Prediction of biodiversity hotspots in the Anthropocene: The case of veteran oaks

Abstract Over the past centuries, humans have transformed large parts of the biosphere, and there is a growing need to understand and predict the distribution of biodiversity hotspots influenced by the presence of humans. Our basic hypothesis is that human influence in the Anthropocene is ubiquitous, and we predict that biodiversity hot spot modeling can be improved by addressing three challenges raised by the increasing ecological influence of humans: (i) anthropogenically modified responses to individual ecological factors, (ii) fundamentally different processes and predictors in landscape types shaped by different land use histories and (iii) a multitude and complexity of natural and anthropogenic processes that may require many predictors and even multiple models in different landscape types. We modeled the occurrence of veteran oaks in Norway, and found, in accordance with our basic hypothesis and predictions, that humans influence the distribution of veteran oaks throughout its range, but in different ways in forests and open landscapes. In forests, geographical and topographic variables related to the oak niche are still important, but the occurrence of veteran oaks is shifted toward steeper slopes, where logging is difficult. In open landscapes, land cover variables are more important, and veteran oaks are more common toward the north than expected from the fundamental oak niche. In both landscape types, multiple predictor variables representing ecological and human‐influenced processes were needed to build a good model, and several models performed almost equally well. Models accounting for the different anthropogenic influences on landscape structure and processes consistently performed better than models based exclusively on natural biogeographical and ecological predictors. Thus, our results for veteran oaks clearly illustrate the challenges to distribution modeling raised by the ubiquitous influence of humans, even in a moderately populated region, but also show that predictions can be improved by explicitly addressing these anthropogenic complexities.


