Incorporating habitat suitability into community projections: Ant responses to climate change in the Australian Wet Tropics

Studies of climate change impacts on animal distributions typically consider only the direct impacts of a changing climate, under the assumption that future areas of suitable climate will otherwise remain ecologically suitable. Here we assess both the direct and indirect impacts of climate change on rain forest ant communities, where substantial shifts are projected to occur for both climate and habitat types.

Community-level responses are more complicated, as changing temperature is likely to have differential effects among species (Elmendorf et al., 2012). Species interactions also may change disrupting mutualisms, trophic interactions, competitive hierarchies and ultimately species occurrence (Blois, Zarnetske, Fitzpatrick, & Finnegan, 2013;Sentis, Hemptinne, & Brodeur, 2014;Sorte & White, 2013). Therefore, forecasting changes in the distribution and composition of biological communities under future climates can be extremely challenging.
The potential decoupling of species associations due to differential climate change sensitivity can have particularly dramatic implications if vegetation and its associated fauna are differentially sensitive (Caddy-Retalic et al., 2018;Carvalho, Brito, Crespo, & Possingham, 2010;Garcia, Cabeza, Rahbek, & Araújo, 2014). Studies of climate change impacts on animal distributions typically consider only the direct impacts of a changing climate, under the assumption that future areas of suitable climate will otherwise remain ecologically suitable. However, this would not be the case if changes in climate indirectly reduce habitat suitability. This might occur as a direct response of vegetation to a changing climate, or indirectly through changes in, for example, fire regimes (Davis, Higuera, & Sala, 2018;Regos, Clavero, & D'amen, Guisan & Brotons, 2018). Changes in vegetation structure can have a greater effect on microclimate than changes in macroclimate (Williams, Bolitho, & Fox, 2003).
Vegetation change would be particularly dramatic for fauna if it rendered the habitat largely unsuitable. This is especially relevant in high-rainfall regions of the tropics where savanna and forest co-occur and readily transition from one to the other with changes in climate and fire regimes (Dantas, Hirota, Oliveira, & Pausas, 2016;Langan, Higgins, & Scheiter, 2017;Murphy & Bowman, 2012). Savanna and forest support highly disjunct faunas (Murphy, Andersen, & Parr, 2016), such that transition from one vegetation type to the other results in profound faunal change. For example, in northern Australia the ant faunas of co-occurring rain forest and savanna are remarkably distinct from each other (Andersen, Ingen, & Campos, 2008), with the rain forest fauna dominated by shade-tolerant, forest specialists, in contrast to the arid-adapted taxa that dominate savannas (Andersen et al., 2008;Reichel & Andersen, 1996). If forest switches to savanna under future climates, then the habitat will be unsuitable for forest species even if the macroclimate is suitable for them. The critical role of the future distribution of habitats upon which species depend has been largely overlooked in studies regarding climate change impacts on biodiversity.
Changes in community composition can be modelled using stacked species distribution models (SDMs), but these have a poor record of predicting change even when environmental correlates of species occurrences are well known (D'Amen, Pradervand, & Guisan, 2015;Pellissier et al., 2013). One reason for this is that SDMs can predict if a species might occur in a general area that is climatically suitable, but they are limited when predicting if they will occur in any particular site, which may only have a fraction of the regional species pool, due to biotic interactions between species. Additionally, commonly used species-level modelling approaches (Elith & Leathwick, 2009) can vary in accuracy. For example, a developed modelling algorithm based on rank-ordered stacking of SDMs, compared to a binary stacking, can reduce variation of error in predicting community composition turnover along environmental gradients, but it still requires caution when predicting the species diversity (Del Toro, Ribbons, Hayward, & Andersen, 2018).
An alternative approach to predicting responses of biological communities to climate change is Generalized Dissimilarity Modelling (GDM), a nonlinear extension of matrix regression that models pairwise biological dissimilarity between sites as a nonlinear function of pairwise site difference in environmental and geographic variables (Brown, Cameron, Yoder, & Vences, 2014;Ferrier, Manion, Elith, & Richardson, 2007). Critically, by working directly with species assemblages, GDM can account for biotic interactions which are not typically included in SDMs. GDM extrapolates patterns of compositional turnover beyond sampled communities and is particularly well suited to communities with high levels of beta diversity.
This analytical method can accommodate almost any measure of geographic or ecological separation as a predictor, including organism-specific representations of barriers to dispersal, or cost of movement/gene flow through unfavourable habitat (Fitzpatrick, Keller, & Vellend, 2015). GDM can also rapidly analyse datasets containing very large numbers of species, regardless of the number of records per species. Furthermore, by studying the emergent rates of spatial turnover along environmental gradients under current climatic conditions, GDM can be used to predict the temporal rate and spatial distribution of turnover under future climates (Blois, Williams, Fitzpatrick, Jackson, & Ferrier, 2013). Compared with SDMs, GDM can predict higher turnover and across a larger contiguous area (Fitzpatrick et al., 2011), where problems such as data limitations or indirect predictors such as habitat changes (Guisan & Thuiller, 2005) are likely to hinder the application of SDMs.

