Data for: Predicting habitat suitability for Townsend’s big-eared bats across California in relation to climate change
Data files
Dec 12, 2022 version files 874.09 MB
-
COTO_EnvironmentalData.csv
17.07 KB
-
Elevation_Cal.tif
127.66 MB
-
README_Hamilton_etal.md
5.04 KB
-
Slope_Cal.tif
746.40 MB
Abstract
Aim: Effective management decisions depend on knowledge of species distribution and habitat use. Maps generated from species distribution models are important in predicting previously unknown occurrences of protected species. However, if populations are seasonally dynamic or locally adapted, failing to consider population level differences could lead to erroneous determinations of occurrence probability and ineffective management. The study goal was to model the distribution of a species of special concern, Townsend’s big-eared bats (Corynorhinus townsendii), in California. We incorporate seasonal and spatial differences to estimate the distribution under current and future climate conditions.
Methods: We built species distribution models using all records from statewide roost surveys and by subsetting data to seasonal colonies, representing different phenological stages, and to Environmental Protection Agency Level III Ecoregions to understand how environmental needs vary based on these factors. We projected species’ distribution for 2061-2080 in response to low and high emissions scenarios and calculated the expected range shifts.
Results: The estimated distribution differed between the combined (full dataset) and phenologically-explicit models, while ecoregion-specific models were largely congruent with the combined model. Across the majority of models, precipitation was the most important variable predicting the presence of C. townsendii roosts. Under future climate scnearios, distribution of C. townsendii is expected to contract throughout the state, however suitable areas will expand within some ecoregions. Main conclusion: Comparison of phenologically-explicit models with combined models indicate the combined models better predict the extent of the known range of C. townsendii in California. However, life history-explicit models aid in understanding of different environmental needs and distribution of their major phenological stages. Differences between ecoregion-specific and statewide predictions of habitat contractions highlight the need to consider regional variation when forecasting species’ responses to climate change. These models can aid in directing seasonally explicit surveys and predicting regions most vulnerable under future climate conditions.
Methods
Study area and survey data
The study area covers the U.S. state of California, which has steep environmental gradients that support an array of species (Dobrowski et al. 2011). Because California is ecologically diverse, with regions ranging from forested mountain ranges to deserts, we examined local environmental needs by modeling at both the state-wide and ecoregion scale, using U.S. Environmental Protection Agency (EPA) Level III ecoregion designations and there are thirteen Level III ecoregions in California (Table S1.1) (Griffith et al. 2016).
Species occurrence data used in this study were from a statewide survey of C. townsendii in California conducted by Harris et al. (2019). Briefly, methods included field surveys from 2014-2017 following a modified bat survey protocol to create a stratified random sampling scheme. Corynorhinus townsendii presence at roost sites was based on visual bat sightings. From these survey efforts, we have visual occurrence data for 65 maternity roosts, 82 hibernation roosts (hibernacula), and 91 active-season non-maternity roosts (transition roosts) for a total of 238 occurrence records (Figure 1, Table S1.1).
Ecogeographical factors
We downloaded climatic variables from WorldClim 2.0 bioclimatic variables (Fick & Hijmans, 2017) at a resolution of 5 arcmin for broad-scale analysis and 30 arcsec for our ecoregion-specific analyses. To calculate elevation and slope, we used a digital elevation model (USGS 2022) in ArcGIS 10.8.1 (ESRI, 2006). The chosen set of environmental variables reflects knowledge on climatic conditions and habitat relevant to bat physiology, phenology, and life history (Rebelo et al. 2010, Razgour et al. 2011, Loeb and Winters 2013, Razgour 2015, Ancillotto et al. 2016). To trim the global environmental variables to the same extent (the state of California), we used the R package “raster” (Hijmans et al. 2022). We performed a correlation analysis on the raster layers using the “layerStats” function and removed variables with a Pearson’s coefficient > 0.7 (see Table 1 for final model variables). For future climate conditions, we selected three general circulation models (GCMs) based on previous species distribution models of temperate bat species (Razgour et al. 2019) [Hadley Centre Global Environment Model version 2 Earth Systems model (HadGEM3-GC31_LL; Webb, 2019), Institut Pierre-Simon Laplace Coupled Model 6th Assessment Low Resolution (IPSL-CM6A-LR; Boucher et al., 2018), and Max Planck Institute for Meteorology Earth System Model Low Resolution (MPI-ESM1-2-LR; Brovkin et al., 2019)] and two contrasting greenhouse concentration trajectories (Shared Socio-economic Pathways (SSPs): a steady decline pathway with CO2 concentrations of 360 ppmv (SSP1-2.6) and an increasing pathway with CO2 reaching around 2,000 ppmv (SSP5-8.5) (IPCC6). We modeled distribution for present conditions future (2061-2080) time periods. Because one aim of our study was to determine the consequences of changing climate, we changed only the climatic data when projecting future distributions, while keeping the other variables constant over time (elevation, slope).
