Variation in dispersal traits and geography predict loss of ranges due to climate change in cold-adapted amphibians

We focused on the six North American cold-adapted amphibians to project range shifts in response to climate change with and without dispersal and demographic parameters included in the models. Current range estimates by the International Union for Conservation of Nature (IUCN) for these species contain regions of northern Canada and Alaska. These estimates are from expert testimony and publications compiled for the Global Amphibian Assessment (available at http://​www.​iucnredlist.​org/​). The six species included were: Long-toed Salamander, Ambystoma macrodactylum; Canadian Toad, Anaxyrus hemiophrys; Western Toad, Anaxyrus boreas; Boreal Chorus Frog, Pseudacris maculata; Wood Frog, Lithobates sylvaticus; and Columbia Spotted Frog, Rana luteiventris. These species likely exhibit unique physiological tolerances, behavior, and occupancy of niche space compared with other North American species that allow them to inhabit colder climates. For example, L. sylvaticus’s northern extent is correlated with freeze tolerance (Store and Storey 1984). In contrast, A. boreas survive freezing air temperatures by moving to mammal burrows to remain between 4.8 and 7 ℃ (Mullally 1952). Three of the species have western North American distributions with high amounts of overlap (R. luteiventris, A. boreas, and A. macrodactylum) compared to those in central North America, allowing for analysis of variation in the response to identical environmental heterogeneity. We created future distribution estimates for the cold-adapted amphibian species in North America using three sets of models: zero, unlimited, or intermediate dispersal. The unlimited dispersal would represent a standard approach of only using an ENM (Seaborn et al. 2020). We used zero dispersal as the scenario with the potential for the greatest reduction in species ranges. We ran models for 100 years, using the machine-learning algorithm Maxent in R (Phillips et al. 2004; Phillips et al. 2006; v4.0.5, R Core Team 2021). All scripts for re-creating the analyses are available at https://​github.​com/​trasea986/​cc_​disp_​amphib.

Presence locations for all species were pulled from the Global Biodiversity Information Facility (www.​gbif.​org) using the ‘rgbif’ R package (v3.6.0, Chamberlain et al. 2021; GBIF Occurrence Download 2021). We used the library ‘tidyverse’ for data management steps and to enhance reproducibility (Wickham et al. 2019). To minimize missing available data due to taxonomic synonyms or other challenges, we used the package ‘taxize’ (v0.9.98, Chamberlain et al. 2020). Global position system (GPS) points of presence that were incomplete for important fields (e.g. missing latitude or longitude or having coordinate uncertainty > 1 km) were removed. We also removed any points before 1970, to align with the averaging done to calculate the present-day environmental variables. In addition, the function clean_coordinates from the package ‘CoordinateCleaner’ was used to check and remove points located exactly in the center of a country, equal latitude and longitude values, the area around the GBIF headquarters, locations near biodiversity institutions, zero values for latitude and longitude, and any points landing in the ocean (v2.0, Zizka et al. 2019). We then retained one random point when two points were within 2.5 km of each other to correct for sampling bias through spatial filtering which can improve model predictions (Kramer-Schadt et al. 2013). The total number of presence points ranged from 71 to 4440, depending on species, with an average of 1252 points (Supplementary Table 1, Supplementary Fig. 1). After downloading points, we reprojected to North America Albers Equal Area Conic, which was the coordinate reference system (CRS) used for all GIS work. All major geographic information system (GIS) manipulations were done using the ‘raster’, ‘terra’, ‘sp’, or ‘sf’ packages (v3.5–11, Hijmans 2021a; v1.4–22, Hijmans 2021b; Pebesma and Bivand 2005; Bivand et al. 2013; Pebesma 2018). We did not do any additional pruning due to disjunct locations (e.g. L. sylvaticus in Colorado) or potential disjunct locations to increase reproducibility and reduce point retention bias.

