Analysing the spatial heterogeneity of emergent groups to assess ecological restoration


*Correspondence author. E-mail:


1. Effective restoration of ecosystem structure and function needs to be built on a strong conceptual basis. The evaluation of restoration is critical in this respect, as it requires an accurate understanding of vegetation dynamics. This study was, therefore, designed to analyse how initial restoration treatments and spontaneous ecological processes act together to produce spatial heterogeneity of plant species at various scales, taking into account that species might respond differentially to these processes in accordance with their biological attributes.

2. In a model system consisting of a large land settlement located on the banks of the river Rhône (France), 85 geo-referenced plots were located, and the abundance of all plant species and seven biotic and abiotic environmental factors that have been modified by restoration were recorded. All species were classified into emergent groups (EGs) based on 12 plant traits, and the abundance of each EG was derived for each plot. Spatial variables that explained the spatial autocorrelation of the EGs at four scales were computed. This was performed using a spatial eigenvector mapping technique. The spatial variables were then linked to a range of selected environmental factors.

3. Large-scale patterning of EGs was explained mainly by the spatial heterogeneity of soil quality and vegetation cover. The attribute combination of the EGs that varied most at this scale suggested that the restoration had influenced vegetation development by inducing harsh physical conditions in some part of the studied area and biotic filtering (through competition) in others.

4. A large amount of the spatial variation of EGs was not explained by the environmental factors providing some evidence for pure spatial autocorrelation effects. At the finest scale, this was linked to poor dispersal abilities of some EGs.

5.Synthesis and applications. By combining a trait-based approach with spatially explicit methods, we explained and quantified the role of environmental changes induced by restoration and spontaneous biotic processes in structuring plant communities on different spatial scales within a large land settlement. We discuss how such an approach offers new opportunities for an improved assessment of induced and spontaneous structuring factors for restoring and for monitoring plant communities during ecological restoration.


Ecological restoration aims to assist the recovery of ecosystems that have been damaged or destroyed. Restoration success depends on setting appropriate objectives (Cairns 2000) and the subsequent use of suitable criteria for evaluating the outcomes. The question of objectives has been the subject of much discussion (Parker & Pickett 1997), but most authors agree that success should be judged on both structural and functional aspects of ecosystems. Ecological restoration should, therefore, be based on an accurate understanding of ecosystem dynamics (Palmer, Ambrose & Poff 1997).

Community succession and assembly rules are two of the most relevant ecological concepts for ecological restoration (Young, Chase & Huddleston 2001). Community succession refers to the predictable turnover of species composition. The position of species along a successional gradient depends on their performance in colonization and biotic interaction processes (competition and/or facilitation), and any modification of the biotope that occurs during the succession (Tilman 1990). Assembly rules are based on the response of organisms to local environmental factors (biotic and abiotic) after their random arrival. To establish and survive, organisms need the appropriate ecological abilities to face the environmental filters/constraints of that biotope (Keddy 1992). Both concepts rely upon the idea that community structure can be predicted from knowledge of organisms’ traits, as these traits affect the response of the species to environmental factors, and any reciprocal effect on ecosystem functioning (Lavorel & Garnier 2002). In turn, trait assembly in local communities is recognized to influence ecosystem-level properties such as primary productivity, biochemical cycling, resistance and resilience to disturbance (Lavorel & Garnier 2002), which are useful ecosystem characteristics for evaluating success of ecological restoration (Ehrenfeld & Toth 1997). For that reason, the trait-based approach has been used to evaluate the restoration of plant communities (Hérault, Honnay & Thoen 2005) and to assess the performance of species in restored communities (Pywell et al. 2003).

Knowledge of the spatial aspects of ecological system functioning and structuring is also important. The importance of biogeographical gradients and landscape structure on the recruitment of species in degraded sites has been identified (Tong et al. 2006), indicating the need to consider a landscape ecology approach beyond the spatial scale of the restored site. However, spatial heterogeneity is also of interest at much finer, within-site scales. Indeed, the spatial heterogeneity of organisms that develops on a restored site will almost certainly reflect interacting abiotic and biotic processes, which together influence community succession and assembly at a wide range of scales (Levin 1992; Tilman 1994). Moreover, the spatial patterning of organisms has also been shown to be important in maintaining both ecosystem structure and function (Pacala & Deutschman 1995). Therefore, in recent studies of spatial heterogeneity in degraded ecosystems, some authors have emphasized its relevance for evaluating restoration success (Maestre et al. 2003; Seabloom & Van Der Valk 2003).

