Guisan, 2001;Scherrer et al., 2019). Thus, it is important to develop an efficient framework, for bird species distribution, to assess the effect of focal predictors and compile information on the neighboring landscape within a focal cell at an appropriate scale (e.g., habitat selection in a proper scale; Bellamy et al., 2013;Bellier, Certain, Planque, Monestiez, & Bretagnolle, 2010;Bucklin et al., 2015;Fournier et al., 2017;Guisan & Thuiller, 2005;Revermann, Schmid, Zbinden, Spaar, & Schröder, 2012). Such a multi-scale modelling method provides information on species distribution, identifies different requirements of bird species within foraging, nesting, and breeding range (Fournier et al., 2017;Guisan & Thuiller, 2005;Pearson, Dawson, & Liu, 2004;Vicente et al., 2014), and contributes to finer-scale identification and planning of biodiversity and protected areas (Fournier et al., 2017;Razgour, Hanmer, & Jones, 2011;Vicente et al., 2014).
One major step forward of this study is the systematic assessment of the best scale within which the landscape (mainly land cover and land use), distance, topographic, and climatic predictors of bird habitat suitability are the most influential. This multi-scale approach which involves linking focal variables with predicting power of ensembles of small models (ESMs) enables us to consider complexities in localization of birds as highly mobile species, especially in mountainous ecosystems with a great topographic ruggedness, which poses the significant modeling challenge of neighborhood analysis (Bellamy et al., 2013;Scherrer et al., 2019). These developments on the spatial modelling should improve our capacity to use models to assist bird monitoring and management in space and time, and ultimately help to identify priority bird conservation areas.
In this study, we employed the occurrence data of bird species to find efficient modelling solutions and build reliable distribution maps for biodiversity and conservation necessities at finer scales.
In order to identify the efficient scale of each environmental predictor, we have performed a multi-scale modelling method and used focal variables at different scales. We defined 12 neighboring window sizes (0.1-5 km) to explore the effect of species-specific scales of influence of each variable, and then we fitted univariate models to identify the best radius for each focal variable (land cover and land use) for each species and the most efficient scale of influence of each predictor. We then employed the final set of variables obtained from the previous stage to build ESMs at a fine spatial resolution of 100 m and generate species distribution maps as a monitoring tool of habitat suitability.

F I G U R E 1
The Mute Swan (Cygnus olor) is an introduced breeding bird to Switzerland (Sattler et al., 2015) The key issues in this research are the following: (a) an analysis of the role of the best scale (size of neighborhood) of effects of focal predictors (land cover and land use) at local (topographic and micro-climatic) scale in bird species distribution across a complex landscape, and (b) providing potential and optimized distribution maps as a monitoring tool of habitat suitability and biodiversity conservation planning. This study followed a similar approach, carried out by Scherrer et al. (2019) to bat species, but we have expanded its capacity to include a diverse group of species (not just rare and cryptic species) and take into account the consequences of climate change and major anthropogenic disruptions in the landscape on bird species distribution for biodiversity planning purposes ( Figure 1).

| Study area and species data
The study area, the western Swiss Alps of Vaud (46°10 to 46°30′N; 6°50′ to 7°10′E; Figure 2), has been a site of high importance for interdisciplinary and transdisciplinary research at the University of Lausanne since 2013 and is now transferred to the Interdisciplinary Center for Mountain Studies (CIRM; http://recha lp.unil.ch). It includes an area of ca. 700 km 2 and elevation gradient extending from Geneva's lake at 372 m to the Pointe des Diablerets at 3,210 m a.s.l. (Descombes, Vittoz, Guisan, & Pellissier, 2017;Scherrer et al., 2019) and it is affected by significant human activity: dense population and intensive farming in the Rhône Valley, tourism and leisure activities and more extensive farming in subalpine regions that makes up a mosaic of meadows, pastures, and forest and woodland patches (see http://recha lp.unil.ch; Pellissier et al., 2012;Randin, Jaccard, Vittoz, Yoccoz, & Guisan, 2009;Scherrer et al., 2019 vulnerable (VU), and 1 as endangered (EN) based on the national red list ( Figure 3). This classification could allow us to apply the research approach to a wide range of bird species from rare to poorly sampled ones. We only considered birds with a sample size greater than 20 presence records (>20 presence) as species with fewer presence records, due to errors connected with very small sample size, are not deemed appropriate for modeling (Thuiller, Lavorel, & Araújo, 2005).