To begin, we used the 19 Bioclimatic environmental layers, all extracted at 30-s resolution from the WorldClim database (www.​worldclim.​org, Hijmans et al. 2005). The WorldClim methods were based on methods originally used in the development of the BIOCLIM package, and now are widely used in conjunction with ecological niche models (Booth et al. 2014). These layers represent a baseline climate scenario for present day derived from averaging from 1970 to 2000 and encompass primarily the pre-climate change state. We clipped geographic range of the layers to the northwest hemisphere of the world and re-projected, setting a latitude range of 20 to 75 degress, and a longitude range of − 180 to − 45. After reprojection, cell size was ~ 575 m (the equal area projection enabled this to be consistent across the full area). We then extracted environmental layer values at the locations of the occupancy points for all species, and calculated Spearman’s correlation (Supplementary Fig. 2). We used a correlation cut off of 0.80 and the ‘findCorrelation’ function from the ‘caret’ package to remove the most correlated variables (v6.0–90, Kuhn 2021). One variable was removed from the pair randomly. The final bioclimatic environmental layers were Annual Mean Temperature (BIO1), Precipitation of Warmest Quarter (BIO18), Precipitation of Coldest Quarter (BIO19), Mean Diurnal Range (BIO2), Isothermality (BIO3), Max Temperature of Warmest Month (BIO5), Temperature Annual Range (BIO7), Mean Temperature of Wettest Quarter (BIO8), Precipitation of Wettest Month (BIO13), and Precipitation Seasonality (BIO15).

We then prepared rasters for future environmental conditions for multiple shared socio-economic pathways (SSPs) for the time periods of 2021–2040, 2041–2060, 2061–2080, and 2081–2100 from WorldClim. We used the SSP2-4.5, SSP3-7.0, and SSP5-8.5 scenarios (Riahi et al. 2017) from the Coupled Model Intercomparison Project 6 (CMIP6) for the WorldClim variables with the 2.5-min raster resolution, the finest resolution available at the time (May 2022), with the CamESM5 circulation model. The notation of the SSPs can be interpreted as the first number after “SSP” being the scenario and the second set of numbers being the forcing levels. These scenarios represent a range of carbon emission scenarios and are the updated versions of the previous representative concentration pathways (RCPs): RCP4.5, RCP6.0, and RCP8.5. With these updated scenarios, SSP3-7.0 represents the intermediate outcome. We then used the ‘resample’ function from ‘terra’ to match the present-day raster resolution.

We used Maxent version 3.4.3 for ecological niche model creation, using cell value as a proxy of habitat quality (Phillips et al. 2004; Philips et al. 2006). We took several steps to optimize the Maxent models based on best practices (Radosavljevic and Anderson 2014). To minimize overfitting by having too large of a background extent, we then cropped all layers for each species by creating a shapefile by buffering a circle around all occupancy points by 500 km. Because of Maxent using random points as pseudo-absence locations, the background extent needs to be within an area that is realistic to avoid poor predictions and performance (Merrow et al. 2013; Bergman et al. 2024). We then used ENMeval to tune the fitting features and regularization multiplier with 10,000 background points and checkerboard partitioning (Muscarella et al. 2014). ENMEval runs a combination of regularization parameters and fitting methods and evaluates each model with AIC to avoid overfitting. For the features, we tested linear; quadratic; hinge; linear and quadratic; linear, quadratic, and hinge options. For the regularization multiplier we tested the values 0.5, 1, 2, 3, and 4. For L. sylvaticus, we used a raster resolution of 2.5-min for just the ENMeval step due to computational limitations with the original resolution requiring more than 1 TB of RAM. We then ran Maxent for 10 replicates for each stepwise value for each species with jackknife and cross validation to evaluate model fit with AUC and individual environmental variable contribution. We then used the best performing model of the replicates based on AUC to establish the proxy for habitat suitability for the dispersal simulations.

After creating these initial models, we used the projection feature to create a raster of the relative probability of suitability for each individual species across all of North America within the bounds of the original cropping of the rasters. We used this probability as a surrogate for habitat quality, assuming that areas within the range with suitable climatic conditions are currently occupied and presents a best case scenario for future distributions. Although predicting outside of the training location adds additional layers of assumptions to the models and simulations, we wanted to constrain future predicted habitat by niche space and not geographic space, outside of the defined dispersal limitations of the simulations and the current range of the species. We scaled relative probability of suitability to a range of 0 to 1000, from the initial scale of 0 to 1. To establish the initial threshold for suitable habitat and the initial distribution of the species, we calculated the 90% quantile for each species’ occupancy points with the present-day distribution model. This threshold was then used to define the present-day distribution and also establish the minimum relative probability of suitability value for suitable habitat during the climate change projections.