K E Y W O R D S
climate change, composition turnover, ecological environments, GDMs, habitat condition, rain forest Generalized Dissimilarity Model commonly accommodates variation in rates of species turnover along environmental gradients Maestri, Shenbrot, & Krasnov, 2017;Pennifold et al., 2017), but, unless programmed to do so, does not account for changes in habitat. However, GDM analyses can produce a variety of metrics to reveal and describe the relative biotic uniqueness of environments, and the proportion of those environments falling within different habitat classes (Jiménez-Alfaro et al., 2018;Johnston, Elmore, Mokany, Lisk, & Fitzpatrick, 2017). Therefore, it is important to simultaneously consider changes in both habitat and climate within a GDM framework, which works in multidimensional environmental space, projected both temporally and spatially.
Invertebrates are often used to monitor ecological change in terrestrial ecosystems because of their critical roles in ecosystem function and dominant contribution to faunal diversity (McGeoch, 1998).
Ants are particularly commonly used in environmental assessments, due to their ecological dominance and sensitivity to environmental change (Andersen, Fisher, Hoffmann, Read, & Richards, 2004;Andersen & Majer, 2004;Underwood & Fisher, 2006). Patterns of diversity and composition in ant communities are strongly related to climatic variables such as temperature, precipitation and humidity (Diamond et al., 2012;Dunn et al., 2009;Kaspari, Alonso, & O'Donnellkwd, 2000;Sanders, Moss, & Wagner, 2003). In this study, we apply GDM to a comprehensive dataset for the diverse rain forest ant communities of the Australian Wet Tropics (AWT) to investigate potential impacts of future climate change based on contemporary spatial patterns of compositional turnover.
Our specific aims are to: (a) identify environmental variables that are most correlated with compositional dissimilarity in rain forest ant communities throughout the AWT, using GDM combined with a theoretic approach; (b) use the GDM to forecast changes in the area of climatically suitable habitat for different ant community types; and (c) predictions to F I G U R E 1 The Australian Wet Tropics bioregion with site locations. Dark shade represents the extent of rain forest within the region account for both the direct effects of climate on ant community changes and the indirect effects of habitat loss on the extent of suitable habitat for the community. Our study will allow the identification of areas in the AWT that are likely to be most sensitive to climate change impacts and are therefore priorities for conservation management.