| Environmental data
Environmental data were derived from different sources, used for generating gridded environmental variables at 100 m spatial reso-  Dormann et al., 2013), to include the variables that are not highly correlated (Bellamy et al., 2013;Progin, 2018;Scherrer et al., 2019).
There are more details on environmental variables in Table 1 and   Table S1.

| Univariate models
For improving our understanding of the ecological connection be- For each species, the best window radius size (i.e. scale) for each land-cover variable was selected using the performance of the GLMs, measured by AUC, that was averaged for a better presentation of all bird data across all species, and run in different habitats.
We selected the best scale (i.e. window size) for focal variables based on the highest AUC score, and all the variables with the highest AUC were selected and used later in the final model (multivariate models).
A fundamental tenet of ESMs, for each species, is to build bivariate small models with all possible combinations of two environmental variables as predictors that results in 105 models defined by (n 2 − n)/2, where n refers to the number of variables (Progin, 2018;Scherrer et al., 2019). Each of the 105 small bivariate models were weighted by Somers'D as a performing parameter D = 2 × (AUC − 0.5) that varies between −1 and +1 and gives

| Species distribution and richness maps
To create distribution maps, for three groups of species (all species, non-threatened species, and threatened species), we combined the ensemble of each group of species into a final ensemble (Figure 8; Breiner et al., 2015Breiner et al., , 2018. In order to select the optimal habitat of bird species which occupy varying home ranges in such a heterogeneous environment, we preferred to rely more on practical binary

| Univariate models-Scale selection
The average predictive power of the focal variables is shown by AUC of univariate models for each focal variable across the scales (100-5,000 m) based on the test datasets of 7 species classes.
Generally, our results showed that the primary predictors of species distribution were land use/cover variables measured at large scales with the neighboring window size ranging from 1,000 to 5,000 m for Agriculture, Swamp, Orchard, Lake, and NDVI variables they had the highest probability of presence/occurrence across large scales (1,000-5,000 m; Figure 5). This (univariate model) approach could improve the selection of the best scale for each focal variable used in the following stage in the final ESMs.

| ESMs-Model performance
The results have shown that both MARS and GAM modelling tech-  Figure 7, shows the contribution of each variable (see Table 1 for its acronyms and descriptions) to all 91 modeled species. In general, the patterns of habitat suitability highlighted by the ESMs are strongly F I G U R E 5 Performance of univariate models for each group of species in different habitats: Association (averaged area under the curve over the 50 runs of each univariate models) between the presence of each group of species and each focal variable at 12 spatial scales linked to bioclimatic variables. For example, the most important relative contribution to model (0.79 and 0.72, respectively)  Eventually, there were the variables of poor performance, including

| ESMs-Species distribution maps
In general, species richness predictions show that the areas close to coniferous forest, agricultural areas, and zone habitats (residential areas) of the highlands are the most suitable for 78 of species listed as LC, NT in the study area. Species richness maps of the 91 bird species and LC or NT in the national red list showed a pattern of moderate species richness distributed across the study area with higher habitat suitability in the eastern, central and northern parts of the study area near the highland area. However, this pattern varies, when 13 endangered species (VU or EN) are taken into consideration, which appear to occur only in lower slopes in the western part of the study area, near the deciduous forest, lake, and wetland ( Figure 8).