Recently, both trait-based and spatial approaches have allowed advances in the theoretical basis for planning and implementing ecological restoration projects. The aim of this paper is, therefore, to combine both these approaches within a model system. The model system is a large land settlement located on the banks of the river Rhône, where diverse initial restoration treatments were applied. More precisely, our objective was to assess how active restoration influenced the spatial patterning of plant traits after 25–30 years of spontaneous vegetation dynamics on the site. This study was performed by classifying the plant species into emergent groups (EGs) (Lavorel et al. 1997), i.e. species that share similar life attributes, and then analysing potential sources of their spatial heterogeneity. As spatial heterogeneity should be considered as both endogenous and exogenous, we hypothesized that: (i) the modification of environmental variables by restoration efforts induced spatial heterogeneity of the EGs at different spatial scales; and (ii) the dispersal abilities of the EGs influenced their spatial heterogeneity over the study site. These hypotheses were tested using a spatial eigenvector mapping technique. We then discuss how such an approach might be used to evaluate and monitor the effects of restoration in creating spatial heterogeneity during ecological restoration.

Materials and methods

Study site, vegetation sampling and environmental factor measurement

Channels diverting the natural course of the Rhone River have been dug along the Rhone valley (France) between the 1950s and 1980s for producing hydro-electricity and promoting river navigation (Fruget & Michelot 2001). Large banks have been built alongside these channels composed of relatively large areas of homogeneous gravels and pebble soils. A range of different restoration approaches were used to establish vegetation in the harsh conditions found on these banks.

This study was conducted on a raised surface (c. 2000 m × 100–150 m) located on the restored banks of the river Rhône at Péage-de-Roussillon (4°45′–4°46′E; 45°21′–4°22′N). The climate is temperate; the mean annual temperature fluctuates between 7 °C and 11 °C, and the mean annual precipitation is between 700 and 800 mm with peaks in April–June and September–November. The site is separated from the river by an earthen embankment, and soil moisture conditions are totally independent of river fluctuations. Unfortunately, there is little detailed documentation of the restoration treatments applied on the site immediately after the bank was constructed; however, attempts were made to initiate vegetation colonization by first improving the abiotic conditions to facilitate plant establishment and growth and second, introducing vegetation material to counteract slow natural colonization (Maman 1984). The treatments varied markedly over the entire site. As a result, the soil structure varied from pebbles to clay loam and species mixtures have been sown heterogeneously, and trees have been sparsely planted. Thirty years later, the herbaceous stratum is now dominated by Elytrigia repens (L.) Desv, and the major woody species is Populus nigra L.

In this area, we tracked spatial heterogeneity of vegetation by sampling 85 plots (5 × 5 m) located on a geo-referenced grid using an a priori rule-based, constrained, random sampling approach. The rules were defined to fit the sampling design to the requirements of the spatial analytical techniques; i.e. the aim was to avoid under- and over-sampling of areas whilst providing a large range of distances between the plot positions, which are required for good practice in spatial autocorrelation analysis (Aubry 2000). This method can be considered as an alternative to stratified-random sampling where there is no obvious criterion for stratification. The plot size was chosen to fulfil two criteria, i.e. each plot was: (i) sufficiently large to represent local species assemblages and (ii) sufficiently small to remain within homogeneous environmental conditions. Within each plot, 10 quadrats (50 × 50 cm) were selected randomly as subsamples, and the species cover in each was assessed visually. The mean cover of each species for these subsamples was then calculated for each plot.

A range of environmental factors (classified into abiotic and biotic) were also measured for each plot (Table 1). These factors were chosen because they corresponded to biotic and abiotic environmental parameters originally imposed during restoration (Maman 1984). The abiotic factors (coarse particle content, coarse particle size, fine particle size) were three ordinal soil variables relating to soil texture, and the biotic factors were related to vegetation structure (number of trees, cover of shrubs, cover of ground vegetation, heterogeneity of ground vegetation cover).

