Books - Non-fiction

Where will species go? Incorporating new advances in climate modelling into projections of species distributions

Where will species go? Incorporating new advances in climate modelling into projections of species distributions
of 18
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Where will species go? Incorporating new advancesin climate modelling into projections ofspecies distributions LINDA J. BEAUMONT * , A. J. PITMAN w , MICHAEL POULSEN z and LESLEY HUGHES ** Department of Biological Sciences, Macquarie University, NSW 2109, Australia, w Department of Physical Geography, MacquarieUniversity, NSW 2109, Australia, z Department of Human Geography, Macquarie University, NSW 2109, Australia Abstract Bioclimatic models are the primary tools for simulating the impact of climate change onspecies distributions. Part of the uncertainty in the output of these models results fromuncertainty in projections of future climates. To account for this, studies often simulatespecies responses to climates predicted by more than one climate model and/or emissionscenario. One area of uncertainty, however, has remained unexplored: internal climatemodel variability. By running a single climate model multiple times, but each timeperturbing the initial state of the model slightly, different but equally valid realizationsof climate will be produced. In this paper, we identify how ongoing improvements inclimate models can be used to provide guidance for impacts studies. In doing so weprovide the first assessment of the extent to which this internal climate model variabilitygenerates uncertainty in projections of future species distributions, compared withvariability between climate models. We obtained data on 13 realizations from three climatemodels (three from CSIRO Mark2 v3.0, four from GISS AOM, and six from MIROC v3.2)for two time periods: current (1985–1995) and future (2025–2035). Initially, we compared thesimulated values for each climate variable ( P   , T  max  , T  min  , and T  mean ) for the current periodto observed climate data. This showed that climates simulated by realizations from thesame climate model were moresimilar toeach otherthanto realizations from othermodels.However, when projected into the future, these realizations followed different trajectoriesand the values of climate variables differed considerably within and among climatemodels. These had pronounced effects on the projected distributions of nine Australianbutterfly species when modelled using the BIOCLIM component of DIVA-GIS. Our results show that internal climate model variability can lead to substantial differencesin the extent to which the future distributions of species are projected to change. These canbe greater than differences resulting from between-climate model variability. Further,different conclusions regarding the vulnerability of species to climate change can bereached due to internal model variability. Clearly, several climate models, each representedby multiple realizations, are required if we are to adequately capture the range ofuncertainty associated with projecting species distributions in the future. Keywords: bioclimatic modelling, climate change, climate models, projected distributions, quantifyinguncertainty, realizations, within-model variability Received 9 May 2006; revised version received 29 September 2006 and accepted 18 December 2006 Introduction Climate change and projections of species distributions Future climate change will have substantial impacts onthe distribution of species. Indeed, the range margins of many species have apparently already responded to theanomalous warming of the past century by shiftingpoleward and/or upward in elevation (Hughes, 2000;Parmesan & Yohe, 2003; Root et al ., 2003). These changeshave been mirrored in projections derived from biocli-matic models (Walther et al ., 2005), which have shownthat under future climate regimes, species distributionsmay alter dramatically (e.g. Peterson et al ., 2002; Correspondence: Linda Beaumont, tel. 1 612 9850 8191,fax 1 612 9850 8245, e-mail: Global Change Biology (2007) 13, 1368–1385, doi: 10.1111/j.1365-2486.2007.01357.x r 2007 The Authors 1368 Journal compilation r 2007 Blackwell Publishing Ltd  Williams et al ., 2003; Beaumont et al ., 2005; Tuck et al .,2006). Bioclimatic modelling has become an increasinglypopular method for understanding the potential impactsof climate change on species distributions. The modelsoperate on the assumption that macroclimate playsa fundamental role in limiting species distributions(Pearson & Dawson, 2003). The ‘bioclimatic envelope’of a species (the range of values for certain climaticparameters to which the species is currently exposed) iseither calculated by correlating its known locations tocurrent climate (e.g. BIOCLIM, Houlder et al ., 2000) or is based on its physiological tolerances (e.g. CLIMEX,Sutherst et al ., 1998). The ‘envelope’ is projected ontogrids of current climate to identify the potential extentof the species current climatic range, and also ontoclimate change scenarios to predict distributionchanges. Bioclimatic models have been used to assesspotential climate-induced