| D ISCUSS I ON
This study supports the role of the best scale of effects of focal predictors (land cover and land use) at local scale (topography and micro-climate) and the exclusion of poor focal variables (Bellamy et al., 2013;Progin, 2018) in bird species distribution in a complex landscape. This systematic approach to variable selection (univariate models with focal variables) and assessment of landscape's neighboring influence increased the power of the model predictions in our heterogeneous study area and our findings confirmed that bird species respond to ecological parameters at particular spatial scale (Graf et al., 2005;Rocchia, Luppi, Dondina, Orioli, & Bani, 2018). The additional important finding is that ESMs approach is highly efficient for modelling species distributions with limited number of species occurrences. ESMs, with a limited number of presence data and a large number of environmental variables, can be effectively used for modelling bird species, the generated models and distribution maps can be used as comprehensive and practical tools in biodiversity conservation decisions (Chamberlain, Pedrini, Brambilla, Rolando, & Girardello, 2016;Scherrer et al., 2019). This innovative study improved our knowledge and understanding of bird species niche variation, ecological priorities and distribution in the western Swiss Alps.
It also supports practitioners to interpret the findings for identifying the environmental niche of the bird species  and taking practical management decisions.

F I G U R E 6
Evaluation of the ensemble of small models (ESMs). Boxplot of the Continuous Boyce Index, Somers'D, true skills statistic (TSS), area under the curve (AUC) and kappa, calculated for each run of the ESMs

| Univariate models
The occurrence of species, predicted by focal variables, significantly changed across all species and the relative importance of focal variables frequently changed with the radius and, in general, most were associated with focal variables measured at large scales between 1,000 and 5,000 m. Focal variables in both univariate and multivariate models (final models) of the majority of species, Hedge and It could be assumed that these 7 groups of species benefit from two variables (forest and hedge) as foraging sources of insect prey that coincide with nesting and feeding in small distances within the home range (Monsarrat et al., 2013). We may infer that, for two variables (Hedge and Forest), habitat suitability of the seven groups of species declines over larger scales (Roth, Amrhein, Peter, & Weber, 2008) and these variables are undesirable or may be disrupted by human activities at larger scales (Graf et al., 2005).
The presence/occurrence probability of majority of species in different habitats increased with increasing the radius of focal variables, "Agriculture area", "Orchard", "Lake", "NDVI", and "Swamp", indicating that the strongest interaction with these focal variables was typically observed at larger scales (1,000-5,000 m). In these radius sizes, a greater number of species ( Figure 5) benefit from these focal variables over larger scales, perhaps because these focal variables are faced with disruption such as urban and agricultural developments, transport growth, loss of connectivity (fragmentation) and loss of habitat at small scales in which bird species are threatened.
Bird species are therefore forced to occupy suitable large-scale habitats of these focal variables (Graf et al., 2005;Grüebler, Schuler, Spaar, & Naef-Daenzer, 2015). We can conclude that bird species, in this mountainous area, are disturbed by small-scale anthropogenic habitat disturbance for these five focal variables and it is not surprising that bird species in different habitats prefer these focal predictors at larger scales (Forman, Reineking, & Hersperger, 2002;Grüebler et al., 2015;Marzluff, 2001).