Table 1.   Environmental variables recorded at each of the 85 sample plots on the study site on the Rhône channel embankment
Environmental factorCodeVariable description
 Coarse particle contentS1Ordinal: 1 – absence, 2 – low, 3 – medium, 4 – important, 5 – very important
 Coarse particle sizeS2Ordinal: 1 – gravel (2–20 mm), 2 – pebbles (20–50 mm), 3 – rock (50–200 mm), 4 – boulder (>200 mm)
 Fine particle sizeS3Ordinal: 1 – sand, 2 – loamy sand, 3 – clay sand, 4 – sandy loam, 5 – sandy clay, 6 – loam, 7 – clay loam, 8 – clay
 Tree numberTreeQuantitative: count
 Shrub coverSchrubQuantitative: percentage
 Mean herbaceous coverHerb1Quantitative: percentage
 Herbaceous cover heterogeneityHerb 2Quantitative: coefficient of variation of percentage cover

Emergent groups’ identification

The EGs were produced by a classification of data on 12 plant traits (Table 2) derived for the entire species complement detected during this study (n = 138 species). These traits were chosen to separate species based on different phases of their life cycle, e.g. dispersion, establishment and then persistence (Weiher et al. 1999). This information was derived from two existing plant trait data bases (Biolflor: Klotz, Kühn & Durka 2002; Clopla: Klimešet al. 1997). As most of the species detected were relatively common, it was possible to derive trait values for the entire species complement.

Table 2.   Plant traits, and their associated attributes, used to classify the species complement detected in the 85 plots on the Rhône channel embankment into emergent groups
TraitsDescriptionType of variable
  1. Data were derived from Biolflor (Klotz et al. 2002) and Clopla (Klimešet al. 1997).

Life-form1: Phanerophyte (Ph); 2: Chamaephyte (Ch); 3: Hemi-cryptophyte (H); 4: Geophyte (G); 5: Therophyte (Th)Qualitative
Life history1: Annual (a); 2: Biennial (b); 3: Perennial (p)Qualitative
Leaf persistence1: Spring green (v); 2: Summer green (s); 3: Persistent green (i)Qualitative
Canopy height1: <100 mm; 2: 101–299 mm; 3: 300–599 mm; 4: 600–999 mm; 5: 1000–3000 mm; 6: >3000 mmOrdinal
Canopy structure1: Rosette (Ro); 2: Semirosette (Sro); 3: Leafy (L)Qualitative
Type of reproduction1: By seeds (s); 2: By seeds rarely vegetatively (ssv); 3: By seeds and vegetatively (sv); 4: Vegetatively rarely by seeds (vvs)Qualitative
Dispersal type1: Endo-zoochory (Ing); 2: Exo-zoochory (Adh); 3: Anemochory (V); 4: Unassisted (Nsp)Qualitative
Seed weight1: <0·2 mg; 2: 0·2–2 mg; 3: 2–10 mg; 4: >10 mgOrdinal
Seed shape (length/breadth ratio)1: <1·5; 2: 1·5–2·5; 3: >2·5Ordinal
Growth form1: Legume (L); 2: Graminoid (G); 3: Forb (F); 4: Woody (W)Qualitative
Flowering phenology1: Flowering complete by end of June (E); 2: Summer flowering (S); 3: Flowering not started until July (L); 4: Auntumn flowering (A); 5: Flowering throughout most of the year (U)Qualitative
Type of clonality1: Nonclonal (Nocl); 2: Infrequent multiplication (Infr); 3: Frequent multiplication with short spacers (<10 cm) (Sh); 4: Frequent multiplication with long spacers (>10 cm) (Lg)Qualitative

The method used to identify the EGs was closely related to that of Verheyen et al. (2003) and applied since by others (e.g. Fukami et al. 2005; Hérault et al. 2005). This method is based on a similarity matrix calculated from the trait data for the species. The similarity coefficient used here was the percentage disagreement, which was particularly useful because most of the 12 traits were categorical variables. A Hierarchical Cluster Analysis (CAH) using an agglomerative process (Ward’s method) was then applied to this similarity matrix and the EGs were derived through interpretation of the resultant dendrogram. The relationship between the individual trait variables and the EGs were tested using the Pearson’s χ2 test for qualitative variables and the Kruskall–Wallis nonparametric test for ordinal variables.

