Trait-based climate change predictions of plant community structure in arid steppes

Authors


Correspondence author. E-mail: cedric.frenette.dussault@usherbrooke.ca

Summary

  1. Global climate change is possibly one of the most important challenges for current and future human populations due to its wide-ranging effects on ecosystems. Global prediction models suggest that in some areas of the world (e.g. Northern Africa, Central America) an increase in aridity might strongly disturb agricultural production and affect food security. To counterbalance these negative effects, reliable predictive models are needed to anticipate ecosystem changes.
  2. We tested the ability of the Community Assembly by Trait-based Selection (CATS) model that is based on the principle of maximum entropy and trait-based environmental filtering, to predict actual and future plant community composition in the arid steppes of eastern Morocco. Specifically, we asked whether this model was adequate for predicting actual community composition, based on predicted community-weighted mean (CWM) traits, and what would be the changes in community composition under various scenarios of climate change for the period 2080—2099.
  3. The CATS model could predict > 90% of actual community composition if the actual CWM traits were known but only ˜40% if the CWM values were predicted from estimated aridity and grazing values. The predictions of community composition for 2080—2099 suggested that, regardless of the climate change scenario considered, the dominant group in grazed and ungrazed sites would shift from ruderal species to stress-tolerant sub-shrubs, which would constitute up over 80% of total community composition in some cases.
  4. Synthesis. Our findings suggest that effects of climate change will strongly modify plant community structure in arid steppes, possibly accentuating the process of desertification, and reducing the pastoral value of the vegetation. Future research efforts should concentrate on the identification of strong trait–environment relationships to improve model predictions.

Introduction

Global climate change is possibly one of the most important challenges for current and future human populations due to its wide-ranging effects on ecosystems. Predicted changes in mean annual temperature, precipitation events and atmospheric CO2 concentrations are expected to influence ecosystem functions, community structure and species interactions (e.g. Hooper & Vitousek 1997; Kardol et al. 2010; Yang et al. 2011; Brodie et al. 2012). Global prediction models suggest that in some areas of the world (e.g. Northern Africa, Central America), an increase in temperature that is coupled with a decrease in precipitation might strongly disturb agricultural production (IPCC 2007a). In such areas where local populations are already coping with harsh socio-economic conditions, climate change is a potential threat to food security (Parry et al. 2004). To counterbalance these negative effects, reliable predictive models are needed to anticipate ecosystem changes. A better understanding of how plant communities might react to climate change is necessary to elaborate sensible natural resource and land management policies, which will help to preserve ecosystem integrity and ensure population well-being.

To understand how changes in climatic conditions might affect plant community structure and ecosystem function, one can (i) experimentally manipulate climatic variables to simulate future environmental conditions on artificial communities and measure ecological variables of interest (e.g. productivity, diversity, soil nutrients) (e.g. Kardol et al. 2010; Miranda et al. 2011) or (ii) use plant species abundance data from observational studies along environmental gradients and extrapolate to future environmental conditions that are predicted by climate change scenarios (Pompe et al. 2010; Bertrand et al. 2011). Both approaches have their advantages and should be used concurrently whenever possible. In the second case, researchers need to carefully think about what kinds of models are appropriate to make sensible predictions. We contribute to the second approach using a statistical model (Community Assembly by Trait-based Selection, CATS), based on the principle of maximum entropy of Jaynes (1957a,b), to predict how changes in temperature and precipitation might affect the structure of plant communities in the eastern Moroccan steppes. These arid ecosystems are strongly structured by water availability and grazing pressure (Frenette-Dussault et al. 2012). According to the Intergovernmental Panel on Climate Change (IPCC 2007b), eastern Morocco should suffer from decreased precipitation and increased temperature in the next decades.