| Species distribution maps
Creating species richness maps by summing species habitat suitability maps (HSMs) probably overestimates species richness because it does not account for limitation factors such as competition and dispersal that can affect species distribution (Bellamy et al., 2013;Graham & Hijmans, 2006). Prediction of habitat suitability in the highland areas (north to east of the study area) was generally higher and these highland areas support higher diversity of species (all species, threatened and non-threatened species) except for VU and EN species. The richness maps of the threatened species signify a different pattern from those of the non-threatened species and the main distribution area of this group of birds is only the western part of the study area located below 1,000 m that can be regarded as the area near the water habitats such as Lake Geneva and wetlands with well-vegetated banks and parts of the deciduous forest that have been used as foraging and feeding sites. Lowland areas can be considered suitable for the VU and EN species due to the warmer climate and proximity to water habitats like lake and wetlands in the study area. Generally, climatic and focal variables "Orchard" and "Forest" play an important role in the distribution of species (Figure 7) and most species seem to respond negatively to global warming as they are forced to shift to the highland area (Maggini et al., 2011(Maggini et al., , 2014. In some parts of lowland areas and the western quadrant of the study area, prediction maps of non-threatened species (mostly passerine bird species in agricultural habitat) were also higher, which can be interpreted as the importance of arable land and wide-leaved forest for feeding and nesting of this group of species. These areas are richer in resources of foraging and nesting. The findings of these study thus emphasize the importance of forests for the survival of bird species contributing to biodiversity conservation. The findings suggest that, despite the high anthropogenic effects in the field, most of the species have a strong relationship with the forest area.
Conservation efforts should therefore concentrate on forest area management and conservation in order to optimize habitat for these species.
Our research allows us to claim that forest regions are the most potential areas for feeding and nesting of these species (Brüngger & Estoppey, 2008;Carpenè et al., 2006). According to our analysis, we are also able to conclude that lower temperatures play a major role in the distribution of most species in this mountainous area, and that these species react negatively to higher temperatures as some are predicted to shift and some already shifted to higher altitudes according to some case studies in Switzerland (see Maggini et al., 2011Maggini et al., , 2014, which indicates that bird species are likely affected in breeding season by climate change due to increasing temperatures in the breeding season. As the available habitat suitability is reduced, the distributional movements can thus be a common response to this situation (Revermann et al., 2012). Our findings shows that climate in the western Swiss Alps, particularly temperature, is the main driving factor for birds and during breeding seasons, minor changes in temperature can push species up. Our results have proved that the bioclimatic variables have the most important effects on species distribution and the important part has been played by "Bio11" and "Bio 4" in shaping the distribution of these species.

| MAIN CON CLUS I ON S
One first shortcoming in this study was the limitations of selecting maximum scales in univariate models due to the heterogeneity and size of the study area and similar responses of some species for many variables. Some species, for example, have different ecological behavior and strategies for feeding, nesting and long-distance flight and probably prefer scales of more than 5,000 m. We eventually selected 5,000 m as a maximum scale for all focal variables to address the problem of the size of the study area. The multi-scale approach that incorporates focal variables with the strength of ESMs was efficient in identifying the complexities of bird localization as species of high mobility in a heterogeneous ecosystem and proved to outperform single scale modelling. This study reveals the significance of lower and seasonal temperatures at higher elevations, as well as forest and farmland edge, for the bird species in the study area. The availability of data for certain species, due to difficulty of detection, rare and threatened status, was also limited in this research, but prediction advantages of ESMs enabled the models to consider poorly sampled bird species and model the distribution of species with limited availability of occurrence data. ESM is a powerful strategy to understand the ecology of complex and heterogeneous environments, and nature reserve managers may make better use of the models to estimate the effects of climate change and land use, such as agriculture utilization and urbanization, on target species.
Distribution maps of the area under study will help to identify the locations of the most threatened and endangered bird species biodiversity and improve conservation efforts toward the identification of bird hotspot. In this study, we only tested the framework with bird species, but this strategy could be applied to other taxa and here we encourage future studies to also examine this modelling approach particularly in case of limited presence data and rare species.

ACK N OWLED G M ENTS
We would like to thank Prof. Antoine Guisane, Dr. Daniel Scherrer, and the Spatial Ecology group (ECOSPAT) at the University of Lausanne who initiated this research, provided environmental data layers, and support the main author at the beginning of this study.
We are also thankful to Dr. Heidi Mod, Dr. Olivier Broennimann and Dr. Rui Fernandes for their assistance with scripts; and Dr. Thomas Sattler from the Swiss Ornithological Institute for providing species data and collaboration. NAT is a doctoral student at the University of Lausanne and was supported by a scholarship from Iran's Ministry of Science, Research and Technology and financial help from the University of Lausanne.

CO N FLI C T O F I NTE R E S T
None declared.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the Swiss Ornithological Institute but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available.