| INTRODUCTION
Global change implies an urgent need to better understand and assess the effects of human land management on biodiversity-rich ecosystems and habitats (Erb et al., 2017;Souza, Teixeira, & Ostermann, 2015;Titeux et al., 2016). Concentrations of biodiversity can be found in many parts of the World (Gaston & David, 1994;Medail & Quezel, 1997;Myers, Mittermeier, Mittermeier, Da Fonseca, & Kent, 2000;Sverdrup-Thygeson, Brandrud, & Ødegaard, 2007). Areas with a large number of species, especially rare, threatened or endemic species often occur in remote and relatively pristine natural areas, like tropical rain forest interior (Mittermeier, Myers, Thomsen, Da Fonseca, & Olivieri, 1998;Myers, 1988) and boreal old-growth forests (e.g., Gjerde, Saetersdal, Rolstad, Blom, & Storaunet, 2004;Sverdrup-Thygeson, Søgaard, Rusch, & Barton, 2014;Timonen, Gustafsson, Kotiaho, & Mönkkönen, 2011). However, areas with high biodiversity are not only confined to remote wilderness. In fact, there is often a high coincidence between people and biodiversity (Araújo, 2003). With the increasing presence and activities of humans, an increasing number of species-rich habitats are found in ecosystems and biomes strongly influenced and transformed by us ("anthromes"; Ellis, Klein Goldewijk, Siebert, Lightman, & Ramankutty, 2010;Ellis & Ramankutty, 2008;Hobbs, Higgs, & Harris, 2009). From a conservation perspective, the most important biodiversity concentrations to keep an eye on are those under pressure, which has led some to include human-induced threats in the definition of biodiversity hotspots (Myers et al., 2000). However, all human actions do not imply biodiversity loss. In Europe, some of the most species-rich habitats outside the Mediterranean basin are seminatural grasslands, partly created and tended by humans today (Cremene et al., 2005).
Similarly, veteran trees in Europe are often legacies from a preindustrial, extensively managed agricultural landscape, and owe some of their qualities to previous human management, like coppicing. Due to their rich microhabitat structures (thick bark, crevices, dead branches, hollows etc.), these veteran trees are not only important landscape elements, but often constitute local hotspots for biodiversity in themselves (Sverdrup-Thygeson, 2009).
Yet finding these hotspots is critical for planning and management at the local level, where most decisions are made and land management is in action every day. Given that complete mapping is way beyond the resources allocated to biodiversity mapping and monitoring in most countries, some form of spatial distribution modeling (e.g., Elith & Leathwick, 2009) is needed. This is especially true in habitat types with a large suite of associated specialized species, where focusing on occurrence and critical properties of the habitat can be a cost-efficient way of locating and protecting several species in one operation (Gjerde, Saetersdal, & Blom, 2007;Lehmann, Overton, & Austin, 2002;Skarpaas, Diserud, Sverdrup-Thygeson, & Ødegaard, 2011). However, the mix of ecological and anthropogenic factors affecting local biodiversity hotspots poses several challenges to ecological and geographical prediction. There are many potentially important predictor variables, distributed across complex landscapes with gradients and thresholds in both ecological and anthropogenic influences over time (Erb et al., 2017). Now, at the beginning of the Anthropocene, there is a rapidly growing need to address these complexities.
In this study, we focus on large and hollow oaks (Quercus spp.; Figure 1)-an important biodiversity hot spot habitat in northern Europe-to address the challenge of interacting ecological and anthropogenic processes in generating spatial patterns in biodiversity hotspots. Our goal is to develop robust process-based predictions of hot spot oak occurrence, for use in conservation management and research. We concentrate on oaks in Norway, where large and hollow oaks (hereafter "veteran oaks") were recently listed as a priority habitat under the Nature Diversity Act, and where comprehensive data sets are being collected as parts of national monitoring efforts and biodiversity studies (Sverdrup-Thygeson, Evju, & Skarpaas, 2013).
Oaks are long-lasting habitats, with some trees thought to be close to 1,000 years old (Drobyshev & Niklasson, 2010). As the years go by, the architectural diversity increases, and the oaks develop patches of decay, broken branches or flaking, deeply creviced bark-and after about 200 years of age, sometimes earlier, internal cavities start to develop (Ranius, Niklasson, & Berg, 2009). In these microhabitats,  exceptionally species-rich communities associated with wood decay and wood mold flourish. For these reasons, veteran oaks are a priority conservation habitat, and for both conservation research and management of these large old trees it is important to know where they are likely to be found and how they are influenced by ecological and anthropogenic processes (Lindenmayer & Laurance, 2016;Lindenmayer et al., 2014).
A critical question for the prediction of veteran oaks is how the distribution of these large and hollow oaks-the biodiversity hot spot oaks-differs from the distribution of oak in general. We know much about the distribution and ecology of oaks from previous studies (e.g., Annighöfer, Beckschäfer, Vor, & Ammer, 2015;Dahl, 1998;Jones, 1959;Stokland & Halvorsen, 2011). To what extent do veteran oaks follow the oak niche? Oaks grow old, large, and hollow when they have sufficient time to grow, age and decay without major disturbances (Ranius et al., 2009). Some 1,000 years ago, in a slightly warmer climate and before the impact of humans, oak forests covered large areas in southern Norway. In the 16th and 17th century, there was a high demand for oak timber for buildings and ships, and large amounts of Norwegian oak were exported to Europe (Moore, 2010;Vevstad, 1998 Dahl, 1998). Elevation and local topography may therefore also be important predictors of oak presence, as temperature generally declines with elevation and south-facing slopes may have a considerably better local climate than north-facing slopes at the same latitude and elevation (Stokland & Halvorsen, 2011). Both oak species seem to tolerate relatively dry habitats (Elven, 2005;Jones, 1959), and their recruitment is limited by light (Annighöfer et al., 2015). These factors may differ strongly between forest types and forestry regimes.
Thus in forests, oaks are related to a combination of processes and variables associated with climate, topography, and land cover.
The distribution of veteran oaks, that is large and/or hollow oaks, may differ from the distribution of oak in general, because the development of large trees and tree hollows requires a long time without major disturbances. Most of the productive oak forests in Southern Norway have been heavily exploited for timber production. Large-scale logging of oak forest in Norway started already in the 1,500s, peaked around 1,650then the oak forests closest to the coast were heavily exploited and logging moved inland-and continued on a reduced scale until approximately 1,900 (Moore, 2010;Vevstad, 1998). The past 100 years or so, little logging of large oaks has taken place, but there has not been enough time for old-growth oak forests to develop. In modern forestry, many oak forests have been logged and replaced by faster-growing Norway spruce. It is unclear; however, to what extent veteran forest oaks are now concentrated in remote, inaccessible, and/or low-productive forest areas.
Veteran oaks are also found in open landscapes. Oaks recruit naturally from trees within the open landscape or nearby forest. However, oaks are also of significant cultural importance and have been planted in courtyards, gardens and parks, and along roads and field margins. As in forests, oaks may have suffered from intensification of production in the agricultural landscape, where edge habitat and other marginal areas have been sacrificed in the creation of larger, more homogeneous production areas. In gardens and urban areas, oaks and other park trees are subject to various pressures from changing gardening trends and preferences related to safety.