| Study sites and compositional data
Mean annual rainfall in the AWT (Figure 1) varies from about 1,500-9,000 mm, with 75%-90% occurring between November and April (McDonald & Lane, 2000). Within the region, mean temperature declines at a rate of about 1°C for every 200 m increase in elevation (Shoo, Williams, & Hero, 2005a). Our study was confined to rain forest and covered a range of structural rain forest types from complex mesophyll vine forest in the fertile lowlands to upland simple notophyll vine forest (Adam, 1992;Hilbert, 2008).
Despite their relatively small area, the AWT rain forests are recognized as a major biodiversity hotspot of global significance due to their extraordinary biological richness and biogeographical uniqueness (Zachos & Habel, 2011). However, this biodiversity is highly threatened by anthropogenic climate change Suppiah, Macadam, & Whetton, 2007;Thomas, Cameron, Green, & Bakkenes, 2004). The rain forests are predicted to decrease in extent by 60% with a temperature rise of 1°C combined with a 10% decrease in precipitation . Future projections show significant decreases in core distributional area of all 65 endemic vertebrates (Williams et al., 2003) and in population size of 74% of rain forest birds (Shoo, Williams, & Hero, 2005b). The insects of the AWT are also highly threatened by climate change, with predictions of the extinction of a large proportion (40%) of schizophoran flies (Wilson, 2010), and 88% reduction in population size of flightless ground beetles (Staunton, Robson, Burwell, Reside, & Williams, 2014).
We used a recent dataset of 5,204 occurrence records of 296 ant species comprehensively sampled using bait traps on the ground and tree trunks, litter extractions and pitfall trapping from 150 sites that cover the full latitudinal and elevational range of rain forest in the AWT (Nowrouzi et al., 2016). A combination of techniques was used to target the distinct faunas associated with litter (cryptic species), the litter surface (epigaeic species) and trees (arboreal species) (Bruhl, Gunsalam, & Linsenmair, 1998), with high sampling intensity to provide reliable estimates of species composition (Ashcroft et al., 2010). Rarefaction curves indicated that most species occurring in the region were recorded (Nowrouzi et al., 2016). The sites were distributed across six mountains, ranging from the Finnegan uplands in the north, to Mt Spec in the south, over a distance of approximately 500 km. All sites were located on granite-derived soils except for those in the Atherton subregion which contains more-fertile basaltic soils (McJannet, Wallace, Fitch, Disher, & Reddell, 2008;Parsons & Congdon, 2008). The elevational range of study sites varied among subregions, due to differences in the availability and accessibility of rain forest habitats. Sampling transects were separated by 200 m elevation from the lowland rain forests to the upland following the elevation contour, with six sites separated by 200 m along each transect. Only three sites were located at each of the 350 m elevation transect at Mt Spec and 100 m elevation transect at Atherton due to limited rain forest cover. All ants were sorted to species and where possible named through comparison with identified specimens held in the CSIRO Tropical Ecosystems Research Centre in Darwin. Unidentified species were assigned species codes that apply only to this study and highly diverse genera were identified to species group following Andersen (2000).

| Environmental variables, habitat condition and climate projections
We generated a comprehensive set of environmental variables including bioclimatic variables (BC01-BC35), soil and topographic layers, along with information on litter depth, canopy cover and disturbance.
Bioclimatic data were used to represent the baseline climate, defined as a 30-year average from 1976 to 2005. These data were derived using the ANUCLIM 6.0 software (Xu & Hutchinson, 2013) at a 9-s resolution, approximately 250-m grids (Hutchinson, Stein, & Stein, 2009  au/en/clima te-proje ction s/clima te-futur es-tool/), which has been built on Climate Futures Framework (Clarke, Whetton, & Hennessy, 2011;Whetton, Hennessy, Clarke, McInnes, & Kent, 2012) to select the climatic models. This tool includes projections from global and regional climate models based on simulations performed for the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (Pachauri et al., 2014). We selected 2 out of 15 global climate models that best reproduced Australian average  patterns of temperature and rainfall (Suppiah et al., 2007): (a) "mild" (Model for Interdisciplinary Research on Climate F I G U R E 2 Current habitat condition of the rain forest in the AWT (a) and a projection in future (b, MIROC 2085). Darker colour represents higher-condition habitats in terms of capacity to support the maximum number of plant species. Site locations are presented for clarity F I G U R E 3 Profile of condition values for AWT rain forest habitat under the current climate (a) and a 2085 projection, MIROC (b) (MIROC) -H; ~125 km resolution) and (b) "harsh" (Geophysical Fluid Dynamics Laboratory [GFDL] 2.0; ~300 km resolution), under the greenhouse gas concentration RCP8.5 scenario. Future projections of rain forest condition were also extracted from SDMs of projected rain forest vegetation layers that were derived using the ANUCLIM 6.0 software (Xu & Hutchinson, 2013) at a 9-s resolution and 250-m grids (Hutchinson et al., 2009), provided by eResearch Centre, James Cook University, Australia.