The relative abundance of each EG in each sample plot was calculated using plot species cover values. The resulting matrix was transformed for use as a response table in constrained ordination. We used the Hellinger transformation that expresses each abundance as a fraction of the total abundance of EG at one site and takes the square root of that fraction (Legendre & Gallagher 2001).

Spatial heterogeneity quantification

Spatial heterogeneity of EGs variation (expressed as the response matrix Y) was quantified by introducing spatial relationships as predictors in the form of a spatial matrix W.

Then we tested whether the spatial matrix could be explained by the environmental variables (summarized in an explanatory matrix X).

The building of the spatial matrix W was critical in this procedure. Originally, such spatial matrices were derived from a trend surface analysis based on the geographical coordinates of sample locations (Borcard, Legendre & Drapeau 1992). However, in doing so, only spatial processes that take place at the largest scale of the sampling scheme could be explored. Another method, the principal coordinates of neighbour matrices (PCNM) method, was developed for building a spatial matrix based on a spectral decomposition of space (Borcard & Legendre 2002). This was performed by diagonalizing a spatial weighting matrix constructed by truncating a pairwise Euclidean distance matrix between sampling locations. Truncation leads to retention of only the closest neighbours. Then, principal coordinates associated with positive eigenvalues provide a set of spatial variables.

Recently, Dray, Legendre & Peres-Neto (2006) have proposed a new method that is a generalized approach of PCNM, called Moran’s eigenvectors map (MEM). The method is similar in that a spatial weighting matrix was diagonalized; however, here its construction is more flexible as different connectivity schemes and various weighting functions can be used. Then the eigenvectors that maximize the Moran’s index of autocorrelation were extracted.

Using both PCNM and MEM methods, spatial explanatory variables track periodic variation in ecological data among sites. Then, period length relies directly on the scales of patterning that can be perceived in the sampling scheme. On the basis of similarity of their periods, spatial variables can be grouped into submodels revealing different scales of the spatial distribution of the response variables Y. Each submodel is calculated as a linear combination of the spatial variables that pertain to the same scale.

First, we used all three methods (polynomial, PCNM, MEM) and included a set of random orthogonal vectors to assess whether the methods provided better fits than by chance alone. Then, the best spatial matrix in regard to our data set was selected. The selection of both the spatial matrix and model was based on an Akaike Information Criterion (AIC)-like criterion (Godinez-Dominguez & Freire 2003 in Dray et al. 2006) for a multivariate response in canonical analysis. In the MEM approach, we tested different schemes of connectivity between plot locations using five procedures (Delaunay triangulation, Gabriel Graph, Relative Neighbourhood Graph, Minimum Spanning Tree and the Distance Criterion). Four weighting functions for each connectivity scheme were also tested, namely binary, linear, concave-down and concave-up functions. The spatial matrix construction, Moran’s eigenvector computation and the spatial model selection were performed using the spacemakeR ( package for r (R Development Core Team 2009).

The next step was aimed to define the submodels separating the different spatial scales on which the EGs were patterned (matrix Y). These submodels corresponded to our final spatial matrix W. The environmental variables of matrix X were then used to explain the spatial variables of matrix W. Our goal was not to predict abundance values of EG in space but rather to highlight which recorded environmental factors (Table 1) might explain observed variations in spatial submodels (i.e. variations of EGs on different scales). To do so, we performed multiple linear regressions of spatial submodels. As automated procedures for selection of appropriate sets of environmental predictors are a critical issue in the analysis and/or modelling of organism–environment relationships, we compared results from three different techniques for each spatial submodel. These techniques were the classical forward and backward stepwise procedures (Venables & Ripley 2002), least angle regression (see Efron et al. 2004 for details) and hierarchical partitioning (Chevan & Sutherland 1991). Then we retained one set of predictors between those offered by the three selection procedures on the basis of its explanatory power and reasonable complexity using information criteria [i.e. AIC; corrected Akaike Information Criterion (AICc) and Bayesian Information Criterion] to perform final multiple linear regressions (Burhnam & Anderson 2004). Lastly, we assessed the relative contribution of abiotic (soil variables) and biotic (vegetation variables) sets of environmental factors in predicting the four spatial submodels using variation partitioning (Borcard et al. 1992).