We created models using the R package MigClim (Engler et al. 2012), which combines ecological niche models with dispersal constraints. These models allowed us to incorporate probabilistic parameters, such as a dispersal kernel and probabilities of long-distance dispersal events, to simulate changes in species’ distributions with shifting habitat suitability through time, although not all parameters were used in every model. MigClim operates by identifying areas that are unoccupied and suitable habitat, and calculates the probability of dispersal by considering all occupied locations, similar to a cellular automaton process where each cell has a single population with the user putting in species-specific dispersal and life history parameters. We used yearly timesteps, with habitat suitability changing every 20 years to represent the time steps of the future environmental conditions, with the first 20 years representing the present-day climate (from WorldClim defined from 1970 to 2000 data). Models were run for 100 years with 20-year environmental increments of change. Habitat was deemed suitable for colonization if the raster cell value was greater than the 90% quantile threshold described above, so that the raster cell functionally had a suitability that would capture the majority of known occupancy points. We used a range of biologically relevant values for dispersal and demographic parameters for a sensitivity analysis to determine the parameters which influenced range dynamics, either contraction or expansion, over the course of the model (Table 1). We did not tailor model sets to each species because individual dispersal and demographic parameters were not available from the literature. We calculated the proportional change in final distribution area as the final area minus the initial area, divided by the initial area, in response to changing one of the parameters individually while keeping the rest constant. We then took this number and multiplied it by the ratio of the initial parameter to the amount of area changed to normalize for the different units. The initial model used one dispersal distance of ~ 575 m, the size of a single raster cell after downscaling and reprojecting, no long-distance dispersal, a probability of dispersal of 0.1, age at first dispersal of 1 and dispersal only occurring during one year of the individual’s lifetime. As an examples of the sensitivity analysis, we calculated amount the range changed (area in m²) for each increase in dispersal distance by the ~ 575 m steps defined by the raster cell size, or amount of range changed when age at first dispersal was increased. Each species had total of 83 simulations run, which occurred due to three SSPs and 27 sets of parameter values used, including one zero and one unlimited dispersal scenario. We did not test every combination of variables in a full factorial sensitivity analysis, and instead changed each parameter value while holding all other values constant to consider each variable independently. When using a dispersal distribution to test the impacts of the dispersal in the distribution curve, we derived the dispersal kernel from a negative exponential function, which is appropriate for the right-tailed distribution of amphibian dispersal events (Breden 1987; Sinsch et al. 2012). We then calculated the ranking of the various species, based on their average range change across all models as a proportion of initial range size. All models were run with 5 replicates. It is important to note that the goal was to test the sensitivity to these parameters, and values may not perfectly mirror reality.

We used a range of values to model dispersal of these species covering the range of amphibian dispersal abilities reported in the literature. For example in A. macrodactylum, salamanders colonized sites up to almost 1.2 km after fish were removed (Funk and Dunlap 1999). In L. sylvaticus, juveniles are the stage at which most dispersal occurs, the maximum distances occur at around 1.5 km, and the rate of dispersal of hatchlings out of the pond has been estimated at roughly 0.2 (Berven and Grudzien 1990). The other ranid in this study R. luteiventris, has a maximum migration distance from 1 to 6.5 km depending on location (Bull and Hayes 2002; Engle 2001; Funk et al. 2005), and may have the highest rate of dispersal, estimated at 0.6 (Funk et al. 2005). In the two Anaxyrus species, A. boreas has traveled up to 6 km among monitored sites (Muths et al. 2003), while A. hemiophrys averaged cumulative movements of about 1.5 km per year, but maximum distance moved in was greater than 4 km (Constible et al. 2010). Boreal chorus frogs (P. maculata) are potentially the least mobile of the group, with maximum movements at just under 0.7 km (Spencer 1964). The dispersal distance of 20 km we used for the maximum is roughly equivalent to half the dispersal modalities of the cane toad (Phillips et al. 2007), an overestimate for these species, but we included it to test whether even the most mobile amphibian would completely fill the newly suitable habitat to the north of the current range.

To estimate the fate of individuals contained in southern edge populations, which represent the trailing edge, we created a second model set of the same Maxent outputs for environmental suitability and parameter values but used the aforementioned southern range initial distribution. We did this to provide greater resolution of the range dynamics along the southern edge. By tracking occupied cells along only the southern edge, we could evaluate whether cells near the core would become occupied by southern range edge individuals at rates faster or slower than the loss of habitat along the southern edge. Again, to create this initial distribution, we shifted the IUCN range north 250-km, and then masked the area between the shifted and original range by the ENM values as in the whole distribution. This initial distribution created large areas of suitable but uninhabited areas near the core of the range, and through time allowed for us to track potential colonization of these areas of individuals from the southern range edge. We then used this to calculate the change in proportional area and sensitivity to dispersal parameters. This presented four potential scenarios. A value of zero represented the loss of range due to climate change being equally compensated by individuals moving and colonizing suitable, vacant habitat closer to the core, whereas value from 0 to -1 would represent overall range contraction. In addition, we ran the same sensitivity analysis as outlined above for the full species’ ranges.