| Environmental predictors of compositional dissimilarity
We first selected 17 out of 28 environmental (19 bioclimatic, six soil and three habitat) variables using Pearson correlation coefficients to exclude highly correlated variables in the dataset (Table A1).
We then assessed which environmental variables best explained dissimilarity of ant communities across the sample sites using the variable selection strategy detailed in Williams et al. (2010) and GDM (Ferrier et al., 2007). This strategy ranks the variables based on their "explained deviance" in the model and selects the best predictors (which reduced the number of variables to 9 out of 17 in this study). To increase the certainty of our variable selection, we also assessed them by using an information-theoretic approach (Burnham & Anderson, 2002). This approach fits multivariate Generalized Linear Models (GLMs) for species composition, developed by Wang, Naumann, Wright, and Warton (2012), using the nine selected predictors. We used a model averaging technique that quantifies relative importance (likelihood) of all possible models that can be generated using the combinations of selected predictor variables (2 9 = 512 models in this study case), based on a modified Akaike Information Criterion (AICc). We then evaluated the standardized effect size of each predictor variable by calculating the differences between observed summed Akaike weight and mean summed Akaike weight derived from 999 null datasets, divided by the standard deviation of summed Akaike weights of null datasets. The calculations were all executed using the mglmn R package developed by Katabuchi and Nakamura (2015). When using the information-theoretic approach, we did not directly include geographic distance as a variable because of correlation between climatic variables (e.g., temperature) and geographic location (e.g., latitude and elevation). However, as excluding geographic distance did not change the impact of other variables in the model, we included it as a driver of composition turnover in the model as suggested by Williams et al. (2010).

| Compositional change under future climates
The selected environmental factors were used as covariates in a GDM fitted using a matrix of pairwise dissimilarities in ant species composition among the 150 locations sampled. Dissimilarity was calculated using the Sørensen index (Roberts et al., 2017).
The GDM assumes that the amount of change in species composition expected for any location as a result of climatic change will be equivalent to the compositional dissimilarity currently observed between that location and another location with a current climate matching that projected for the first location (Ferrier, Harwood, & Williams, 2012;Ferrier et al., 2007). Predictions of the community's compositional change were obtained by using the model to transform future climate layers as described below. Analyses were conducted using the ecodist and GDM R packages (Manion, Lisk, Similarly, we calculated the proportion of species retained in the entire rain forest of the region as P i = [A future climate & habitat change / A current ] Z (using z, the exponent of the species-area curve, which is typically 0.25 (Allnutt et al., 2008) and therefore estimate regional species loss. Calculations and projections were conducted using the.

| Environmental predictors of compositional dissimilarity
Nine out of 17 candidate explanatory variables (Table A2) Table 1). The greatest turnover was explained by the gradient of soil bulk density (BD), followed by mean temperature of coldest quarter (MTCQ), mean temperature of warmest quarter (MTWQ), geographic distance and temperature seasonality (TS) ( Table 1). The information-theoretic approach also suggested that turnover could be explained by isothermality, TS, MTWQ, MTCQ, precipitation of wettest period, BD, available water capacity of soil , litter depth and habitat disturbance (Table 2). In contrast to the GDM, the highest effect sizes (>20) were achieved by MTWQ and MTCQ when using the information theory (Table 2).

| Direct effects of climate
Under a mild climate model (MIROC), GDM predicted that the effective habitat area was largely maintained and even increased (particularly in coastal regions) up to 2055 (Figure 4a,b). However, by 2085 the extent of suitable habitat for ant communities had declined by >85%, and 50% of the rain forest ants would lose more than half of their SEE, with the declines occurring primarily inland (Figure 4c; Table 3). Under the harsher GFDL climate model, at least 60% of the communities in the region were predicted to lose more than half of their SEE by 2085, with only some coastal areas remaining essentially unchanged (Figure 5c; Table 3).

| Incorporating indirect effects of changes in habitat suitability
When changes in habitat suitability are incorporated, >90% of the current rain forest, with higher suitability values of 0.3-0.65 ( Figures   2a and 3a), is projected to transition to categories lower than 0.2 by 2085 under MIROC (Figures 2b and 3b). Our modelling indicates that this will result in almost all ant communities in the region experiencing a reduction in their similar habitats (Figures 4 and 5; Table 3).  (Figures 4f and 5f; Table 3). However, despite the major turnover of ant composition at local scales, only 9 (3%) species were predicted to be lost from the AWT as a region under both climate models when habitat suitability was incorporated.