Variation sources on the spatial heterogeneity of the EGs

The main goal of this study was to analyse and quantify the relative effects of different sources of spatial heterogeneity of the EGs. Our hypotheses focused on the combination of the spatial heterogeneity of environmental factors and pure spatial autocorrelation effects (Legendre 1993). Therefore, the variation of the EGs abundance matrix was partitioned (response table Y) between the environmental matrix (explanatory table X) and the spatial matrix (explanatory table W). Following Borcard et al. (1992), the variation in Y was divided into four independent components: (i) local variation of the EGs explained by the environmental variables independent of spatial structure; (ii) spatial structure of the EGs that is shared with the environmental variables; (iii) spatial patterns of the EGs that is not shared by the environmental variables; and (iv) unexplained variation from the variables included in the analysis. The partitioning was performed using canonical redundancy analyses (RDA).


EG identification

Thirteen EGs were derived from the cluster analysis (see Table S1). The main traits that discriminated species into groups were growth form, life form, canopy height and type of clonality. The first three groups consisted of hemi-cryptophyte graminoids varying in canopy height and type of clonality, i.e.: (1) short tufted graminoids, (2) tufted graminoids and (3) graminoids with long internodes. The next two groups represented short perennials: (4) short hemi-cryptophytes and (5) short chamaephytes. The next two were geophytes: (6) vernal geophytes and (7) summer-green geophytes. One group consisted of tall perennial legumes: (8) hemi-cryptophyte legumes. Four groups represented nonperennial species: (9) biennials, (10) annual graminoids, (11) annual forbs and (12) annual legumes. The final group consisted of shrub and tree species: (13) woody species. Thus, as there were 12 groups of herbaceous species and only 1 group of woody species, the subsequent analyses were focused on the groups of herbaceous species (i.e. excluding group 13).

Spatial heterogeneity of the EGs

First, a detrended correspondence analysis highlighted the pattern of variation of the 12 herbaceous EGs over the 85 sampling plots. The first axis, that explained 32·4% of the total variation of EGs, was mainly associated positively with EG3 (graminoids with long-internodes) and negatively with EG5 (short chamaephytes). The second axis accounted for 9·7% of the total variation of EGs and was associated positively with EG1 (short tufted graminoids) and negatively with both EG7 (summer-green geophytes) and EG8 (hemi-cryptophyte legumes).

The spatial matrix that best explained the distribution pattern of the EGs in these 85 plots – designated on the basis of its lowest value of AICc – was then constructed using the MEM approach (see Table S2). The connectivity scheme of this matrix was based on the distance criterion (dnn). This criterion corresponds to the maximum distance of the minimum spanning tree (Dray et al. 2006). Here, γ = 494 m was used to keep all plots connected with a concave up-weighting function. The MEM procedure selected 11 eigenvectors that were grouped into four spatial scales on the basis of the similarity of range of their semi-variogram (Table 3; Fig. 1): a very broad-scale (c. 1 km) submodel (MEM 1, 3); a broad-scale (300–500 m) submodel (MEM 6, 7); a meso-scale (80–300 m) submodel (MEM 14, 36); and a fine-scale (<80 m) submodel (MEM 38, 51, 67, 70, 77).

Table 3.   Canonical coefficients for standardized variables of emergent groups distribution extracted from canonical redundancy analyses (RDA) computed for each of four spatial scales
MEM vector no.Spatial scaleCanonical coefficient axis 1Canonical coefficient axis 2% Variation of EGs
  1. Percentage of the variation of emergent groups (EGs) that was explained by the submodels and the whole spatial model are reported.

  2. MEM, Moran’s eigenvectors map.

 1Very broad0·6100·08021
 3Very broad−0·148 0·330
 6Broad−0·4570·029 8
14Meso−0·2990·229 5
Whole model52
Figure 1.

 Maps of the 85 vegetation sample plots located on the Rhône channel embankment. (a) Samples scores on the first axis from the detrended correspondence analysis (DCA) of the emergent groups in the sampling plots (centred on 0); (b) samples scores on the second axis from the DCA of the emergent groups in the sampling plots (centred on 0); (c) very broad-scale submodel; (d) broad-scale submodel; (e) intermediate-scale submodel; (f) fine-scale model. Black bubbles are positive values, white bubbles are negative values except for DCA scores where black bubbles are values superior to mean sample scores and white bubbles are values inferior to mean sample scores.

