CHAPTER 28 Estimating the exposure of carnivorous plants to rapid climatic change

Matthew C. Fitzpatrick and Aaron M. Ellison

28.1 Introduction carnivorous plants distributions constrained by cli- mate; and second, how readily, if at all, might car- Forecasting how carnivorous plant species will nivorous plants disperse to colonize new habitat as respond to climatic change is a key issue in their it becomes climatically suitable? conservation and management (Chapter 27) but In this chapter, we estimate the vulnerability of presents a number of challenges. These challenges carnivorous plants to climatic change in light of derive from interactions between the relatively challenges identified with SDMs in general and their simplistic statistical methods typically used to fore- particular application to these unique species. The cast species responses to climatic change, which to modeling approaches we use partially overcome date have been limited mainly to species distribu- some of these challenges, and may be applicable to tion models (“SDMs;” Elith and Leathwick 2009, other sparse or rare species. We begin by reviewing ­Franklin 2009), and particular aspects of the ecology the basics of SDMs. We then highlight specific eco- of carnivorous plants, including their rarity, habi- logical characteristics of carnivorous plants and their tat specialization (Chapter 2), and limited dispersal geographic distributions that limit the utility of clas- ability (Chapter 22). sical SDMs for forecasting their future distributions. The small ranges and oftentimes low local We combine two approaches: “ensembles of small abundance of carnivorous plants provide few oc- models” (Breiner et al. 2015), which attempt to deal currence records, which increase the potential for with the challenges of fitting SDMs for data-limited poorly or over-fitted SDMs and misspecification species; and “bioclimatic velocity” (Serra‐Diaz of relationships with their “optimal” environ- et al. 2014), which is an estimate of how fast a species ments. The unique habitats in which carnivorous would have to migrate to track its climatic niche (as plants often grow (Chapter 2) also are difficult to opposed to a prediction of the potential shift in dis- characterize using the basic temperature and pre- tribution, the typical output from SDM projections), cipitation data that often undergird SDMs. Rather, to provide initial assessments of the vulnerability of habitats in which carnivorous plants are common carnivorous plants to climatic change. often are decoupled from broader climatic patterns (e.g., many retain high soil moisture even during seasonal drought) and may be associated with fre- 28.2 The basics of species distribution quent disturbance (e.g., fire; Chapter 2). Last, dis- models persal limitation also may constrain range shifts of carnivorous plants as the climate changes. These Efforts to quantify the vulnerability of species to three issues raise two related questions that are crit- climatic change typically rely on SDMs. These mod- ical for understanding and forecasting the future of els (also called bioclimatic envelope models, habi- carnivorous plants. First, to what extent are current tat suitability models, or ecological niche models;

Fitzpatrick, M. C., and Ellison, A. M., Estimating the exposure of carnivorous plants to rapid climatic change. In: Carnivorous Plants: Physiology, ecology, and evolution. Edited by Aaron M. Ellison and Lubomír Adamec: Oxford University Press (2018). © Oxford University Press. DOI: 10.1093/oso/9780198779841.003.0028 390 CARNIVOROUS PLANTS