range shifts (Arau´ jo et al .,2006; Gritti et al ., 2006), to estimate extinction rates(Williams et al ., 2003; Thomas et al ., 2004), to examinethe efficacy of existing reserve systems (Rutherford et al ., 1999; Te´llez-Valde´s & Da´vila-Aranda, 2003; Arau´ jo et al ., 2004), and to identify priority areas for conserva-tion (Pyke et al ., 2005).Bioclimatic modelling is the most commonly used toolfor assessing the potential impacts of climate change onspecies, and for policy making in relation to speciesconservation (Hannah et al ., 2002; Harrison et al ., 2003).However, there are a number of challenges that requireinvestigation to improve the accuracy of models. Theseinclude improved techniques for sampling species loca-tionrecords,selectingpredictorvariables,andevaluatingmodels, andmethodstoincorporatemigrationand bioticinteraction (Guisan & Thuiller, 2005; Arau´ jo & Guisan,2006; Guisan et al ., 2006). The modelling process requiresa series of decisions to be made, each of which contri- butes to the level of uncertainty that cascades through tothe resulting projections of species distributions. Theprimary choices are:1. The bioclimatic model : there are a broad range of models to choose from (see Guisan & Zimmermann,2000), and generally models that allow the use of  both presence and absence data result in more accu-rate projections of the current potential distributionsof species than those that utilize presence-only data(e.g. see Ferrier & Watson, 1997; Elith et al ., 2006;Pearce & Boyce, 2006). The accuracy of differentmodels may also vary with the characteristicsof individual species (Thuiller, 2003; Segurado &Arau´ jo, 2004). Further, slight differences betweenthe current distributions of species derived fromdifferent bioclimatic models can be magnified whenprojecting distributions under future climates (Thuil-ler, 2003). This can result in significant differencesamong projections obtained from alternate biocli-matic models(Lawler et al ., 2006; Pearson et al ., 2006).2. Selection of climatic or geographic data for inclusion inmodels : the choice of predictor data can stronglyinfluence the outcome of models. For example,Heikkinen et al . (2006) assess the impact of ‘seasonalfine-tuning’ of climate variables on projections of thedistributions of migratory birds. Temperature andprecipitation variables for individual species weretailored to match arrival and departure dates inFinland. Projections derived from fine-tuned datawere statistically more accurate than those derivedfrom a ‘baseline’ set of variables for summer tem-perature and precipitation that were applied to allspecies. The number of predictor variables can alsoinfluence the size of predicted distributions. Usingthe bioclimatic model GARP, Peterson & Cohoon(1999) projected the distributions of North American birds using combinations of one to eight geographicdatasets. Maximum predictive accuracy was reachedafter the inclusion of five datasets, with annual meantemperature being a critical variable. Conversely, thecorrelative model BIOCLIM continually projectssmaller potential distributions with the inclusion of more variables (e.g. Beaumont et al ., 2005).3. Emission scenarios : The Intergovernmental Panel onClimate Change (IPCC) has developed a set of sce-narios that describe future greenhouse gas (GHG)emissions (Nakicenovic et al ., 2000). These scenariosuse ‘storylines’ that incorporate judgements on howpopulation growth, technological change, and eco-nomic growth may drive future emissions of GHGs.Each ‘storyline’ results in emission scenarios that are judged as equally likely (Nakicenovic et al ., 2000). Ithas been recommended that more than one emissionscenario be included in impacts assessments to quan-tify some of the uncertainty associated with esti-mates of future GHG emissions (Nakicenovic et al .,2000), and indeed a number of studies of potentialshifts in species bioclimatic ranges have done so (seeAppendix A). Differences in emission scenarios ac-count for approximately half of the 1.4–5.8 1 C rangein the projections of global warming due to aneffective doubling of GHGs (Houghton et al ., 2001).4. Climate model : Differences between predictions fromvarious climate models make identification of themost appropriate models difficult (Martı´nez-Meyer,2005). Climate models differ for a wide range of reasons (see McAvaney et al ., 2001). These include(a) technical issues such as spatial and vertical reso-lution (b) parameterization issues including howclouds, water vapour, ocean mixing, terrestrial pro-cesses, etc. are represented, and (c) feedback issues CLIMATE MODELS AND SPECIES DISTRIBUTIONS 1369 r 2007 The Authors Journal compilation r 2007 Blackwell Publishing Ltd, Global Change Biology , 13, 1368–1385  (models simulate a range of feedbacks relating towater vapour, clouds, snow, with differentstrengths). Differences between climate models ac-count for approximately half of the 1.4–5.8 1 C rangein the projection of global warming due to an effec-tive doubling of GHGs (Houghton et al ., 2001). Theoutput of climate models also varies due to differentdownscaling approaches (see Timbal, 2004). To ac-count for uncertainty among