| D ISCUSS I ON
Our study represents the first analysis of climate change impacts on ant communities using comprehensive compositional data and macroecological models. It is also the first study to integrate climateinduced changes in habitat suitability into future projections of a diverse invertebrate group. Both the GDM and information-theoretic approaches indicated that temperature is the most important climatic driver of compositional change in rain forest ant species in either dry (lower temperature) or wet (higher temperature) seasons of the year in the AWT. This result reinforces results from previous studies showing temperature to be a primary driver of ant species diversity patterns, both at global and regional scales (Andersen, 1995;Bishop, Robertson, Rensburg, & Parr, 2014;Diamond et al., 2012;Dunn et al., 2009).
After temperature, habitat disturbance had the second highest effect on dissimilarity of community composition according to the information-theoretic models, and it significantly contributed to the spatial pattering in the GDM. A common effect of habitat disturbance is simplification of habitat structure (Gibb & Parr, 2013;Hoffmann & Andersen, 2003), so disturbance can have especially strong effects in complex habitats like tropical rain forest (Andersen, 2018). Climate change is predicted to increase the frequency of extreme weather events (Williams et al., 2003), including cyclones, which are the major agent of disturbance in the AWT.
Based on our models, 50%-60% of the ant communities across the AWT will likely lose more than half of their SEEs when considering just the direct impacts of climate change. Regardless of differences among the climate models, this increases to >90% when changes in habitat suitability are also incorporated. Such a finding is consistent with the greater magnitude of projected changes in ant community assemblages compared with that of plants reported by Caddy-Retalic et al. (2018). Thus, projections based on only direct effects of climate change can significantly underestimate the expected changes in ant communities.
Most ant communities currently occurring in the AWT's coastal zone are projected to continue to do so over the next few decades. Coastal ant communities are expected to be buffered from the effects of climate change in the near future due to the close proximity of their habitat to the ocean, which has a moderating influence on temperatures (Dowdy, Abbs, & Bhend, 2015). However, a high degree of turnover in ant composition is expected to occur by 2085 even in coastal areas due to the effects of rising temperatures and altered rainfall patterns on rain forest vegetation. Faster and greater impacts can be expected inland of the mountain ranges, where changes in rainfall and temperature will be more severe (Hughes, 2003;Reside et al., 2013;Suppiah et al., 2007;Welbergen et al., 2015). Throughout much of this region, rain forest is expected to convert to savanna, which is entirely unsuitable for specialist rain forest species. Our analyses of forest conditions are unable to differentiate changes in habitat suitability due to transition within rain forest from those due to the conversion of rain forest to savanna. However, up to 90% of current rain forest in the AWT is expected to convert to sclerophyll open forest and savanna woodland by 2085 (Hilbert, Graham, & Hopkins, 2007;Mokany, Westcott, Prasad, Ford, & Metcalfe, 2014).
Some major uplands and mountain peaks of the AWT, including the Windsor, Carbine and Atherton uplands, have maintained complex notophyll vine forests throughout the Pleistocene (Hilbert et al., 2007) and represent key refugia for rain forest flora and fauna (Reside et al., 2013;Welbergen et al., 2015). These uplands support many endemic ant species; for example, 11 ant species are known only from Carbine uplands (Nowrouzi et al., 2016). Our models reinforce the high sensitivity of these uplands to future climate change Mokany et al., 2014) unlikely that such inland refugia will be maintained under future climates.
Our modelling indicates that even with changing rain forest condition, sufficient suitable habitat will remain at the regional level to avoid high extinction rates even under an extreme climate scenario. However, our analyses of compositional change based on SEEs can be considered conservative because it does not account for dispersal limitations. In particular, new SEEs on different mountains might not be accessible for some species, especially for those without winged reproductive, such as within Lioponera and Pseudoneoponera (Peeters & Ito, 2001). Our analyses also do not account for other factors that can affect extinction rate, such as interactions within metapopulations (Morrison, 2002) and allee effects (Luque, Giraud, & Courchamp, 2013). Additionally, our models do not consider the potential increased incidence of habitat disturbance through cyclones, which will potentially have a major impact on rain forest suitability, including through increased risk of weed invasion (Turton, 2012).
In conclusion, our study has demonstrated that rain forest ants in the AWT are likely to be severely impacted under climate change. The direct impact of a changing climate is likely to be substantial, as has been concluded for well-studied groups such as birds, reptiles, mammals, frogs, butterflies and dragonflies (Bush et al., 2014;Shoo et al., 2005b;Williams & Pearson, 1997;Williams, Williams, Alford, Waycott, & Johnson, 2006). However, the changes in ant communities can be expected to be far more profound when changes in habitat suitability are also considered. Our study shows that projections of reduction in ecological

ACK N OWLED G EM ENTS
We are grateful for the help from Kristen Williams and Simon

DATA ACCE SS I B I LIT Y
Data for this study will be available on request and may be found through James Cook University library (https ://doi. org/10.4225/28/5ac2f ba61317f).