Guisan and Zimmermann 2000, Elith and Leath- footing, especially when used to forecast species wick 2009, Franklin 2009) are relatively simple sta- distributions under scenarios of climatic change tistical models that predict habitat suitability across (Hampe 2004, Heikkinen et al. 2006, Dormann 2007). an area of interest using empirical relationships be- The primary critique centers on whether empirical tween the distribution of a species (expressed as a species–­climate relationships derived solely from set of point locations at which the species is known observations of point-occurrence records and asso- to occur) and coincident environmental variables ciated environmental conditions can reliably predict (typically derived from digital maps of interpolated responses of species to climatic change. Detractors climatic data). When applied to current climates or argue that multiple interacting abiotic and biotic simulations of future climate, SDMs respectively in- factors determine range limits (Gaston 2003), yet fer current species distributions or forecast potential SDMs typically use only a small number of climatic changes in species distributions from predictions of variables (most frequently temperature and precipi- habitat suitability (Guisan and Thuiller 2005). tation) in model fitting and often ignore other causal The field of species distribution modeling has abiotic drivers, biotic interactions (Wisz et al. 2012), grown rapidly over the last few decades (Guisan dispersal processes (Fitzpatrick et al. 2008), or adap- et al. 2013), fueled by three primary factors: in- tation (Fitzpatrick and Keller 2015). creased availability of species occurrence and Although carnivorous plants are found world- environmental datasets (Graham et al. 2004); the de- wide and several species have large geographic velopment of more powerful statistical techniques ranges, all of them occupy patchy, restricted mi- and user-friendly software packages (Phillips crohabitats within their geographic range (Juniper et al. 2006, Thuiller et al. 2009); and an overall need et al. 1989; Chapter 2). Local environmental con- for comprehensive information on species distribu- ditions including soils and hydrology likely are tions, including quantitative assessments of the vul- at least as important, if not more so, than climate nerability of species to climatic change (Chapter 27). per se in determining habitat suitability and occur- rence patterns. In terms of dispersal constraints, few species have any obvious mechanisms for long- 28.2.1 Challenging species distribution models distance seed dispersal. Moreover, the successful es- with sparse or rare species tablishment of some carnivorous plants far outside Despite advancements in data and algorithms, their native ranges suggests that distributions of modeling the impacts of climatic change on sparse some species are limited by their ability to disperse species (sensu Rabinowitz 1981), rare species, or and colonize suitable, but distant, habitats (e.g., El- habitat specialists like carnivorous plants using lison and Parker 2002; Chapter 22). These dispersal SDMs remains a major challenge. These challenges constraints suggest carnivorous plants may not be arise primarily from fitting models with insufficient able to track rapid climatic shifts by shifting their or unreliable point-occurrence records and envi- geographic ranges. ronmental predictors that inadequately character- Even if all habitat factors could be included in ize habitats. Whereas rare species may be under perfectly calibrated SDMs, fitted species–climate greatest threat from climatic change and thus might relationships still may not reflect true distribu- benefit most from SDM-based climate-impact as- tional constraints. There also is no reason to expect sessments, they are often are most difficult to model that current species–climate relationships will re- using SDMs, a conflict Lomba et al. (2010) summa- main constant in altered ecological contexts (e.g., rized as the “rare species modeling paradox.” ­Fitzpatrick et al. 2007, Veloz et al. 2012) or novel climates of the future (Williams and Jackson 2007, Fitzpatrick and Hargrove 2009). 28.2.2 Critiques of species distribution models Last, when applied to future scenarios of climatic Even when data are adequate and statistical is- change, SDMs forecast potential changes in species sues can be minimized, numerous critiques have distributions, not actual changes in where popula- argued that SDMs stand on weak theoretical tions occur on the landscape. The extent to which Estimating the exposure of carnivorous plants to rapid climatic change 391 species will be able to follow these forecasted range Besides limiting information content, low numbers shifts over the next several decades will depend on of point-occurrence records can bias fitted statisti- dispersal and population dynamics, both of which cal relationships that are extrapolated across the are uncertain and stochastic. Moreover, the increas- study area to map habitat current and future suit- ing isolation and fragmentation of natural habits ability (Barry and Elith 2006). Loss of populations and the rapid rates of projected climatic change because of habitat conversion or over-collecting likely will make rapid dispersal unfeasible for all (Chapter 27) only adds to the challenges of reliably but the most vagile and widespread species (Hill applying SDMs to carnivorous plants: not only are et al. 1999, Malcolm et al. 2002, Loarie et al. 2009). the number of occurrences reduced, but also rela- Pragmatic users of SDMs are quick to acknowl- tionships between their distributions and climate edge these, and other, shortcomings of the models, may be altered or obscured. but counter that SDMs are one of the only tools A small geographic distribution is not necessar- available that can be applied across multiple taxa, ily problematic; SDMs often perform best