Variation partitioning showed how the whole spatial model and environmental variables interacted in the explanation of the variation of EGs among plots (Fig. 2). The whole spatial model explained 52% of the variation of EGs. This was split into two fractions: the spatial patterns of EGs that were shared by the environmental variables included in the analyses (21%) and the spatial patterns of EGs that was not shared by those same environmental variables (31%). Finally, a fraction of 10% represented the local variation of EGs that was explained by the environmental variables independently of any spatial structure, leaving c. 38% of the EGs variation unexplained.

Figure 2.

 Variation partitioning of the emergent group data derived from 85 sample plots located on the Rhône channel embankment. The variation accounted for by spatial and environmental factors have been identified.

The EGs were patterned at different scales (Fig. 3) whereas the set of environmental variables tested explained mainly the very broad scale (Table 4; Table S3). This very-broad-scale submodel explained 21% of the variation of EGs between sampling plots. Three EGs varied markedly on this scale: EG3 and EG4 contributed to the positive part of the first axis of the RDA whereas EG5 and EG11 contributed to the negative part. The environmental variables were significant in explaining this scale of variation (adjusted R² = 0·48), with, in decreasing order of strength, the fine particle size of soil (S3, b = 0·48), the number of trees (Tree, b = −0·24) and the mean cover of herbaceous vegetation (Herb1, b = 0·21). This set of environmental variables was selected whichever selection technique was used (see Table S3). Additionally, abiotic (soil variables) and biotic (vegetation variables) factors explained distinct fractions of this very-broad-scale submodel. The fine-scale submodel explained 18% of the variation of EGs, EG3 and EG1, and varied distinctly at this scale. However, none of the environmental variables used here explained this spatial patterning. The broad-scale and meso-scale submodels explained only 8% and 5% of the variation of EGs respectively. Competing sets of environmental variables emerged from the different selection techniques we used to explain those both submodels (see Table S3). Moreover, the adjusted R2 for the reported set of predictors (Table 4) were only 0·03 for the broad-scale submodel and 0·12 for the meso-scale submodel.

Figure 3.

 Scores of the 12 herbaceous emergent groups on the first canonical axis of the redundancy analyses (RDA) performed for each scale submodel previously defined from the 85 sample plots located on the Rhône channel embankment. Black points show the three emergent groups of which variation was best explained by each scale submodel.

Table 4.   Standardized regression coefficients of the environmental variables (see Table 1) that explained the different scales of spatial patterning of the 12 studied emergent groups in the 85 sample plots located on the Rhône channel embankment
 Very broadBroadMesoFine
  1. Sets of environmental variables have been defined by combining results from three different selection techniques: backward and forward stepwise, least angle regression and hierarchical partitioning (see Table S1). Adjusted R2 for the whole models and for fraction explained only by abiotic (soil) factors or biotic (vegetation) factors and shared by both are reported.

  2. ***≤ 0·001; **0·001 ≤ ≤ 0·01; †0·05 ≤≤ 0·1.

 S2  0·16 
 Schrub  0·27 
 Herb2 0·23  
Adj R2 (soil only)0·2900·01(−)
Adj R2 (vegetation only)0·090·030·08(−)
Adj R2 (vegetation and soil)0·1000·03(−)
Adj R2 (whole model)0·48***0·03†0·12**(−)


The emergent groups

The set of traits used for analysis has been chosen for characterizing species abilities in the main phases of plant life cycle. However, one might consider that species classification in EGs may be altered by the use of different set of traits (Lavorel et al. 1997).

In this study four main types of trait were implicated in the delineation of the EGs. First, life-form, which is recognized as correlated with numerous other morphological and physiological traits (McIntyre, Lavorel & Tremont 1995), and has been associated with plant longevity, ability to occupy space and tolerance to disturbance (Weiher et al. 1999). Secondly, growth form which is related to other traits of the persistence phase of the plant life cycle (Leishman & Westoby 1992) and is often used to predict the response of plants to environmental factors and effects on ecosystem functioning (Dorrepaal et al. 2005). Thirdly, canopy height has been recognized for a long time to be positively correlated with plant growth and competitive ability (Weiher et al. 1999). Finally, type of clonality, which is related to asexual recruitment, plant longevity, short mobility and resource storage, influences both spatial patterning and competitive relationships at the community level (Oborny & Bartha 1995). The attribute combinations that characterized the EGs here are typical of those found in other studies, and as such could help to interpret the response of the EGs to spatial heterogeneity in interaction with measured environmental factors in this study, and perhaps more generally.

Effects of biotic and abiotic environmental factors

On one side, niche theory is based on the idea that species performance varies along environmental gradients. As these environmental gradients can occur over space, this concept has been used to explain the spatial distribution of species (Austin, Nicholls & Margules 1990) and to model the impact of environmental change on it (Guisan & Thuiller 2005). Here, the environmental factors explained 31% of the variation of the EGs. Two-thirds of this environmental fraction was shared with spatial predictors while the one-third remaining was independent of spatial predictors (Fig. 2). This might indicate local effects of environmental factors at a spatial scale smaller than the sampling design could detect (Borcard et al. 1992). Most of the environmental effects on spatial patterning of the EGs occurred at the very broad scale and were mostly due to soil conditions. Independent effects of biotic factors accounted for 10% of the variation of the very-broad-scale submodel. Environmental effects were expressed as an axis of variation with short chamaephytes (EG5) on one side, and graminoids with long internodes (EG3) on the other. The former were associated with coarse particles in the soil and a low vegetation cover, i.e. conditions where there had been no restoration treatment to aid vegetation growth and establishment (Maman 1984). The group of short chamaephytes (EG5) had attributes typical of species that can tolerate harsh physical conditions (chamaephyte life-form, short canopy height, long leaf persistence, investment to vegetative reproduction). This suggests that the abundance of EG5 on this part of the axis of variation arose from appropriate attribute combinations for harsh environmental conditions. In contrast, EG3 was associated with soils with a fine particle size and high vegetation cover, and typically found where restoration efforts had enhanced soil water retention and plant establishment (Maman 1984). The group of graminoids with long internodes (EG3) has the ability to invade at the local scale (large investment in rapid vegetative spread) and a high competitive ability (higher canopy height with long life span). The success of this group on this part of the main axis of variation of EGs over space could be explained by the more productive habitat. In turn, this would indicate that there would be greater biotic constraints in these local communities. In this way, EG11 (annual forbs) varied inversely with EG3 at this very broad scale (Fig. 3). This group has weak competition abilities and this suggests that the spatial heterogeneity of biotic constraints might limit the spatial expansion of this group at the very broad scale. The group of short hemi-cryptophytes EG4 also varied markedly at this scale, but was much more difficult to interpret in regard to interaction between environmental conditions and trait attributes.

These results also highlight that the modification of environmental variables by restoration has imposed a source of spatial heterogeneity for some of the EGs at the very broad scale, which translates into a gradient ranging from one where physical constraints are most important through to one where biological constraints predominate. This supports the need to consider the responses of species to both physical variables of environment and the impact that some species might have on biotic constraints in the understanding, and by extension the modelling, of the spatial heterogeneity of plants (Guisan & Thuiller 2005).

Effects of dispersal abilities

Before growing and surviving in a local community, individual plants need to arrive. Arrival of a species at a given site is limited both by local site isolation and the dispersal abilities of that species according to regional dynamics of population and community theories (Freckleton & Watkinson 2002; Leibold et al. 2004). In homogenous environments, the dispersal kernel of individual species, influenced by its dispersal abilities, is a key factor in determining its spatial structure (Nathan & Muller-Landau 2000; Ozinga et al. 2005). Here, c. 30% of the variation of EGs over the studied area was explained by spatial patterning, independent of environmental variables. This could be interpreted, either as spatial heterogeneity induced by environmental factors that have not been introduced in the analyses or as biologically generated spatial heterogeneity determined by dispersal and biotic interactions (Borcard et al. 1992; Ozinga et al. 2005). The environmental variables explained the very-broad-scale variation of EGs but not the fine scale. This provided some evidence that spatial patterning between the nearest local communities occurred in the context of local environmental homogeneity. The short tufted graminoids (EG1) and graminoids with long internodes (EG3) were the two groups that varied most at the fine spatial scale. These groups were characterized by investment in vegetative spread and by the heaviest seeds among the 12 EGs of herbaceous species, thus their abilities to disperse over long distance would be limited. The variation of EG1 and EG3 at the fine scale suggests a possible effect of their weak dispersal abilities on their spatial patterning. However, such inference needs to be put in perspective because the consequences of dispersal mechanisms on the spatial patterning of species are very complex to elucidate from field data (Wichmann et al. 2008).