Across species, there was variation in model fit and best model parameters based on tuning using ENMEval. Optimal features for fitting were either hinge or a combination of linear, quadratic, and hinge. The regularization multiplier ranged from 0.5 to 2 (Supplementary Table 2). After tuning, the test AUC of the best model used for the dispersal simulations ranged from 0.817 to 0.985, with a mean AUC of 0.915 (Supplementary Table 2, Fig. 1). Predicting outside of the geographically constrained range highlighted similar geographic space based on niche, even for present day predictions, and these areas outside of the known range were often the locations where relative probability of suitability values substantially increased with the climate change scenarios (Supplementary Figs. 3, 4 and 6). Permutation importance of the individual variables varied greatly with species, although BIOCLIM19 (‘Precipitation of the Coldest Quarter ‘) was the most important for 3 species (Supplementary Table 3; Supplementary Fig. 7).

Overall changes in range area were most dependent on dispersal scenario, but also varied with carbon emission scenario (Table 2). In the zero dispersal scenario, all ranges declined in range size regardless of carbon emission scenario. In contrast, under the unlimited dispersal scenario, ranges almost always expanded for all three SSPs (Fig. 2 for SSP5-8.5; Supplementary Fig. 8 for SSP2-4.5; Supplementary Fig. 9 for SSP3-7.0; Table 2). This was generally most apparent in the more central North American species (L. sylvaticus, P. maculata, and A. hemiophrys) as opposed to the western species (A. macrodactylum, A. boreas, and R luteiventris). Loss along east to west lines was even for most species except for L. sylvaticus and P. maculata. Generally, the difference between SSP5-8.5 and the lower emission scenarios was the same pattern but increased effect size under SSP5-8.5.

When dispersal was incorporated and averaged across all model parameter sets, the average dispersal model was generally closer to the zero-dispersal model than the unlimited dispersal scenario (Table 2) although the amount of area varied greatly across species and SSPs, which highlights variation within the cold-adapted amphibian group (Fig. 3). The full potential new habitat was never completely colonized, in particular when using the base model values which are likely biologically more relevant than some of the longest dispersal parameters tested. Under SSP5-8.5, three of six species, A. macrodactylum, P. maculata, and L. sylvaticus, declined in range size on average across all dispersal models, but the average range size of all cold-adapted amphibians increased (Fig. 4). Conversely, under the SSP2-4.5 two of the cold-adapted amphibian ranges, P. maculata, and L. sylvaticus,declined with the average range of the dispersal models and overall results were less drastic (Table 2, Supplementary Fig. 10). The most negatively affected species was always L. sylvaticus regardless of emission scenario or dispersal scenario. SSP3-7.0 was intermediate, but closer to SSP5-8.5 (Table 2, Supplementary Fig. 11). The least affected species were generally the Pacific coastal species.

Sensitivity analysis revealed that the parameter driving how much the area of the range changed across SSPs was most frequently dispersal distance, long-distance dispersal distance, or the probability of dispersal for all species (Table 3; Supplementary Table 4 for raw values). The order also varied within a species across SSPs with two of the species. This highlights that initial starting distribution and the rate of environmental change interact with demographic and dispersal factors to determine range dynamics.

The ability for individuals from the southern range edge populations to move toward the core and colonize at a rate to offset population loss in many parts of the southern range was dependent on the dispersal model and scenario. Individual dispersal model sets were not always able to match the rate of shifting appropriate habitat across the range, even if the average proportional change in range size was positive, indicating longitudinal variation in climate change response (Fig. 5). Contraction along the southern edge was highest with SSP845 (Table 2). The most impacted species by far was L. sylvaticus, which experienced complete extinction under some dispersal models. This could be due to the higher threshold of appropriate habitat due to high overall relative probability of suitability values at occupancy points. Similar to the full range models, changes were highest for SSP845, followed by SSP3-7.0, and then SSP2-4.5 (Supplementary Figs. 12 and 13).

The sensitivity analysis of the southern range edge showed similar parameters as the full range, with some variation within a species. The differences between the southern range edge models and the full range models may represent differences in the response along the northern range edge or rate of environmental change occurring at the southern range edge. Unlike the full range models, there were no between-species differences once simulations with extinctions were dropped.