for nar- regions, times, and spatial scales (Guisan and Thu- rowly distributed species with specialized climate iller 2005). Proponents also contend that both the niches. However, performance of SDMs under such fossil record and contemporary observations pro- circumstances is high only when species are well vide evidence that, at broad spatial scales, species sampled (Guisan et al. 2007). Good predictive per- often predictably respond to climatic change by formance based on current distributions also may shifting their geographic ranges to track suitable be indicative of over-fitting, which is most likely in habitat (Davis 1989, Parmesan et al. 1999, Chen models fit with fewer than approximately ten oc- et al. 2011) and SDMs have been shown to predict currence records per predictor variable (Franklin such responses under modest degrees of extrapo- 2009). Over-fit models may perform poorly when lation in the distant (Maguire et al. 2016) and re- extrapolated to new climatic conditions (Maguire cent past (Araújo et al. 2005, Hijmans and Graham et al. 2016) and could lead to errors of omission 2006, Kearney et al. 2010, Dobrowski et al. 2011). (failure to predict suitable habitat where the species Although these suggestions may be true for cer- in reality can survive). tain taxa in certain regions (e.g., temperate trees), the same cannot necessarily be said of carnivorous 28.3.2 Habitat specialization plants or other sparse or rare species. Carnivorous plants are adapted to unique micro- 28.3 Characteristics of carnivorous habitats that exist as distinct, sometimes minute, patches on the landscape (Chapter 2). Identifying plants that challenge SDMs the environmental or habitat factors that most con- 28.3.1 Rarity and sparse distributions strain a species’ distribution is central to modeling it accurately; failure to include true and direct driv- Although carnivorous plants are found worldwide ers of habitat suitability may lead predictions of and several species are abundant or have large geo- suitable habitat where the species in reality cannot graphic ranges, all occupy patchy, restricted micro- survive (Guisan and Zimmermann 2000). For car- habitats within their geographic range (Chapter 2). nivorous plants, these limiting factors are relatively Typically, carnivorous plant species are rare sensu well understood (Chapters 2, 18), but environmen- Rabinowitz (1981): they tend to have small ranges, tal datasets describing microhabitats are rare and high habitat specificity, and some either have small such habitats are poorly represented by the spa- population sizes or are sparse within their range. tially and temporally homogenized climatic data Each of these components of rarity limits the abil- commonly used as predictors in SDMs. ity of SDMs to forecast reliably future distribu- Studies often include non-climatic predictors such tions of carnivorous plants (Sbragia et al. 2014). as land cover, substrate, geology, soils, topography, Small ranges reduce the number of locations (oc- or remotely sensed data when fitting SDMs (re- currence records) that can be used in fitting SDMs. viewed in Franklin 2009), but these can be properly 392 CARNIVOROUS PLANTS employed only if habitats containing carnivorous plants have been identified and such patches are as large as the resolution (pixel size) of gridded data, 20 and if the spatial precision and accuracy of the oc- 200 15 currence data are sufficiently high to ensure that point-occurrence records are assigned to the cor- 10 rect grid cells. Because occurrence data for many 100 5 species, including carnivorous plants, were col- lected before the advent of modern GPS technology, Number of species 0 many occurrence records have low precision and 25 50 75 100 accuracy. Thus, it is impossible to assign accurately 0 occurrence records to precise spatial locations. A re- 0 5000 15000 25000 35000 lated problem for historic occurrence records is that Number of records land cover may have changed and populations no longer exist where they once did. Many available Figure 28.1 Frequency distribution of occurrence records obtained for carnivorous plant species from regional and national herbaria predictor (climatic, habitat) variables are collected and global biodiversity databases. Inset expands the range of records at widely spaced geographic intervals. Subsequent between zero and 100. interpolation to fill intervening spaces between col- lection locations at best only poorly represents suit- ranges. For example, Dionaea muscipula is now es- able microhabitats. Although variables (e.g., slope, tablished in areas within the US states of Florida, soils, vegetation) might be combined to improve New Jersey, Pennsylvania, and Washington (Giblin characterization of particular habitats, the use of 2016). Sarracenia species have been introduced in numerous predictor variables only exacerbates the parts of Europe and New Zealand (Walker 2014). problem of over-fitting SDMs from a small num- Several species of Utricularia, as well as Aldrovanda ber of point-occurrence records. In general, SDMs vesiculosa, occur well outside of their native ranges should have a minimum of ten point-occurrence re- (Compton et al. 2012, Lamont et al. 2013), and have cords per predictor variable, plus another ten for the become species of management concern in areas model intercept (Franklin 2009, Feeley and Silman where they have been accidently or deliberately in- 2011). Based on our experience, the number of avail- troduced. The extent to which these introductions able occurrence records for most carnivorous plants are outside of the climatic limits implied by their seems to be well below this threshold (Figure 28.1). native distributions is not known; dispersal limita- tion for these and other carnivorous plants might be 28.3.3 Are carnivorous plant distributions determined by geography and history rather than constrained by climate? climate.

Just because an SDM performs well at fitting the distribution of a carnivorous plant species is not 28.4 Species distribution models evidence that its range is actually determined by for carnivorous plants and other rare climate. Besides strong habitat constraints, geo- species graphic distributions of carnivorous plants are constrained by biotic factors (e.g., Kesler et al. 2008; The need to develop reliable methods for modeling Paniw et al. 2015), including dispersal limitation the distributions of rare species and forecasting (Ellison and Parker 2002; Chapter 22), competition their response to ongoing and intensifying climatic (Chapter 2; but see Brewer 2003), pollination (Chap- change is gaining greater attention. We use two ter 22), and trophic mutualisms (Chapters 23–26). such approaches to assess the vulnerability of car- Of these, dispersal limitation may be the most nivorous plants to climatic change: ensembles of important, as many carnivorous plants have been small models (Breiner et al. 2015); and estimation of introduced successfully well beyond their native bioclimatic velocity (Serra‐Diaz et al. 2014). Estimating the exposure of carnivorous plants to rapid climatic change 393

28.4.1 Ensembles of small models improve their transferability to other time periods (Warren and Seifert 2011). Lomba et al. (2010) described a strategy for avoid- ing over-fitting when modeling rare species that Breiner et al. (2015) termed “ensembles of small 28.4.3 Estimating bioclimatic velocity models” (“ESM”). Instead of fitting one, potentially over-fit, SDM with numerous predictor variables, The vulnerability of species to climatic change can numerous bivariate models are fit, evaluated, and be assessed with three primary components: expo- then averaged to an ensemble weighted by model sure, sensitivity, and adaptive capacity (Dawson performance. Breiner et al. (2015) tested this method et al. 2011). Exposure refers to the rate and magni- on 107 species and found that ESM performed sig- tude of climatic change where a species currently nificantly better than standard SDMs, and that the occurs; exposure is positively related to vulnerabil- performance gains were greatest for rare species. ity. In contrast, sensitivity considers the extent to They concluded that an average of many simple which a species is dependent on prevailing climatic models avoids over-fitting while retaining explana- conditions; sensitivity also is positively related to tory power. vulnerability. Last, adaptive capacity refers to how well a species may persist in situ as the climate changes; adaptive capacity is negatively related to 28.4.2 Controlling complexity and over-fitting vulnerability. SDMs such as MaxEnt, whether used individually or in ensembles, assess only exposure In addition to using ESM, it can be beneficial to con- to climatic change. trol the complexity of the statistical model used to Because exposure includes the rate and mag- relate occurrence with climate. A statistical approach nitude of climatic change where the species now known as MaxEnt is one of the most popular tech- occurs, one method of estimating exposure is to niques for fitting SDMs. It is an implementation of ignore species entirely and instead simply estimate a statistical approach called maximum entropy that the magnitude of climatic change in that particu- characterizes probability distributions from incom- lar region. Loarie et al. (2009) estimated exposure plete information (Phillips et al. 2006). Modeling as being related to the velocity of climatic change distributions of species using MaxEnt assumes that (measured in km/year): the local rate of displace- occurrence data represent an incomplete sample of ment of climate expressed as the ratio of the rate of an empirical probability distribution; this unknown climatic change through time and the local rate of distribution can be estimated most appropriately climatic change across space. For temperature, cli- as the distribution with “maximum entropy:” the matic change velocity is: probability distribution that is most uniform, sub- ject to constraints imposed by environmental varia- °C year km bles; and that this distribution of maximum entropy = year C approximates the potential distribution of the spe- ° km cies (Phillips et al. 2006, Elith et al. 2011). Software for implementing MaxEnt models Biologically, the velocity of climatic change esti- (­Phillips et al. 2006, Hijmans et al. 2016) allows users mates the speed at which species must migrate to to control model complexity, but most studies rely track constant climatic conditions (in the context on the default settings instead of taking advantage of a single dimension of climate such as tempera- of this feature. Warren and Seifert (2011) demon- ture). All else being equal, larger velocities of cli- strated that optimizing model complexity can help matic change represent greater exposure because avoid models that are overly complex or simple the magnitude of a species’ response (in terms of and that exhibit poor inference regarding habitat rate and/or distance of migration) must be greater suitability. Controlling model complexity also can so that it can track its niche. Indeed, some studies help to identify correctly the relative importance of have suggested that the magnitudes or velocities variables in constraining species’ distributions and of past climatic changes are related positively to 394 CARNIVOROUS PLANTS rates of species extinction and evolution (Dynesius modeling. First, we used literature searches, books, and Jansson 2000, Nogués-Bravo et al. 2010, Sandel and online databases to update and standardize spe- et al. 2011), including the diversification ofDrosera in cies names to the latest taxonomy (Chapters 4–10) southwestern Australia (Yesson and Culham 2006). and to identify hybrids, which were removed from Although estimation of the velocity of climatic further consideration. Second, species records pro- change can be considered a “species-free” method vided without geographic (e.g., latitude–longitude) of estimating exposure because it does not require coordinates but with collection location information any data on species occurrences, several studies were georeferenced using GEOLocate 2.0 (Rios and have estimated species-specific velocities of cli- Bart 2010). Third, we took four additional steps to matic change. For example, Serra‐Diaz et al. (2014) ensure that only data of high spatial integrity were estimated the species-specific exposure of oak spe- used in modeling fitting: we removed occurrence re- cies to climatic change in California by calculating cords with low spatial precision, whose coordinates the velocity of change in habitat suitability from an were far outside the known range, whose coordi- SDM instead of the velocity of change of a single cli- nates matched the centroids of political boundaries, matic variable (see also Ordonez and Williams 2013, or whose coordinates fell within botanical gardens Carroll et al. 2015). Substituting habitat suitability or museums. Last, we removed all spatial replicates for climate allows the calculation of exposure to be (i.e., records for the same species with identical co- scaled specifically to the change in overall habitat ordinates). The remaining points were those that fell suitability for each species rather than to a single within the known native distribution of each spe- dimension of climate (e.g., temperature alone). As cies and which had sufficient spatial precision given such, the species-specific bioclimatic velocity ac- the resolution of the climate data. counts for the multidimensional nature of climatic change (i.e., multiple variables are changing at once 28.5.2 Climate data and at different rates) and appropriately weights these changes given the strength of their relationship We obtained gridded data layers for both current to the current distribution of a particular species. and potential future climatic conditions. To describe current climate (average for the period 1971–2000), 28.5 Modeling exposure of carnivorous we worked with the 19 bioclimatic variables from plants to climatic change WorldClim (http://www.worldclim.org; Hijmans et al. 2005), a database of globally contiguous, grid- We estimate the exposure of carnivorous plants to ded representations of climate developed from in- climatic change by fitting ESM for a large set of car- terpolations of observed data. Bioclimatic variables nivorous plants. We then project these models using include estimates of minima, maxima, and season- numerous scenarios of climatic change for 2050 and ality in temperature (°C) and precipitation (mm). calculate bioclimatic velocity for each species to es- We used data at 2.5 arc-minute spatial resolution. timate their potential exposure to climatic change. We reduced the full set of 19 variables by retaining those that had low multicollinearity (r < 0.7), but retained uncorrelated pairs of variables that were, 28.5.1 Species occurrence data in our opinion, most biologically informative. This We assembled a comprehensive database of occur- selection process reduced the 19 covariates to four rence records from multiple sources, including the variables that we used in model fitting and predic- Global Biodiversity Information Facility (GBIF), the tion: maximum temperature of the warmest month, Australia Virtual Herbarium, and numerous regional mean temperature of the coldest quarter, annual herbaria (see Acknowledgments for herbaria that precipitation, and precipitation seasonality. contributed data). We sought records for all currently For scenarios of future climates, we used decadal recognized carnivorous plant genera (Appendix). averages of the same four bioclimatic variables for We applied numerous quality control measures 2050 from 32 future climate simulations statistically to the occurrence records before using them for downscaled to 2.5 arc-minute spatial resolution. Estimating the exposure of carnivorous plants to rapid climatic change 395

These climate simulations allowed us to project our 28.5.4 Ensembles of small models (ESM) models for a range of possible future conditions. All possible combinations of bivariate MaxEnt mod- The simulations included output from numerous els (i.e., models that consider only two predictors at general circulation models and one representative a time out of the set of all possible predictors) were concentration pathway (RCP 8.5) developed as part fit for each species. Our set of four total climatic vari- of the IPCC AR5 and available from the Research ables resulted in six possible bivariate models, which Program on Climate Change, Agriculture and Food were then combined into a weighted ensemble based Security.1 on predictive performance (Breiner et al. 2015).

28.5.3 Species distribution modeling 28.5.5 Model projections, bioclimatic velocity, We modeled habitat suitability for carnivorous and exposure metrics plants using MaxEnt version 3.3.3E as implemented For MaxEnt models with a Boyce Index of at least in the dismo package (Hijmans et al. 2016) in R (R 0.4 (model appreciably better than random), we Core Team 2016). Except for optimizing model com- projected the models for each of the 32 global sce- plexity (§28.4.2; Warren and Seifert 2011), we fit narios of future climate for 2050. Using the ensem- models using the default values for all settings. We ble maps of current and future habitat suitability, fit MaxEnt models for any species with at least 10 we calculated the species-specific bioclimatic veloc- spatially unique occurrence records and constrained ity (Serra‐Diaz et al. 2014) as: selection of background data (points selected at ran- dom from the study region and also used to inform Δ HS model fitting) to those areas within any terrestrial year km = year ecoregion in which the species has been observed ΔHS km and immediately adjacent ecoregions, as defined us- ing the World Wildlife Fund’s definitions of terres- where ΔHS is the projected change in habitat suit- trial ecoregions (Olson et al. 2001). We partitioned ability between current and future climate from the occurrence data for each species randomly 10 MaxEnt (i.e., ΔHS/50, the number of years between times into calibration (70%) and evaluation (30%) 2000 and 2050). If the spatial gradient in climatic datasets, and models were run on each of the 10 suitability, ΔHS/km = 0, we removed that point to resulting datasets. The multiple models for each avoid an infinite bioclimatic velocity. species resulting from different random splits of We quantified the exposure of carnivorous plants the occurrence data into training and test partitions to climatic change using two metrics: bioclimatic were combined into a single ensemble by weighted velocity at each occurrence point for each modeled averaging (§28.5.4). To evaluate model perfor- species between current climate and 2050 (averaged mance, we used the continuous Boyce Index (Hirzel across the 32 future climate scenarios); and the per- et al. 2006), calculated using the ecospat package cent change in range area. (­Broennimann et al. 2016) in R (R Core Team 2016). Bioclimatic velocities can be either positive or neg- The Boyce Index varies from −1 to 1. Values greater ative, indicating the rate of loss or gain of climatically than zero indicate that predictions from the model suitable habitat through time respectively. Negative are consistent with the distribution of the species bioclimatic velocity indicates a decline in habitat (i.e., high habitat suitability predicted at known suitability where the species is known to occur: i.e., a presences), values near zero indicate a model no loss of climatically suitable habitat from which popu- better than random, and negative values indicate a lations must migrate or possibly face local extinction. model that predicts low habitat suitability at pres- Positive bioclimatic velocity requires more cautious ences and higher suitability in locations where the interpretation. Positive values at locations where the species is not known to occur (Hirzel et al. 2006). species has been observed represent habitats where populations may persist in the future, but such in- 1 http://www.ccafs-climate.org creases are not necessarily an improvement from a 396 CARNIVOROUS PLANTS vulnerability perspective. Projected increases in hab- sufficient data to fit models (number of spatially itat suitability can offset habitat losses only if indi- unique occurrence records ≥ 10). The majority viduals are able to disperse to and establish in newly of carnivorous plants had fewer than 70 occur- suitable habitats beyond the current distribution. rence records (Figure ­ 28.1) but a few had many To estimate the extent to which habitat losses thousands of records (e.g., Drosera rotundifolia, might be offset through increases in habitat suitabil- D. intermedia, and Pinguicula vulgaris). The high- ity, we calculated the percent change in range area as: est densities of occurrence records are in Western Europe, followed by eastern North America and − AAgainedlost Australia (Figure 28.2). North Africa, the Arabian Δ=Arange × 100, Apresent Peninsula, Russia, and China were lacking in oc- currence records for carnivorous plants, likely where A is the area of suitable habitat in 2050 gained due to both a lack of suitable habitat (e.g., north but not at present, A is the area of suitable habi- lost Africa, Arabian Peninsula) and a lack of reporting tat at present but absent in 2050, and A and is present (e.g., Russia, China). the predicted range area based on current climate.

To calculate ∆Arange, we converted the continuous habitat suitability values (0–1) from MaxEnt to suit- 28.6.2 Performance of species distribution able/unsuitable (0 or 1) using the threshold that models for carnivorous plants maximized model performance based on the true In general, SDMs were able to predict high suitabil- skill statistic (Allouche et al. 2006). Positive values ity habitat at known presences (as quantified using of ∆Arange indicate spatial expansion of climatically the weighted average Boyce Index across ESMs for suitable habitat that offer opportunities for estab- each species) for most carnivorous plants. Of the 223 lishment outside the current range should those species for which we could fit models, 180 (80.7%) locations provide other aspects of habitat that car- had a Boyce Index greater than 0.4 (­Table 28.1)—our nivorous plants require. Negative values indicate selected cutoff value for a model to be considered a net loss of suitable range area. Based on these of high enough quality to be projected to future metrics, species with greatest exposure to climate scenarios. Seventy-five carnivorous plants (33.6%) change are those with both bioclimatic velocities had a Boyce Index of at least 0.75, whereas 18 (8.1%) and ∆Arange less than zero. carnivorous plants had a Boyce Index less than 0, indicating a model worse than random. 28.6 Results

28.6.1 Occurrence data for carnivorous plants 28.6.3 Vulnerability of carnivorous plants to climatic change We obtained more than 160,000 occurrence re- cords for 296 carnivorous plants (Table 28.1; The majority of modeled carnivorous plants ­Figure 28.1). Of these, 223 carnivorous plants had (68.3%) had a negative median bioclimatic velocity (­­Figure 28.3), indicating that most current locations

Figure 28.2 Spatial distribution and 500 density of occurrence records obtained for 50 carnivorous plant species from regional and 10 national herbaria and global biodiversity 1 databases. Estimating the exposure of carnivorous plants to rapid climatic change 397

Table 28.1 Summary of occurrence data for carnivorous plant species obtained from regional and national herbaria and global biodiversity databases. The number of records is the total number of spatially unique records for each genus at 2.5 arc-minute resolution remaining after quality control (§28.5.1). The number of modeled species indicates the number of species with at least ten occurrence records and the number of projected species is the number of modeled species with an ensemble weighted average Boyce Index of at least 0.4.

Family Genus Number of Species Number of Records Number of Modeled/ with Data Projected Species

Bromeliaceae Brocchinia 2 104 2/0 Catopsis 1 143 1/0 Byblidaceae Byblis 3 791 3/3 Cephalotaceae Cephalotus 1 97 1/1 Dioncophyllaceae Triphyophyllum 1 131 1/1 Droseraceae Aldrovanda 1 370 1/1 Dionaea 1 503 1/1 Drosera 104 72,064 73/63 Drosophyllaceae Drosophyllum 1 256 1/1 Lentibulariaceae Genlisea 6 170 4/2 Pinguicula 33 34,663 22/19 Utricularia 124 52,099 96/74 Martyniaceae Ibicella 1 139 1/1 Nepenthaceae Nepenthes 5 388 5/2 Sarraceniaceae Darlingtonia 1 758 1/1 Heliamphora 3 130 3/2 Sarracenia 8 4,989 8/8

TOTAL 296 167,795 223/180 where carnivorous plants have been observed are spatial distribution, magnitude, and nature of cli- projected to decrease in climatic suitability through matic change relative to the current distribution of time; declines exceed projected increases in habit the species (Figure 28.4). suitability. Median bioclimatic velocities ranged Bioclimatic velocities also varied by genus and from a minimum of −4.62 km year−1 for Utricularia latitude. When averaged within genera, only Byb- olivacea to a maximum of 4.80 km year−1 for U. flori- lis, Nepenthes, and Sarracenia were projected to have dana, with bioclimatic velocity depending on the more increases than losses in habitat suitability in 398 CARNIVOROUS PLANTS

(a) Drosera linearis capillaris brevifolia petiolaris ordensis filiformis lanata peltata banksii dilatatopetiolaris brevicornis paradoxa sessilifolia derbyensis roraimae arcturi anglica aliciae stenopetala capensis nitidula binata natalensis fulva dielsiana barbigera pygmaea burkeana spatulata stricticaulis androsacea neesii platystigma

Species whittakeri leucoblasta intermedia parvula erythrogyne andersoniana huegelii paleacea ramellosa indica pulchella macrantha bulbosa erythrorhiza heterophylla fimbriata subhirtella glanduligera scorpioides gigantea pallida stolonifera collinsiae kaieteurensis subtilis spilos intricata pycnoblasta dichrosepala platypoda –5 05 Velocity (km/year)

Figure 28.3 The bioclimatic velocity between current climate and year 2050 for (a) Drosera species, (b) Utricularia species, and (c) all other projected carnivorous plant species. Points indicate the median velocity across all occurrence locations and 32 future climate scenarios and lines indicate the 95% confidence interval. Estimating the exposure of carnivorous plants to rapid climatic change 399

(b) Utricularia floridana resupinata inflata cornuta chrysantha odorata muelleri radiata limosa purpurea minutissima quinquedentata involvens juncea fulva leptoplectra striata lasiocaulis arnhemica tenuicaulis hydrocarpa leptorhyncha capilliflora caerulea circumvoluta kimberleyensis macrorhiza praetermissa trichophylla bifida triflora unifolia jamesoniana uniflora uliginosa inflexa stygia ochroleuca foliosa Species lateriflora guyanensis livida dichotoma amethystina intermedia minor hispida pubescens viscosa longeciliata multifida violacea flaccida tenella kamienskii vulgaris prehensilis subulata singeriana inaequalis beaugleholei fistulosa simulans dunlopii tubulata geminiscapa arenaria welwitschii reflexa andongensis praelonga bremii australis olivacea –5 05 Velocity (km/year)

Figure 28.3 (continued) 400 CARNIVOROUS PLANTS

(c) Other Sarracenia minor Pinguicula pumila Sarracenia rubra Byblis filifolia Byblis liniflora Sarracenia flava Sarracenia alata Pinguicula primuliflora Nepenthes mirabilis Pinguicula ionantha Sarracenia oreophila Sarracenia psittacina Pinguicula caerulea Pinguicula planifolia Pinguicula lutea Sarracenia leucophylla Darlingtonia californica Pinguicula vallisneriifolia Pinguicula macroceras Sarracenia purpurea Pinguicula nevadensis Pinguicula moranensis Pinguicula alpina Species Heliamphora tatei Pinguicula grandiflora Pinguicula vulgaris Pinguicula oblongiloba Pinguicula calyptrata Heliamphora minor Pinguicula corsica Drosophyllum lusitanicum Aldrovanda vesiculosa Nepenthes madagascariensis Byblis gigantea Pinguicula villosa Cephalotus follicularis Pinguicula lusitanica Pinguicula longifolia Genlisea africana Genlisea aurea Dionaea muscipula Triphyophyllum peltatum lbicella lutea –5 05 Velocity (km/year)

Figure 28.3 (continued) Estimating the exposure of carnivorous plants to rapid climatic change 401

(a) (b)

60.0 0.0 Figure 28.4 (Plate 19 on page P16) 45.0 Maps of (a, c) positive and (b, d) negative −1 30.0 –1.8 bioclimatic velocity (km yr ) between current climate and year 2050 for (top 15.0 row) Sarracenia flava and (bottom row) Cephalotus follicularis averaged across 0.0 –3.6 32 future climate scenarios. Black circles 170.0 0.0 represent occurrence locations used in 127.5 model fitting. ForS. flava, most locations where populations have been observed 85.0 –0.4 are projected to experience increases in 42.5 habitat suitability and low velocity declines –0.9 only at southernmost locations. In contrast, 0.0 C. follicularis is projected to experience relatively low velocity decreases in suitability across all locations where populations have (c) (d) been observed.

5

0 locity (km/year )

Ve –5

Figure 28.5 Genus-level boxplots of the bioclimatic velocity (km yr−1) between current climate and year 2050 across 32 –10 future climate scenarios. Values less than zero indicate declines in habitat suitability, a ia while values greater than zero indicate llum cen Byblis y uicula a Drosera Ibicella increases in habitat suitability at locations Dionaea Genlisea epenthes tricularia liamphora N Ping Sarr U Aldrovanda CephalotusDarlingtoni where populations of each genus have been Drosoph He Triphyophyllum observed. For clarity, outliers are not shown. locations where carnivorous plants in these genera near even mix of positive and negative bioclimatic have been observed, implying opportunities for velocities. The northern hemisphere exhibited a persistence in current habitats (Figure 28.5). When near equal distribution of positive and negative viewed as a function of latitude (Figure 28.6), bio- bioclimatic velocities, likely reflecting temperature- climatic velocities tended to be mainly negative in driven declines in habitat suitability in the southern the southern hemisphere, with the exception of portion of carnivorous plants’ ranges and increases between ≈10° and 25° south latitude, a range that in climatic suitability in the northern portion of their includes northern South America, south central ranges. All latitudinal bands between 50° south and Africa, and northern Australia, where there was a 70° north contained at least one occurrence record of 402 CARNIVOROUS PLANTS

50

25

0 Velocity (km/year) Figure 28.6 The bioclimatic velocity (km yr−1) between current climate and year 2050 –25 by latitude. Each point indicates the average bioclimatic velocity across 32 future climate –50 –25 0725 50 5 scenarios at that latitude. Gray shading Latitude indicates the density of occurrence records. a carnivorous species, but the extent to which these distribution of change in habitat suitability. Be- results reflect spatial sampling bias (Figure 28.2) is cause we quantified vulnerability mainly in the unclear. context of bioclimatic velocity at locations where Because bioclimatic velocity tends to be nega- carnivorous plants have been observed, we con- tive for most species in locations where carnivo- sciously discounted the potential for establish- rous plants occur at present, a majority of species ment in areas outside the current distribution. (57.8%) also were projected to experience overall This approach seems reasonable given most car- declines in range area, even after factoring in any nivorous plants are unlikely to track rapid changes potential gains (Figure 28.7). Thirty-nine carnivo- in climate over the next few decades. Nonetheless, rous plants were projected to undergo > 50% re- expansion of climatically suitable habitat does ductions in range area and 13 were projected to represent potential opportunities for range expan- experience ≥ 80% declines in range area. The model sion, which we quantified using percent increase projections also suggested potential opportunities in range area (Figure 28.7). for range expansion under future climate, should For > 65% of the species we considered, cli- carnivorous plants be able to disperse to, and estab- matic change would lead to substantial declines lish in, those regions that are projected to increase in habitat suitability at locations where carnivo- in climatic suitability. rous plants have been observed. These projected declines suggest that many populations could be 28.7 Discussion under threat of local extinction because of climatic change. If these projections are correct, then the Our goal was to provide a first-pass estimate of large percent reductions in range area for many the vulnerability of carnivorous plants to climatic carnivorous plants might increase the number of change, given their unique ecology, and the sta- at-risk species and perhaps increase the rank of cli- tistical and other limitations of SDMs when fit- matic change as a threat to their conservation (Jen- ting models for small-ranged, often rare, habitat nings and Rohr 2011). specialists. We used ESM (Breiner et al. 2015) to Because our models considered only climate- overcome the statistical challenges of fitting SDMs related aspects of habitat, the extent to which the for species with few occurrence records and biocli- SDM-based projected reductions in habitat suitabil- matic velocity (Serra‐Diaz et al. 2014) as a means ity may translate into population declines depends to provide an integrated, species-specific metric on the amount of coupling between microhabi- of exposure to climatic change that simultane- tats and the regional climate. The extent to which ously considers the magnitude, rate, and spatial habitat losses may be offset by the emergence of Estimating the exposure of carnivorous plants to rapid climatic change 403

(a) Drosera peltata dilatatopetiolaris paradoxa ordensis brevicornis banksii fulva linearis capillaris sessilifolia derbyensis lanata petiolaris filiformis brevifolia nitidula intermedia anglica aliciae androsacea stenopetala binata indica arcturi neesii pygmaea andersoniana pulchella stricticaulis spatulata whittakeri spilos parvula

Species platystigma barbigera macrantha dichrosepala glanduligera leucoblasta erythrorhiza scorpioides paleacea ramellosa bulbosa erythrogyne subhirtella huegelii heterophylla gigantea intricata pallida collinsiae stolonifera pycnoblasta natalensis platypoda burkeana dielsiana capensis roraimae subtilis fimbriata kaieteurensis –100 0 100 200 300 Percent change in potential range area

Figure 28.7 Percent change in range area between current climate and year 2050 for (a) Drosera species, (b) Utricularia species, and (c) all other projected carnivorous plant species. Points indicate the median projected percent change in range area across all occurrence locations and 32 future climate scenarios and lines indicate 95% confidence intervals. 404 CARNIVOROUS PLANTS

(b) Utricularia tenuicaulis fulva floridana capilliflora leptoplectra involvens circumvoluta arnhemica singeriana triflora quinquedentata minutissima odorata lasiocaulis muelleri inflata kimberleyensis leptorhyncha chrysantha limosa striata radiata hydrocarpa resupinata purpurea macrorhiza fistulosa cornuta juncea australis dunlopii caerulea bifida vulgaris ochroleuca foliosa uliginosa olivacea intermedia Species minor longeciliata inflexa geminiscapa subulata flaccida lateriflora stygia dichotoma violacea pubescens praelonga trichophylla unifolia uniflora tenella beaugleholei simulans arenaria guyanensis multifida reflexa hispida inaequalis amethystina jamesoniana tubulata andongensis praetermissa livida bremii prehensilis viscosa kamienskii welwitschii –100 0 100 200 300 Percent change in potential range area

Figure 28.7 (continued) Estimating the exposure of carnivorous plants to rapid climatic change 405

(c) Other Sarracenia alata Pinguicula primuliflora Sarracenia minor Sarracenia psittacina Sarracenia leucophylla Pinguicula caerulea Pinguicula planifolia Sarracenia flava Pinguicula macroceras Pinguicula ionantha Pinguicula pumila Pinguicula lutea Byblis filifolia Sarracenia rubra Byblis liniflora Sarracenia oreophila Sarracenia purpurea Darlingtonia californica Byblis gigantea Drosophyllum lusitanicum Nepenthes mirabilis Pinguicula vallisneriifolia

Species Aldrovanda vesiculosa Pinguicula grandiflora lbicella lutea Heliamphora tatei Pinguicula villosa Dionaea muscipula Pinguicula lusitanica Pinguicula vulgaris Pinguicula alpina Pinguicula moranensis Cephalotus follicularis Pinguicula calyptrata Pinguicula oblongiloba Triphyophyllum peltatum Pinguicula nevadensis Nepenthes madagascariensis Heliamphora minor Genlisea aurea Genlisea africana Pinguicula corsica Pinguicula longifolia –100 0 100 200 300 Percent change in potential range area

Figure 28.7 (continued) 406 CARNIVOROUS PLANTS climatically suitable habitat beyond the current dis- may become less suitable for their persistence in tribution will depend strongly not only on the abil- the future. Given other, arguably more immedi- ity of individuals to reach these habitats, but also ate, conservation and management challenges fac- on whether unoccupied areas also provide unique ing carnivorous plants (Jennings and Rohr 2011; habitat properties, including soil characteristics, ­Chapter 27), climatic change may be best viewed hydrology, and disturbance regimes, none of which as an additional stressor. Managing climate itself is were included in our modeling framework. Finally, not a near-term solution, so a good strategy for con- we used only a single algorithm to fit SDMs. While serving carnivorous plants as the climate changes MaxEnt is regarded as one of the more robust al- would be to decrease other stressors on populations gorithms when dealing with small sample sizes to the greatest extent possible, thereby potentially (Hernandez et al. 2006; Wisz et al. 2008), it is good increasing resilience to climatic change. Actions practice to consider multiple algorithms. Algorithm should focus on immediate threats: loss of popula- type represents one of the primary sources of uncer- tions due to habitat conversion; poaching; native tainty in future projections of species distributions and non-native invasive species that compete with (Diniz-Filho et al. 2009). carnivorous plants; hydrological regimes; and natu- Fifty-seven of the carnivorous plants modeled in ral disturbance such as fire that no longer occur with this study had projected increases in habitat suit- some degree of regularity. A more controversial ap- ability at occurrence locations, mainly poleward proach would be to identify rear-edge populations of 25° north latitude; what do these increases actu- that may contain unique genotypes pre-adapted ally mean in the context of vulnerability to climatic to future climate (Hampe and Petit 2005) and con- change? They could suggest either that current sider these populations for transplant to new areas climatically marginal habitats will improve in the (Vitt et al. 2010) projected to become suitable in the future or that current high-quality microhabitats future. On the other hand, because the habitats in are not identified as such by broad-brush climatic which carnivorous plants occur are naturally frag- variables. Alternatively, they could be indicative mented, there may be little to be gained by increas- of populations adapted to local conditions beyond ing connectivity. the mean climatic conditions experienced by the Different carnivorous plants often co-occur in the species as a whole. We are unable to distinguish same habitat or have similar habitat requirements, among these alternatives because of challenges so one of the most robust strategies for conservation with modeling habitat specialists such as carnivo- would be to protect habitats that currently support rous plants and the shortcomings of SDMs more carnivorous plants. Such a strategy might allow for generally for assessing the exposure of species to the establishment of future range-shifting species climatic change. with similar habitat requirements. For example, in Beyond direct effects, climatic change also could eastern North America, many of the habitats that have indirect effects on carnivorous plants. For currently support cold-tolerant species such as Sar- example, changes in habitat suitability could alter racenia purpurea and Drosera rotundifolia likely could interactions with co-occurring species, leading to support other carnivorous plants as winter tem- declines in population growth rates (Nordbakken peratures continue to warm rapidly. Some habitats et al. 2004). In the southeastern United States, some already may be suitable for more southerly species. of the most diverse carnivorous plant habitats occur For example, Dionaea muscipula already is estab- in low-lying coastal areas (Chapter 2). Carnivorous lished in the New Jersey Pine Barrens, well north of plants endemic to such regions may be particularly its native distribution (Chapters 4, 27). That numer- susceptible to sea level rise or storm surges (Abbott ous carnivorous plants are established well outside and Battaglia 2015) associated with climate change. their native ranges suggests that these species may Overall, our results suggest that the climatic con- be more dispersal limited than constrained by nar- ditions where many carnivorous plants grow now row climatic tolerances. Estimating the exposure of carnivorous plants to rapid climatic change 407

28.8 Future research distributions of carnivorous plants, such data would be useful only for vulnerability assessments if we How might we improve our ability to estimate the also understand relationships between climate, vulnerability of carnivorous plants and other rare hydrology, disturbance regimes, and soil charac- plant species to climatic change? One strategy, teristics in carnivorous plant habitats. Last, a better which would require little in the way of additional understanding of the climatic tolerances of different data or modeling techniques, would be to use com- carnivorous plants and their ability to adapt to new mon plants as indicators for carnivorous plants climatic regimes, perhaps through the coupling of (Smart et al. 2015) in a community-level modeling emerging genomic techniques (Chapter 11) and spa- framework (Maguire et al. 2016). Although envi- tial modeling (Fitzpatrick and Keller 2015), would ronmental data that characterize unique microhabi- help assess whether we can expect these species to tats could help improve our ability to map current tolerate or adapt to climatic change in situ.