The CATS model relies on functional trait data to make predictions about species relative abundances (Shipley 2009). It is the direct mathematical translation of ‘trait-based habitat filtering’ into a quantitative framework (Shipley 2009). Both biotic and abiotic conditions act as selection filters on vegetation through interactions with species functional traits (Keddy 1992). Depending on their trait values, some species will persist, grow and reproduce in a given environment better than others, giving rise to specific species assemblages.

The CATS model has shown potential to become a powerful predictive tool (Sonnier, Shipley & Navas 2010; Laughlin et al. 2011; Merow, Latimer & Silander 2011; Shipley et al. 2011; Laliberté et al. 2012; Shipley, Paine & Baraloto 2012). The model was used to predict changes in herbaceous (Shipley et al. 2011) and forest communities (Laughlin et al. 2011) along an altitudinal gradient in semi-arid and grazed conditions, but the model has yet to be applied to the vegetation of extremely arid and heavily grazed regions where local populations directly depend on such ecosystems for subsistence.

The primary objective of this article was to explore the capacity of the CATS model to make predictions of species abundances under both current and future environmental conditions in arid steppes. More specifically, we addressed the following two questions: (i) how well can the CATS model predict actual species assemblages in ecosystems driven by gradients of both aridity and grazing, given community-weighted traits predicted by an established abiotic aridity index? (ii) Under various predicted scenarios of increased aridity, how might this affect plant community structure? This modelling exercise is important because even though qualitative predictions of desertification of Northern Africa have been suggested (Le Houérou 1995), there is still a need for quantitative predictions of community structure given different quantitative climate change scenarios. Equally importantly, changes in climate potentially interact with changes in grazing intensity in our system, and so it is important to include both factors in any quantitative predictions. This is difficult because even qualitative predictions are not straightforward under such a scenario. Further increases in aridity in such already dry ecosystems should select for more ‘stress-tolerant’ species according to Grime (2001), but increased grazing could select either for species more tolerant of grazing (presumably more ‘stress-tolerant’) or for more ‘ruderal’ species able to escape grazing events by timing their germination and growth to coincide with unpredictable rainfall events (Frenette-Dussault et al. 2012). Of course, such qualitative predictions, even when possible, cannot predict by how much each species will respond. Thus, the use of a quantitative prediction model could help to quantify the relative importance of aridity and grazing under a climate change scenario.

Materials and methods

Study Sites

We carried out this study in the arid steppes of the provinces of Boulemane and Figuig, eastern Morocco. The climate in this area has been described as Mediterranean arid, with cold winters and hot dry summers (Le Houérou 1995); average yearly minimum and maximum temperatures are about 10 °C and 25 °C, respectively, and annual precipitation range lies between 150 and 250 mm with high inter-annual variability (Frenette-Dussault et al. 2012). These steppes, which are collectively owned, have become degraded due to overgrazing by sheep and goats, the harvesting of woody species and the absence of fallow periods (Ait 1996; Msika et al. 1997). Land use for grazing is year-round, and average stocking rates are estimated to be around 1.6 animal unit ha−1 year−1 (Laouina et al. 2001). The steppes can provide as much as 75% of herd fodder requirements during wet years, but this amount can drop to 30% during dry years (Lazarev 2008).

We sampled a total of 50 20 × 20 m2 vegetation plots in five sites (10 plots per site) differing in aridity and in grazed and ungrazed conditions (Table 1). For each site, we calculated an aridity index as the ratio of potential evapotranspiration to precipitation (Frenette-Dussault et al. 2012). The intensity of grazing by sheep and goats was divided into two classes: intense grazing or no grazing. The ungrazed plots were located inside permanent exclosures maintained by the Emirates Center for Wildlife Propagation (Missour, Lamjalil, Enjil and Maatarka sites) and the ‘Haut Commissariat aux Eaux et Forêts et à la Lutte Contre la Désertification’ (Tirnest site, Table 1). We assessed species abundances based on plant cover at each of the 50 plots using the point-intercept method (Daget & Poissonnet 1971). At each site, we sampled five grazed plots and five ungrazed plots at the peak of standing biomass (April–May 2010). The plots were separated by at least 800 m in grazed conditions and 500 m in ungrazed conditions.