| Data
We used data on veteran oak presences and absences from the pilot study behind the recently established national monitoring of veteran oaks (Sverdrup-Thygeson et al., 2013). The data were collected as a stratified random sample within seven regions of approximately 200 km 2 subjectively selected to represent the variation in the oak region in south-eastern Norway. Each region was divided into 3 × 3 km blocks. We first randomly selected three blocks, and then 20 plots of 500 × 500 m within each block ( Figure 2). Based on existing knowledge on occurrence of hollow oaks (primarily the database of the Norwegian Environment Agency; http://kart.naturbase. no/) we sorted the plots into "oak present"-plots and "oak presence unclear"-plots. All "oak present"-plots and 25% of the "oak presence unclear"-plots were visited in the field, and the position of all large and hollow oaks in the plot was determined with a hand-held GPS. To prepare for analysis, each plot was gridded to obtain information on the presence and absence of veteran oaks in 10 × 10 m cells matching the geographical resolution of the predictors.
Potential predictors were collected from digital maps in a geographical information system (GRASS Development Team, 2015).
Because the goal was prediction of oak presence across the landscape, we could only use variables for which (more or less) full-cover maps were available. The variables considered were of two main kinds: (i) Geographical variables, reflecting large-scale geographical gradients (e.g., latitude, longitude, elevation) and local topography (e.g., slope and aspect), and (ii) land cover variables, reflecting vegetation, landscape structure and human land use (e.g., forest, open landscape, distance to roads). When relevant, we considered variables at different spatial scales. Based on the literature and initial screening of about 80 potential predictor variables, we decided to test a set of 13 variables (Table 1) representing more or less independent aspects that could potentially affect the occurrence of veteran oaks (e.g., Dahl, 1998;Elven, 2005;Jones, 1959;McEwan et al., 2011;Moore, 2010;Stokland & Halvorsen, 2011;Vevstad, 1998), and for which data with more or less full geographical cover could be obtained (see Appendix S1 for further details on the variable screening process).

| Analysis
We developed prediction models for the full data set across all landscape types and for each landscape type separately by means of logistic regression (GLMs; McCullagh & Nelder, 1989). Visual inspection of the data indicated that some of the predictors could be nonlinearly related to oak presence. However, we found no support for strong nonlinearities in the initial model testing (GAM yielded essentially the same linear models as GLM), except for one predictor (slope), where the inclusion of a squared term (slope 2 ) was necessary to account for a nonlinear response.
We developed a suite of models to elucidate key patterns and ensure robustness of the results with respect to our predictions.
We analyzed a few single regression models with selected key predictors to test for human-modified ecological responses (prediction 1) and different effects in forests and open landscapes (prediction 2). To evaluate the relative importance of these and other predictor variables we developed multiple regression models with several predictors. To evaluate model uncertainty and assess the need for multiple models (prediction 3) we also calculated AICc-weighted average parameter estimates across multiple alternative models (Burnham & Anderson, 2002), and compared those to the best model (in terms of AICc

| RESULTS
We found that the distribution of veteran oaks was influenced by a When we broadened the perspective to look at regression models with multiple predictors, we found that several variables were important. Although no single combination of variables was clearly better than the others for any of the data sets (the best models receive relatively weak support with AIC weights <0.5, and the confidence sets of models consist of 4-18 models; Table 2), there was a good correspondence between the best models (Table 3) and coefficient estimates based on AIC-weighted model averaging across the confidence set of models ( For statistics on the variables in the different data sets, see Appendix S1: Table S1.