climate models someresearchers have projected species envelopes ontofuture climates derived from more than one climatemodel (see Appendix A).5. ‘  Idealized ’ scenarios : An alternate approach to usingfuture climates projected by climate models has beento use ‘idealized’ combinations of temperature andrainfall (i.e. a specific change in temperature, e.g. 1, 2,and 3 1 C, or rainfall, e.g. Æ 5%, 10%, and 20%) toassess the impact of possible changes in bioclimaticranges (see Appendix A). This approach provides amethod to identify thresholds where species may become threatened in an environment where appro-priate climate projections are not available. ‘Idea-lized’ scenarios are the method of choice in placeswhere climate model simulations compare poorlywith observations, such as in areas where the geo-graphy is spatiallyheterogeneous (mountain regions,islands, locations with strong gradients in climate,coastal areas, etc.) or when climate models cannot bedownscaled appropriately.It is inevitable, therefore, that any projection of a futurespecies distribution will have an associated level of uncertainty (see Dessai et al ., 2006). Identifying andquantifying this uncertainty is a necessary step towardsimproving the credibility of all projections. A numberof recent studies have focused on assessing the accuracyand reliability of different bioclimatic models (e.g. Thuil-ler, 2004; Arau´ jo et al ., 2005; Elith et al ., 2006; Lawler et al .,2006; Pearson et al ., 2006). These have shown that(a) there can be substantial differences in the accuracyof different bioclimatic models (e.g. Elith et al ., 2006)(b) changes in range size in response to climate changecan vary in both direction and magnitude due to biocli-matic model choice (Lawler et al ., 2006; Pearson et al .,2006), and (c) the size of the calibration vs. validationdatasets and the rules used to transform probabilities of occurrences into presence/absence can, along withmodel choice, also substantially affect the projection of species distributions (Arau´ jo et al ., 2005).Numerous studies have accounted for uncertaintyamong future climates by projecting species distribu-tions onto multiple climate scenarios (Appendix A).Few studies, however, have assessed variability infuture distributions arising from different bioclimaticmodels and different climate scenarios. Thuiller (2004)projected the potential distributions of 1350 Europeanplant species for 2050 using four bioclimatic models,climate data from two GCMs (HadCM3; A1, A2, B1, B2scenarios and CSIROMk2 A2 scenario) and two meth-ods to transform probabilities of occurrences into pre-sence/absence. He found that variability in projecteddistributions due to climate scenarios was less thanvariability arising from the use of different bioclimaticmodels. Similarly, more variation in the projected dis-tributions of European amphibians and reptiles wasattributed to different bioclimatic models than climatescenarios (Arau´ jo et al ., 2006).Recent advances in climate models mean that we arenow able to identify the impact of an additional sourceof uncertainty on projections of species future distribu-tions, that is, internal climate model variability. Thispaper provides a preliminary assessment of how inter-nal climate model variability translates into uncertaintyin projections of species distributions using the BIO-CLIM component of DIVA-GIS (Hijmans et al ., 2005a).  Internal-climate model variability and its potential impacton projections of species distributions Climate models are the basic tool for projecting thefuture climate (Houghton et al ., 2001). The models havedeveloped considerably over the last few decades(McGuffie & Henderson-Sellers, 2001) and now provide‘credible simulations of climate, at least down to sub-continental scales and over temporal scales from seaso-nal to decadal’ (McAvaney et al ., 2001). As part of theIPCC Fourth Assessment Report (AR4), a suite of com-mon climate model simulations was archived by theProgramme for Climate Model Diagnosis and Intercom-parison (PCMDI) (see A large number of specific ex-periments were performed using climate models thatincluded a coupled ocean. A list of the experimentsperformed, the experimental protocol and a listof output requirements are available at the IPCC Third Assessment Report (TAR) waswritten (Houghton et al ., 2001), the standard experi-mental protocol for climate modelling was to perform atransitory simulation, where the model was initializedat around 1850 and then run through to 2100. Tosimulate the impact of increasing GHGs two simula-tions were run. For the first run, atmospheric CO 2 wasincreasedto reflect the contribution made by each GHG,while the second run consisted of a single simulationwhere it was assumed that GHGs remained at a fixedlevel (preindustrial or present day). The difference 1370 L. J. BEAUMONT et al . r 2007 The Authors Journal compilation r 2007 Blackwell Publishing Ltd, Global Change Biology , 13, 1368–1385   between the simulated climates from these two runswas then used to assess the impact of increasing GHGs.Climate modellers have known for many years that if multiple simulations within a single climate model