Looking back to our hypothesis, we first accept that the modification of environmental conditions by initial restoration have induced the spatial heterogeneity of EGs at different spatial scales. However, results showed that such exogenous spatial heterogeneity of EGs mainly occurred at the broader scales. Secondly, results highlighted a potential effect of limiting dispersal abilities as a source of endogenous spatial heterogeneity at least for two EGs.

Conclusions and perspectives for application

Both functional composition/diversity of plant communities and their spatial heterogeneity have been shown to be related to ecosystem functions such as productivity, rates of biochemical cycles, resistance to invasion or resilience after disturbance (Lavorel & Garnier 2002). Therefore, they must be considered when restoring damaged land or newly created land settlement as in the present case (Ehrenfeld & Toth 1997; Maestre et al. 2003; Seabloom & Van Der Valk 2003).

Here we studied the effects of the restoration plan in the spatial patterning of EGs. Because active restoration was applied irregularly the spatial heterogeneity of EGs was patterned at the very broad scale. However, a large amount of spatial variation of EG occurred at finer spatial scales and was independent from environmental variables. The trait-based approach allowed us to infer the response of the EGs to environmental factors, some assembly rules and a potential effect of dispersal in some EGs. Therefore, we conclude that after 25–30 years, the restoration plan succeeded in creating spatial heterogeneity of groups of plant species characterized by different set of ecological attributes.

However, it is likely that such spatial heterogeneity is not sustainable at the very broad scale over the long term, because local community successions in the harshest areas – and more precisely the amelioration of physical factors mediated by vegetation – might attenuate environmental variations originally induced by the restoration. The speed of these changes would obviously depend on the rate of community succession. In such a scenario, biotic constraints imposed by most competitive groups of species (long internodes graminoids and ultimately woody species) might prevail, whereas less competitive ones that tolerate stressful conditions of the physical environment might disappear. Additionally, an important part of the spatial variation that we quantified was not related to such environmental heterogeneity and took place at finer scale. If dispersal limitation in some groups is effectively responsible for vegetation heterogeneity on such scales, it is likely that with time natural colonization process will allow those groups to establish in most suitable areas. This would threaten groups that cannot establish and grow in severe abiotic conditions or in competitive situations. In that case, application of disturbance, through irregular mowing or grazing, might take place at finer scales than scales of variation due to environmental heterogeneity and would probably help in maintaining spatial patterning. However, these conclusions remain speculative and further studies are required to turn from inferences to knowledge. We argue that attention should be paid to two critical issues. First, the impact of the regional dynamic that has become better understood recently (Wichmann et al. 2008). Secondly, the complexity of process combination that could be treated by comparing patterns simulated by spatially explicit models and those observed in restored areas (Seabloom et al. 2005).

Homogenization of functional features of vegetation due to local community succession and natural colonization of competitive groups could also lead to a decrease of beta and gamma diversity of EGs at the study site. Maintaining the spatial heterogeneity of plant traits at multiple scales might be a future challenge for practitioners. Therefore, monitoring such patterns would be of primary interest. We argue that our results offer great potential for mapping functional composition and diversity over large land settlement areas by combining the niche modelling approach (Guisan & Thuiller 2005), community assembly modelling (Shipley, Vile & Garnier 2006) and spatial autocorrelation modelling of ecological data (Dormann 2007). Such maps could be used for assessing spatial structure and functional properties of vegetation over the entire sites, but also for managing targeted local ecological services. This would help practitioners to identify where they should focus their attention in areas that are placed under their responsibility, especially when the size of land settlements becomes large as in our study.


We thank the French Ministère de l’Ecologie, de l’Energie, du Développement durable et de l’Aménagement du territoire for funding, and the editors, B. Bolker and an anonymous referee, for their valuable comments on early drafts.