Species distribution modeling
We generated distribution maps for total occurrences (maternity + hibernacula + transition, hereafter defined as “combined models”), maternity colonies , hibernacula, and transition roosts. To estimate the present and future habitat suitability for C. townsendii in California, we used the maximum entropy (MaxEnt) algorithm in the “dismo” R package (Hijmans et al. 2021) through the advanced computing resources provided by Texas A&M High Performance Research Computing. We chose MaxEnt to aid in the comparisons of state-wide and ecoregion-specific models as MaxEnt outperforms other approaches when using small datasets (as is the case in our ecoregion-specific models). We created 1,000 background points from random points in the environmental layers and performed a 5-fold cross validation approach, which divided the occurrence records into training (80%) and testing (20%) datasets. We assessed the performance of our models by measuring the area under the receiver operating characteristic curve (AUC; Hanley & McNeil, 1982), where values >0.5 indicate that the model is performing better than random, values 0.5-0.7 indicating poor performance, 0.7-0.9 moderate performance and values of 0.9-1 excellent performance (BCCVL, Hallgren et al., 2016). We also measured the maximum true skill statistic (TSS; Allouche, Tsoar, & Kadmon, 2006) to assess model performance. The maxTSS ranges from -1 to +1:values <0.4 indicate a model that performs no better than random, 0.4-0.55 indicates poor performance, (0.55-0.7) moderate performance, (0.7-0.85) good performance, and values >0.80 indicate excellent performance (Samadi et al. 2022). Final distribution maps were generated using all occurrence records for each region (rather than the training/testing subset), and the models were projected onto present and future climate conditions. Additionally, because the climatic conditions of the different ecoregions of California vary widely, we generated separate models for each ecoregion in an attempt to capture potential local effects of climate change. A general rule in species distribution modeling is that the occurrence points should be 10 times the number of predictors included in the model, meaning that we would need 50 occurrences in each ecoregion. One common way to overcome this limitation is through the ensemble of small models (ESMs) (Breiner et al. 2015., 2018; Virtanen et al. 2018; Scherrer et al. 2019; Song et al. 2019) included in ecospat R package (references). For our ESMs we implemented MaxEnt modeling, and the final ensemble model was created by averaging individual bivariate models by weighted performance (AUC > 0.5). We also used null model significance testing with to evaluate the performance of our ESMs (Raes and Ter Steege 2007). To perform null model testing we compared AUC scores from 100 null models using randomly generated presence locations equal to the number used in the developed distribution model. All ecoregion models outperformed the null expectation (p<0.002).
Estimating range shifts
For each of the three GCMs and each RCP scenario, we converted the probability distribution map into a binary map (0=unsuitable, 1=suitable) using the threshold that maximizes sensitivity and specificity (Liu et al. 2016). To create the final maps for each SSP scenario, we summed the three binary GCM layers and took a consensus approach, meaning climatically suitable areas were pixels where at least two of the three models predicted species presence (Araújo and New 2007, Piccioli Cappelli et al. 2021). We combined the future binary maps (fmap) and the present binary maps (pmap) following the formula fmap x 2 + pmap (from Huang et al., 2017) to produce maps with values of 0 (areas not suitable), 1 (areas that are suitable in the present but not the future), 2 (areas that are not suitable in the present but suitable in the future), and 3 (areas currently suitable that will remain suitable) using the raster calculator function in QGIS. We then calculated the total area of suitability, area of maintenance, area of expansion, and area of contraction for each binary model using the “BIOMOD_RangeSize” function in R package “biomod2” (Thuiller et al. 2021).
Usage notes
We cannot provide original bat occurrence points used in analyses in the interest of protecting sensitive colony locations from disturbance or vandalism.