Table 1. Environmental variables characterizing the five study sites
Environmental variablesSites
Missour 33.006° N 4.102° WLamjalil 32.894° N 3.957° WTirnest 33.414° N 3.788° WMaatarka 33.257° N 2.724° WEnjil 33.160° N 4.586° W
  1. a

    PET was computed with the climate information tool (CIT).

  2. (http://www.fao.org/nr/water/aquastat/gis/index3.stm). See Frenette-Dussault et al. (2012) for additional details.

  3. b

    The Lamjalil site was not fenced at the time of sampling but has been during second half of the 20th century. See Frenette-Dussault et al. (2012) for additional details concerning its grazing history.

Annual precipitation (P) (mm)164164190183264
Potential evapotranspiration (PET)a (mm)1546.41546.01512.61492.91406.4
Elevation (m)960100098013001600
Mean maximum annual temperature ( °C)24.123.124.422.620.1
Mean minimal annual temperature ( °C)10.79.711.29.36.6
Peak of standing biomassAprilAprilAprilApril/MayMay
Fenced since1996Not fencedb199820082004
Aridity index (PET/P)9.4109.4327.9768.1515.331

At each plot, we measured a set of 14 functional traits (Table 2) on species making up at least 85% of total relative abundance (Pakeman & Quested 2007), giving a total of 34 species. We measured traits following standardized protocols (Weiher et al. 1999; Cornelissen et al. 2003). We chose those traits because of their relationships with water stress and grazing. We measured continuous trait variables on a minimum of 25 individuals (i.e. five individuals per plot*five plots). We considered only a single trait value per species: intraspecific variability should not affect results much (Frenette-Dussault et al. 2012). We computed CWM traits following Garnier et al. (2004). Onset of flowering (i.e. the first month when the vast majority of individuals of a species are flowering) was obtained from regional floras. Pastoral value is an integrative index of palatability that is used by agronomists to judge the quality of pastures and rangelands and is based on a combination of productivity and nutrient content (P. Daget, personal communication). Additional methodological details are available in Frenette-Dussault et al. (2012).

Table 2. A list of the 14 functional traits and their ecological relevance used in the maximum entropy analysis to predict species relative abundances. Annual life cycle, presence of a woody stem, clonal reproduction and succulence of leaves or stem are binary traits. Pastoral value is an ordinal trait on a 1—10 scale. R2 represents explained deviance from the generalized additive models of CWM traits as functions of aridity and duration of exclosure. Traits are ordered in decreasing order of explained deviance. Detailed methodology of trait measurement is available in Frenette-Dussault et al. (2012)
Functional traits (Abbrev.)VariableNumber of replicatesa/categoriesEcological relevance R 2
  1. a

    Each replicate corresponds to an individual plant.

Succulence of leaves or stem (Succ)Categorical0: absence, 1: presenceStress avoidance0.696
Onset of flowering (OF)OrdinalOrdinal (first month of flowering)Stress and disturbance avoidance0.633
Woody stem (Wood)Categorical0: absence, 1: presenceStress and disturbance avoidance0.606
Specific leaf area (SLA)Continuous25Resource acquisition/retention0.549
Pastoral value (PV)OrdinalOrdinal (on a 1–10 scale)Palatability, productivity0.521
13C isotope ratio (δ13C)ContinuousFive pooled samples of five individuals eachStress avoidance0.509
Leaf dry matter content (LDMC)Continuous25Resource acquisition/retention0.393
Clonality (Clon)Categorical0: absence, 1: presenceCompetitive ability0.305
Annual life cycle (Ann)Categorical0: absence, 1: presenceStress and disturbance avoidance0.303
Vegetative plant height (Hveg)Continuous50Competitive ability0.182
Leaf carbon: nitrogen ratio (C : N)ContinuousFive pooled samples of five individuals eachResource acquisition/retention, decomposition0.172
Leaf area (LA)Continuous25Stress avoidance, light acquisition0.134
Leaf nitrogen content (LNC)ContinuousFive pooled samples of five individuals eachResource acquisition/retention, palatability0.051
Seed mass (SM)ContinuousFive samples of 20–100 seedsDispersal strategy, establishment success0.022

Predicting Current Species Assemblages with CATS

The CATS model predicts relative abundances given a regional species pool and average trait values (i.e. the CWM traits) that are expected at a given point along an environmental gradient using the principle of maximum entropy of Jaynes (1957a,b). The logic of maximum entropy is rather simple: it is the criterion to make the least biased choice among a multitude of probability vectors for which only partial information is available. In the case of plant functional ecology, maximizing entropy provides the vector of species' relative abundance that constitutes our prediction. Our criterion for this selection is the information contained in each vector, as quantified by relative entropy (H):

display math(eqn1)

where pi represents the predicted relative abundance for species i and qi the prior probability distribution for species i and S is the total number of species in the regional species pool (Shipley 2009). Relative entropy represents the amount of information that we gain after learning the CWM values of the traits, relative to the information available before knowing these values (i.e. the information encoded in the prior probability vector q). To make the least biased selection, we have to choose the vector p with the highest entropy, that is, the vector that minimizes the difference between the vectors p and q. The vector p that will be closest to the prior q will have the highest entropy (i.e. the least new information) among the possible vectors that all agree with the known CWM trait values. Doing otherwise would incorrectly imply that we had additional information beyond that included in the CWM values plus the prior. We used a ‘maximally uninformative’ prior, that is, a uniform distribution giving the same probability to all species (i.e. 1/S), meaning that we had no prior information allowing us to prefer one species over any other. See Shipley (2009), Jaynes (1957a,b) and Della Pietra, Della Pietra & Lafferty (1997) for a more in-depth treatment of the maximum entropy CATS model, including its mathematical aspects.

The first tests of the CATS model were concerned with measuring the degree to which the model could reproduce the observed relative abundances, and so the observed CWM trait values were used. However, to use the CATS model to predict community composition under new climatic conditions, we first had to obtain general predictive relationships between CWM traits and the two independent environmental variables of the study: the aridity index and the duration of exclosures preventing grazing. These general CWM traitenvironment relationships could then be used to obtain the predicted CWM trait values under different scenarios, and these could then be used as constraints in the CATS model to obtain predicted relative abundances. To do this, we used generalized additive models (GAMs) of measured CWM traits as nonlinear functions of aridity index and duration of exclosure with the gam function from the mgcv package in R (R Development Core Team 2011). We used a normal error structure and we estimated the optimal amount of smoothing with the generalized cross-validation (GCV) criterion (Zuur et al. 2009). The GCV criterion is an estimate of prediction error. It is similar to a leave-one-out cross-validation procedure but with computational advantages (Zuur et al. 2009). The optimal amount of smoothing is found by minimizing the GCV criterion (Wood 2006). The GCV criterion has a tendency to overfitting on occasion. To reduce the risk of producing overfitted GAM models, one can use a penalty on each degree of freedom. A practical way of doing so is to force effective degrees of freedom to be counted as 1.4 degrees of freedom (the gamma argument in the gam function) in the GCV criterion (Kim & Gu 2004). For all our GAM models, this penalty did not change the qualitative interpretation of results, so we considered our models adequate to predict CWM values. To obtain the predicted CWM values for each plot, we extracted the fitted values from the GAMs. These predicted CWM traits were input as constraints in the CATS model using the maxent function with a uniform prior to obtain predicted species relative abundances. This function is included within the FD package (Laliberté & Shipley 2011). We tested the CATS model with varying numbers of CWM constraints (from 1 to 14) by adding them in decreasing order of explained deviance (Laughlin et al. 2011). For example, the first model only included succulence because it was the CWM trait that was best predicted from the GAMs (Table 2).