areperformed from 1850 with an identical emission scenar-io but the initial state of the model is perturbed by avery small amount, a different simulation results. Thedifference in the result is caused by climate variabilitygenerated internally within the climate model. Thisinternal variability (‘noise’) results from the complexityof the climate system and the nonlinear interactions between components whereby tiny initial differencesare amplified through time. The term ‘realization’ isused to describe each independent simulation thatdiffers only in initial conditions while ‘ensemble’ de-scribes the average output of multiple realizations. Theperturbation to the initial state can be achieved in manyways, but a typical method is to take the climate modelsimulation that has been used to equilibrate the ocean(typically several hundred years long) and initialize thenew experiment from the end of this equilibrationintegration. To obtain realization one ( 1 R), a simulationis initialized using, for example, 1 January 1860. Toobtain 2 R, a new simulation is initialized using 2 Jan-uary 1860, and so on. This tiny perturbation (1 day overseveral centuries) is sufficient to generate a measurabledifference in the climate simulation due to internalmodel variability. It is important to note that all of theserealizations are equally valid simulations of future climates .According to Cubasch et al . (2001) estimates of futureclimate change are likely to be more robust whenensembles are created. This is because multiple realiza-tions sample a higher proportion of the internal modelvariability than any single realization. In a sense, asingle realization is analogous to using a single sampleto estimate species richness of a particular area. A singlesample might grossly under/over estimate the actualrichness, and thus a more comprehensive samplingregime is desirable.At the time of the TAR, only the Canadian ClimateCentre’s model had been run for multiple realizationsfor SRES emission scenarios. In the AR4, however, theissues of emission scenarios and realizations have beensignificantly advanced. Multiple emission scenarioshave commonly been run (typically A2, B1, and A1B)with some models running multiple realizations. Thispresents new opportunities to explore levels of uncer-tainty that affect projections about future species dis-tributions. If internal climate model variability did notexist, one would assume that 1 R 5 2 R 5 3 R, etc, for agiven climate variable, where R is a separate realizationwithin a given climate model generated via a tinyperturbation to the initial condition. Thus, in exploringthe impact of climate change on species one would only be concerned with differences between climate models.However, as internal model variability does exist, 1 R 6¼ 2 R 6¼ 3 R. The question, therefore, is whether internalclimate model variability translates into large differ-ences in projections of species distributions comparedwith variation in species distributions derived fromdifferent climate models. Why is this important? Figure1 illustrates two situations. Both examples in this figureshow, for a given climate variable, a hypothetical rangeof variation due to the within-model variability of thedifferent climate models. This range of variation has been recorded by running a number of realizations foreach model, each time perturbing the model slightly asdescribed previously. In Fig. 1a, differences across thethree climate models (where M 5 model) are largecompared with the differences between realizationswithin a single model, that is, although 2 M is verysimilar to 3 M, the mean and variance of  1 M is substan-tially different. In this case, systematic errors among theclimate models dominate compared with internal varia- bility. This would mean that studies of climate impactsshould systematically and objectively assess differentclimate models. Conversely, the alternate situation oc-curs in Fig. 1b, where differences within a set of realiza-tions for one model ( 5 M) are large compared withdifferences between models ( 4 M and 6 M). In this situa-tion, internal climate model variability will pose agreater source of uncertainty for impacts studies than between-model variability.In this paper, we use the AR4 climate data to addresswhether the range of variability in projections of poten- M MMMM (a)(b)    C   h  a  n  g  e   i  n  c   l   i  m  a   t  e  v  a  r   i  a   b   l  e Fig. 1 Hypothetical representation of climate model variation.For each model (M) a series of realizations has been achieved fora climate variable. Black boxes represent the mean of the climatevariable while vertical lines represent Æ 1SD. Dotted linesindicate the extent of variability among the three models in bothfigures. (a) Situation in which systematic errors between climatemodels lead to greater between-model than within-model varia- bility. (b) Alternate situation where means of models are similar but within-model variability is greater than between-modelvariability. CLIMATE MODELS AND SPECIES DISTRIBUTIONS 1371 r 2007 The Authors Journal compilation r 2007 Blackwell Publishing Ltd, Global Change Biology , 13, 1368–1385  tial changes in the