| DISCUSSION
A veteran oak is not just any forest tree. Veteran oaks are essential carriers of biodiversity, and of rich cultural traditions, like many other biodiversity hotspots (Habel et al., 2013;Lindenmayer & Laurance, 2016;Myers et al., 2000;Timonen et al., 2011). Moreover, like an increasing number of small and local hotspots, they are far apart, difficult to locate, and strongly influenced by anthropogenic processes.
Developing models to predict their occurrence is, therefore, critically important and timely, but challenging. In both landscape types, the probability of encountering veteran oaks is low-even in the best spots. Our forest model predicts probabilities of about 0.05 for the best 10% of 10 × 10 m cells. This is still high compared to the ex-  The table shows the four best models (based on AICc) for each landscape type, the number of parameters (k, including geographical parameters), and AICc statistics. All models include the eight geographical variables below the table header "Model" in addition to the variables listed (see Table 1 for variable definitions).
T A B L E 3 Model coefficients of the best logistic regression models for each data set, based on AICc  (Table 1, Appendix S1: Table S1). p-values for coefficient estimates (z-tests): ***<.001, **<.01, *<.05, ˄<.1. (See Appendix S1: Table S2, for extended results.) T A B L E 4 Model coefficients averaged across the 95% confidence set of logistic regression models for each data set and standardized by the SD of the predictor variables (Table 1, Appendix S1: Coefficients with 95% confidence intervals not including zero (see Appendix S1: Table S3, for extended results). One solution to this problem is to split the ecological predictions by landscape type, as we did here (see also Meineri, Skarpaas, & Vandvik, 2012): landscape-specific models can be developed for different landscape types, and predictions merged geographically (i.e., on a map).
This approach requires that landscape types are clearly defined and that information on the spatial distribution of landscape types is readily available.
Finally, the third challenge highlighted by our study is the multitude and complexity of processes and factors affecting biodiversity hotspots in human-influenced landscapes (prediction 3). Ecological processes are complex, anthropogenic processes even more so. This challenge goes far beyond the time-and space-dependent relevance and importance of single-predictor variables discussed above. It is clear from our study of veteran oaks that no predictor can be singled   (Framstad & Lid, 1998), but many oaks may have been spared as ornamental trees, especially toward higher elevations where oaks are less common (see further discussions of transitional landscapes in Appendix S1). Thus, the forest area variable seems to capture anthropogenic landscape structures and processes of importance to veteran oaks. Distance to road and water, on the other hand, did not have significant predictive power (Tables 3 & 4). This result was unexpected, as timber is largely transported by road in present-day forestry, whereas log driving on rivers and lakes was the main means of moving timber from the forest to the sawmills in the past (Sandmo, 1951). Steepness (and productivity) of the terrain seem to better represent effects of logging activities on oak than distance to roads and water. However, direct data on past logging activities would clearly have been much more informative than proxies based on current landscape structures.
This underscores the need for keeping track of major human land use activities for future studies of their effects, especially for systems involving slow ecological processes.
Despite the challenges discussed above, our spatial models for veteran oaks provide clear results of relevance of biodiversity management and conservation as well as further research. It is evident from our results that veteran oaks are influenced by more than the natural factors shaping the fundamental oak niche. Our models suggest that elevation, terrain wetness, and landscape structure (forest area) are important predictors of veteran oak presence, in accordance with the fundamental oak niche (Dahl, 1998;Jones, 1959;Stokland & Halvorsen, 2011), but with differing responses and additional variables playing major roles in forested and open landscape types, as discussed above. We now have a workable set of prediction models that can help us design mapping and monitoring efforts, improve estimates of veteran oak abundance, guide conservation management (Lindenmayer et al., 2014), and support research addressing issues such as cost-effective probability-based sampling (Yoccoz, Nichols, & Boulinier, 2001), effects of landscape structure and connectivity (Evju, Blumentrath, Skarpaas, Stabbetorp, & Sverdrup-Thygeson, 2015;Evju & Sverdrup-Thygeson, 2016;Sverdrup-Thygeson, Skarpaas, Blumentrath, Birkemoe & Evju, in press), predictions of species richness (Skarpaas et al., 2011), and spatial community dynamics (Engen, Saether, Sverdrup-Thygeson, Grøtan, & Ødegaard, 2008).
To summarize and conclude, we find that veteran oaks are predictable despite the complexity of processes in human-influenced landscapes and that considering how different human-related processes operate in different landscapes helps both understanding veteran oak responses to environmental variables and prediction of distribution patterns. We expect predictions to be further improved with the extensive monitoring data set under establishment, especially after repeated visits. This will document recruitment and mortality patterns in veteran oaks and can be used to develop increasingly refined process-oriented models. Regardless of the modeling approach, finding ways to account for human influence on ecological systems and address the challenges illustrated by the veteran oak case are likely to become increasingly important in the Anthropocene.

ACKNOWLEDGMENTS
This study was carried out under the projects "Survey and monitoring of red-listed species" (ARKO, funded by the Norwegian Environment Agency), and "Management of biodiversity and ecosystem services in spatially structured landscapes" (funded by the Norwegian Research Council, grant 208434/F40).