Predicting Future Species Assemblages under Various Climate Change Scenarios

According to the IPCC (2007b), Northern Africa will experience an increase in mean annual temperature and a decrease in precipitation for the period between 1980—1999 and 2080—2099. This will lead to an increase in aridity for the next decades, but there is uncertainty about the magnitude of the change. According to global circulation models, Northern Africa will likely experience a 2 °C increase but this could be even more severe (IPCC 2007b). In terms of precipitation, predictions are less precise but could decrease roughly between 10% and 30% (IPCC 2007b). To provide a general assessment of future climatic conditions and how this could affect plant communities in eastern Morocco, we examined nine scenarios that combined three levels of temperature increase (2 °C, 3 °C and 4 °C) crossed with three levels of precipitation decrease (−10%, −20% and −30%). For each temperature-precipitation scenario, we also considered two grazing scenarios: (i) the maintenance of grazing activities for grazed plots and the maintenance of exclosures for ungrazed plots (the ‘Keep Grazing’ scenario) and (ii) the cessation of all grazing activities in grazed plots and the maintenance of exclosures for ungrazed plots to the 2080—2099 period (the ‘No Grazing’ scenario).

To predict future plant species abundances, we first computed the expected aridity indices for the 2080—2099 period, based on the climate information tool from the FAO website (http://www.fao.org/nr/water/aquastat/gis/index3.stm), by increasing monthly temperature and decreasing monthly precipitation accordingly for each scenario. Second, we used these new aridity indices to predict future CWM values from the GAMs with the predict.gam function (mgcv library). Finally, we used the predicted future CWM values as constraints within the CATS model. We limited ourselves to a uniform prior, because we cannot predict the composition of the future species regional pool.

To facilitate the interpretation of future community structure, we classified species into one of Grime's (2001) three C-S-R primary strategies. We classified species as C-, S- or R-strategists based on their functional traits: (i) perennial species with clonal reproduction, large lateral spread and high stature were ascribed to C-strategists (competitive species; mostly perennial grasses), (ii) perennial species with drought resistance adaptations (e.g. succulence, C4 metabolism) were ascribed to the S-strategists (stress-tolerant species; mostly sub-shrubs) and (iii) annual species with short stature, high SLA and LNC values were ascribed to R-strategists (ruderal species). We used a likelihood ratio test on contingency tables (chi-square statistic) to assess whether differences between observed C, S, and R abundances in 2010 and predicted C, S, and R abundances for the 2080—2099 period were significant. We used point-intercepts as count data.

Results

Predicting Current Species Assemblages with CATS

The percentage of the deviance in the 14 observed CWM traits that could be explained given the aridity index and duration of exclosure from grazing varied from 70% (succulence of leaves or stems) to 2% (seed mass, Table 2). Figure 1 plots each relationship along with the fitted values. In general, traits related to the tolerance or avoidance of water limitations or grazing were relatively well predicted (R2 > 0.5), while traits related to dispersal or the capture of nutrients or light were poorly predicted.

Figure 1.

Regression models (generalized additive models) to illustrate the relationships between aridity and CWM traits. White dots represent grazed plots and black dots represent ungrazed plots. Shaded area is standard error. Smoothed lines are fitted values. Explained deviance (R2) in bold represents a significant model (< 0.05).

The predictive ability of the CATS model increased with the number of CWM traits included in the model and reached an asymptote at nine CWM traits. The model accounted for over 90% of the deviance using the nine observed CWM traits but only 40% using the values predicted from the GAM fits (Fig. 2).

Figure 2.

Explained variance (i.e. correlation between observed and predicted species relative abundances as measured by Pearson's R2) as a function of the number of functional traits included in the CATS model. Trait abbreviations are indicated to represent the order, in which they were included in the model based on the explained deviance from the GAMs. For example, succulence was included first because it is the GAM that had the highest explained deviance (0.696, Table 2). Predictions based on observed CWM traits were added to the graph to illustrate the difference when using observed or predicted CWM traits.

Predicting Future Species Assemblages under Various Climate Change Scenarios

We used the nine CWM traits with the highest explained deviance in the GAMs (Table 2) to make predictions for the 20802099 period. Predictions for the 20802099 period showed a similar shift in plant community structure for all climate change scenarios (Table 3). For both grazing scenarios (i.e. the ‘Keep grazing’ and ‘No grazing’ scenarios), the model predicted a dramatic increase in stress-tolerant sub-shrubs from an actual relative abundance of ~30% to a range of between 67 and 85% with the dominance of this group increasing as the future aridity increased. Currently, the relative abundance of competitors and ruderal species depends on whether the vegetation is grazed or not, with ruderals being especially dominant (59%) in the grazed plots. Both of these groups are predicted to decrease in abundance under all climate change scenarios with ruderals accounting for no more than a maximum of 27% under the least severe future aridity scenario and maintenance of grazing. Competitive species are predicted to make up no more than 13% to 16% of the vegetation with the maximums occurring in the least severe levels of future aridity and no grazing. We illustrated this shift in community composition with the least severe climate scenario (2 °C temperature increase, 10% precipitation decrease) and the most likely social scenario (maintenance of grazing in grazed plots) in Fig. 3.

Table 3. Predicted relative abundances of C-, S- and R-strategists for the period 2080–2099 under various climate and grazing scenarios. Each climate scenario represents a temperature increase (2 °C, 3 °C or 4 °C) and a precipitation decrease (−10%, −20% or −30%) for (a) the cessation of grazing in grazed plots (‘No Grazing’), (b) the maintenance of grazing in grazed plots (‘Keep Grazing’) and (c) the maintenance of exclosures in ungrazed plots. The chi-square statistic indicates whether or not the predicted abundances differ from the observed abundances in 2010
Observed abundances in 2010 CSR 
Grazed plots0.1110.3020.587
Ungrazed plots0.2790.3380.383
 Climate scenarioCSRχ2 P
(a) No grazing2 °C −10%0.1260.7090.1653482.4< 0.0001
3 °C −10%0.1150.7480.1374247.4< 0.0001
4 °C −10%0.0980.7770.1254690.9< 0.0001
2 °C −20%0.0710.8130.1165163.4< 0.0001
3 °C −20%0.0800.8240.0965851.0< 0.0001
4 °C −20%0.0750.8280.0975857.2< 0.0001
2 °C −30%0.0600.8520.0886402.9< 0.0001
3 °C −30%0.0620.8490.0896342.8< 0.0001
4 °C −30%0.0590.8190.1225097.0< 0.0001
(b) Keep grazing2 °C −10%0.0560.6700.2742083.0< 0.0001
3 °C −10%0.0490.7020.2492503.7< 0.0001
4 °C −10%0.0370.7350.2283000.1< 0.0001
2 °C −20%0.0270.7880.1854024.3< 0.0001
3 °C −20%0.0400.7970.1634210.1< 0.0001
4 °C −20%0.0360.8050.1594373.7< 0.0001
2 °C −30%0.0190.8400.1415328.1< 0.0001
3 °C −30%0.0160.8200.1644853.0< 0.0001
4 °C −30%0.0240.8410.1355328.2< 0.0001
(c) Ungrazed plots2 °C −10%0.1610.6350.2042280.5< 0.0001
3 °C −10%0.1600.6740.1663029.9< 0.0001
4 °C −10%0.1410.7160.1433925.3< 0.0001
2 °C −20%0.0910.7750.1345457.2< 0.0001
3 °C −20%0.1020.7940.1046173.9< 0.0001
4 °C −20%0.0890.8100.1016685.0< 0.0001
2 °C −30%0.0580.8480.0948340.5< 0.0001
3 °C −30%0.0580.8420.1008069.6< 0.0001
4 °C −30%0.0530.8220.1257434.4< 0.0001
Figure 3.

Community composition by functional group (C-, S- and R-strategists) for 2010 and 2080–2099 in grazed and ungrazed plots. The scenario represented here consists of an increase of 2 °C in mean annual temperature, a decrease of 10% in mean annual precipitation and the maintenance of grazing in grazed plots. The numbers in the columns represent the observed (2010) and predicted (2080–2099) relative abundances of each functional group per year and condition (grazed or ungrazed). Proportions between 2010 and 2080–2099 were significantly different for both grazed and ungrazed plots (< 0.0001). For other climate change scenarios, please refer to Table 3.

Discussion

Predicting Current Species Assemblages with CATS

Traitenvironment relationships are thought to reflect environmental filtering based on species functional traits (Keddy 1992). The relationships reported in this study were characterized by nonlinear tendencies and contrasted explained deviance among functional traits (Fig. 1). The accuracy of CWM trait values for a given community strongly affects the prediction of species abundances when using the CATS model. The use of measured CWM trait values as model constraints in CATS is both valid and useful to determine the degree of association between traits and relative abundances and to establish an upper limit on the predictive value of such traits. However, as our measured CWM trait values require the relative abundances of the species in their calculation, their use as constraints in the CATS model has no practical predictive value. In order to predict composition in communities whose CWM values are unknown but whose position along environmental gradients is known, it is necessary to replace such unknown CWM trait values by less precise estimates based on general relationships between CWM values and environmental variables. In our study, these less precise estimates were obtained from GAM regressions of the CWM values on aridity and duration of exclosures. Clearly, as the differences between the actual and estimated CWM values increase (i.e. as the residual errors in the GAM models increase), the predictive ability of the CATS model necessarily decreases as well. The decrease in the predictive success of CATS using the actual and predicted CWM values (Fig. 2) is due to the errors introduced when replacing the actual CWM values by the expected values given the measured levels of aridity and duration of exclosures.

This loss of predictive ability, when using environmental variables as input to the model, is likely due to a number of different causes. First, our measure of ‘aridity’ was based on yearly totals of precipitation and temperature at each site and so only imperfectly measured the differences in ecologically relevant levels of water availability in the soil and its variation during the year that actually filter species via their traits. Such a measure of aridity might be acceptable when comparing sites differing widely in these climatic variables (e.g. including sites in arid, semi-arid and subhumid zones for example) but not over the short ranges occurring in our study where all sites were strongly water-limited. Second, the variable ‘duration of exclosures’ was only an imprecise measure of the ecologically relevant levels of the intensity, duration and type of grazing experienced by the plants. The vegetation in arid ecosystems tends to respond very slowly to changing grazing conditions (Valone et al. 2002). Third, we presumably missed other important environmental variables associated with our traits. For instance, although traits that are functionally linked to water availability (succulence, SLA, δ13C, onset of flowering) or grazing (woody stems, pastoral value) were moderately well predicted by our environmental variables (0.7 > R2 > 0.5), traits linked to nutrient acquisition/retention (leaf C : N, LNC), competition (height, clonality) or dispersal (seed mass) were poorly predicted. Thus, future research should focus on identifying and accurately measuring important environmental variables responsible for trait-based filtering and on developing strong general relationships between environmental conditions and functional traits.

Finally, as previously shown (Frenette-Dussault et al. 2012), plants of arid steppes might not have evolved towards a single ‘optimal’ ecological strategy: rather, they have adopted two distinct strategies, namely avoidance of water shortages (annual species whose phenology is timed to respond to rainfall) and tolerance of water shortages (sub-shrubs). The contrast in functional trait values between those two functional groups probably reduced the correlation between CWM trait values and environmental variables. Our application of the CATS model used community-weighted means (a measure of central tendency) as input but this could be modified to include community-weighted variance (CWV; Sonnier, Shipley & Navas 2010).

Predicting Future Species Assemblages under Various Climate Change Scenarios

The CATS model predicted that stress-tolerant sub-shrubs will become the dominant functional group in coming decades regardless of which climate change scenario is considered. If true, then the pastoral value of the vegetation will decline. This shift in community composition reflects a process of desertification. Desertification consists of a reduction in herbaceous cover, an increase in shrub cover and a general increase in bare soil surface area (Asner et al. 2004). Heavy grazing and drought are usually considered to be major factors contributing to desertification (Asner et al. 2004). Grazing exclusion may help to restore arid grasslands and steppes, but the very slow response of these ecosystems may delay visible changes in community structure, especially the return of perennial grasses (Valone et al. 2002). Such perennial grasses (i.e. the C-strategists) are predicted to be the least abundant group even in ungrazed sites (Table 3 and Fig. 3). Perennial grasses are an important component of these ecosystems to maintain grazing activities, which are vital to local populations’ subsistence. If plant community changes are in line with our predictions, serious actions will have to be undertaken to maintain long-term sustainability of the arid steppes of eastern Morocco.

It is important to highlight some limitations of the CATS model under various climate change scenarios. For instance, our predictions of the CWM trait values in these scenarios assume that the GAM regressions can be extrapolated beyond the environmental ranges of the data used to construct them. We have no strong evidence for this assumption beyond the fact that the resulting CWM trait values, when input into the CATS model, plausibly predicted a shift from a steppe to desert vegetation. A major limitation affecting the validity of our predictions of future species assemblages is our inability to predict how the species regional pool will change with increasing aridity. Because we maintained the same species pool as exists presently, we implicitly assumed that no new species will migrate into the region. In fact, it is probable that some new species will migrate due to changing climatic conditions (e.g. Walther et al. 2002; Bertrand et al. 2011), although the pool of species that are capable of migrating and their respective speeds of colonization remain unknown. Adding such new species into the model is mathematically simple once such potential invaders are identified and their traits measured. For example, Fredolia aretioides (Coss. & Dur. ex Bunge) is a drought-resistant endemic sub-shrub that is currently distributed just north of the Sahara and which can persist in rocky steppes receiving ≤100120 mm of rain per year. We did not include F. aretioides in our species pool because it is not currently found in our study sites but occurs further south where aridity is more severe. Consequently, it is not unrealistic to think that F. aretioides may be found in our study sites in future decades. Future work should focus on defining ‘extended’ species pools, as it has been shown that using a ‘restricted’ species pool may yield misleading conclusions (Barbet-Massin, Thuiller & Jiguet 2010).

Another uncertainty concerns the fate of grazing activities in future years. Depending on how socio-economic conditions evolve in Morocco, the importance of pastoralism may evolve as well. Pastoralism has often been associated with precarious living conditions (Maatougui, Acherkouk & Bouayad 2005). Recent changes in pastoral activities have increased the pressure on the arid steppes of eastern Morocco and have further accentuated their current state of degradation (Bechchari et al. 2005). Given the importance of arid steppes for ecological processes and socio-economic well-being, future management policies for such ecosystems will have to find a balance between conservation and production to ensure their long-term use.

Acknowledgements

Funding and supervision of this study was provided by the Emirates Center for Wildlife Propagation (ECWP) under the leadership of the International Fund for Houbara Conservation (IFHC). We are grateful to H.H. Sheikh Mohamed bin Zayed Al Nahyan, Crown Prince of Abu Dhabi and Chairman of the IFHC, and H.E. Mohammed Al Bowardi, Deputy Chairman of IFHC, for their support. We thank F.O. Ezza, A.S.R. Sanz, M. Bidat and H. Hdidou for their assistance during field work. We also thank Dr. G. Sonnier, Dr. W.F.J. Parsons, Dr. Jason Fridley and anonymous referees for valuable comments that substantially improved the manuscript. We are grateful to the Haut Commissariat aux Eaux et Forêts et à la Lutte Contre la Désertification for authorizing access to the Tirnest site.

Ancillary