distributions of nine Australian butterfly species results primarily from internal climatemodel variability or from systematic biases between themodels. In doing so, we initially explore the extent towhich predictions of current and future climate differwithin and among climate models. We address thefollowing questions:1. For simulations of the current climate, is within-climate model variability less than between-climatemodel variability?2. Do patterns of variability within and between mod-els remain similar when climates are projected intothe future?We then explore the extent to which differences inprojections of future climate result in variability inprojections of species distributions, that is:3. To what extent does internal climate model variabil-ity matter? Is internal climate model variability largeenough to cause differences in projections of speciesdistributions that may be ecologically meaningful orhave significantly different implications for conser-vation management? Data and methodology Climate model data To explore differences in projections of species distribu-tions resulting from both internal- and between-climate model variation, the monthly mean AR4 datain the PCMDI archive ( were analysed. The climate vari-ables selected for inclusion in the bioclimatic model(BIOCLIM) were annual and seasonal maximum tem-perature ( T  max ), minimum temperature ( T  min ), meantemperature ( T  mean ), and total rainfall ( P ). We chosethe A1B emission scenario because more realizationswere available for more climate models for this scenariothan any of the other scenarios, thereby providing alarger sample size for analysis. The A1B scenario de-scribes a future world of very rapid economic growthand global population that peaks in mid-century anddeclines thereafter. It also assumes the rapid introduc-tion of new and more efficient technologies and a balance between fossil fuels and other energy sources.The emissions of CO 2 are relatively high (comparedwith other scenarios) through to 2020 and are mid-range by 2050. We averaged climate data over twoperiods. For the ‘present-day’ scenario we averageddata from 1985 to 1995 while for the ‘future’ scenariowe averaged from 2025 to 2035. Within this latter timeperiod, the A1B scenario is almost identical to the A1scenario and both are towards the upper end of therange of emission scenarios reported by Nakicenovic et al . (2000).Three climate models in the PCMDI data archivereport T  max , T  min , T  mean , and P for both time periodsand for multiple realizations. These are the AustralianCommonwealth Scientific and Research Organization(CSIRO) model (version Mark 3.0), the NASA/GoddardInstitute for Space Studies (GISS) model (AOM) and theModel for Interdisciplinary Research on Climate (MIR-OC version 3.2) co-developed by the Center for ClimateSystem Research, the National Institute for Environ-mental Studies and the Frontier Research Center forGlobal Change in Kanagawa, Japan. Brief technicaldetails of these three models are provided in Table 1.Although a number of realizations were used in thisstudy, we assumed that (a) the emission scenario usedfor forcing the climate model are realistic and (b) thereare no systematic errors common to the three models.Most climate impact studies calculate the change inclimate variables such as temperature and rainfall asthe difference between the quantity at some point in the21st century (here a decade centred on 2030, i.e. 2025–2035) minus a similar length period in the 20th century(here a decade centred on 1990, i.e. 1985–1995). Table 1shows the number of realizations for the decades 1985–1995 (R 20 ) and 2025–2035 (R A1B ) for each model. In total,we assessed 13 combinations of 20th century and 21stcentury climates. For example, climate data from thesingle realization for A1B from the CSIRO model can besubtracted from data for each of the three ‘present-day’realizations( 1 R 20 , 2 R 20 and 3 R 20 ) from the same model toproduce three different estimates of each climatic vari-able. Similarly, for the MIROC model, there are threerealizations for R 20 and two for R A1B , resulting in sixpermutations for each variable, while for GISS therewere two realizations each for both R 20 and R A1B resulting in four permutations for each variable. Current Climate Current climate data were obtained from WorldClim(version 1.4, Hijmans et al ., 2005b, WorldClim generates data layers through inter-polation of mean monthly climate data ( T  min , T  mean , T  max , total P , and 19 derived bioclimatic variables) fromweather stations onto a 30 arc-second resolution grid.Major climate databases used by WorldClim include theGlobal Historical Climatology Network, FAO, WMO,International Center for Tropical Agriculture, R-Hydro-net; elevation data were provided by the STRM eleva-tion database. WorldClim consists of long-term dataaveraged over 1950–1990. The interpolated layers weremade using the ANUSPLIN software (Hutchinson, 2004) 1372 L. J. BEAUMONT et al . r 2007 The Authors Journal compilation r 2007 Blackwell Publishing Ltd, Global Change Biology , 13, 1368–1385

Formal ReSpecT

Mar 10, 2018
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks