Des aires protégées aux réseaux écologiques

Mesurer et cartographier la spécialisation des communautés à l’aide de données d’atlas

L’utilisation de la valeur moyenne de traits spécifiques des espèces présentes dans une communauté a récemment été proposée comme un indicateur écologique pertinent. Plus particulièrement, un Indice de Spécialisation des Communautés (CSI) reflétant la spécialisation moyenne des espèces présentes dans un assemblage donné a été proposé pour quantifier la réponse des communautés aux perturbations paysagères ou pour cartographier des zones d’intérêt pour la conservation. Cependant, cette approche n’a été testée que sur la base de données de suivi permettant d’évaluer précisément les variations d’abondance des espèces co-occurrentes. D’autre part, l’évaluation des niveaux de spécialisation des espèces nécessite des informations précises et quantifiées quant à leurs exigences, qui sont rarement disponibles. Ici, nous analysons dans quelle mesure des données d’atlas peuvent être utilisées pour calculer et spatialiser des valeurs de CSI. Pour cela, nous utilisons des données concernant la flore d’une région Méditerranéenne française afin d’évaluer la spécialisation des espèces sur la base de leur co-occurrence à l’échelle du paysage. Nous testons par ailleurs la relation entre la richesse des assemblages et la valeur de CSI ainsi que le lien entre CSI ou richesse et degré d’artificialisation du paysage. Nous observons ainsi que les assemblages les plus riches sont composés de nombreuses espèces spécialistes et que la richesse et le CSI sont maximisés à des niveaux de perturbation intermédiaires. Nos résultats suggèrent que l’estimation du niveau de spécialisation à l’échelle du paysage fournit un nouvel éclairage sur la distribution spatiale des espèces et des communautés. Par ailleurs, l’estimation du CSI avec des données d’atlas représente une approche complémentaire à la cartographie traditionnelle d’indices de diversité tant du point de vue de la conservation que du point de vue théorique.

INTRODUCTION

Searching the best indicator has been a long-lasting goal for ecologists. Consequently, many ecological indicators based on “species diversity” are now very popular, although their relevance was early questioned up to consider species diversity as a non-concept (Hurlbert, 1971). More recently, developing relevant indicators of biodiversity has become a scientific, political and societal issue of great importance (Balmford et al., 2005). Instead of searching the “best” indicator, authors have recognized that indicators are not “good” or “bad” but that their relevance depends on the question asked and on the data available. To assess the large-scale impacts of landscape degradation on communities, ecological metrics reflecting the dynamics of “loosers” versus “winners” within species assemblages was proposed as a promising approach (Devictor and Robert, 2009). In particular, the replacement rate of specialist species by generalists has been viewed as a direct signature of community response to large-scale habitat degradation, so-called biotic homogenization (Clavel et al., 2010). Indeed, one generally expects that specialists will benefit from stable and undisturbed habitats whereas generalists should be positively affected (Colles et al., 2009). Ideally, this process should be measured in tracking the replacements of individuals belonging to specialist species by those belonging to more generalists using large and standardized community monitoring. In practice, monitoring data are currently running for only few groups (mostly birds, butterflies and mammals) and presence-absence or distribution maps are often the only data available. Moreover, measuring species’ specialization is often difficult so that species are often simply categorized as being specialist or not. Indeed, species specialization has multiple components and should reflect species requirements on resources and/or habitats, which are difficult to measure for many species. Biotic homogenization was therefore mostly quantified at global or at national scales using crude classification of species into specialist versus generalist groups, or, on the contrary, using high-resolution data on detailed species requirements measured in controlled experiments (Devictor et al., 2010). Interestingly, (Fridley et al., 2007) have recently proposed a method to estimate species specialization that only requires presence-absence data. This method only assumes that species co-occuring with similar species are generally those found in similar habitats whereas generalists should be widely distributed across habitats and thus co-occur with many different species. In other words, the similarity in the identity of co-occuring species is taken in this Chapitre 6 – Mesurer et cartographier la specialisation des communautés 138 approach as a continuous proxy for species‘specialization. From this assumption, one can simply deduce a Species Specialization Index (SSI) for each given species using the identity of the species co-occurring with that species. This approach can be applied to any dataset where different species assemblages have been recorded in different locations. Beyond specialization, this algorithm also provides the number of plots where the species is found but also the average species richness of the plots where the species occurs. Then these characteristics can in turn be used to test interesting predictions. For instance, one expects generalist species to have large distributions (so called Brown hypothesis, (Brown, 1984). But whether specialist species tend to co-occur in rich or poor assemblages is less clear and depends on complex interacting processes. Once a SSI is available for each species, then a Community Specialization Index (CSI) of species assemblages can be calculated as the average of each SSI of species present in the assemblage (Devictor et al., 2008). One expects the CSI to be higher for species assemblages mostly composed of specialist species. Then, CSI can in turn be used as an interesting metrics complementary to more traditional indicators based on diversity. For instance, mapping the CSI provides a picture of spatial variation in the specialization level of communities, which can then be related to sources of disturbance or as a spatial guideline to identify sites of conservation interest. In particular, this approach should be useful in complex mosaic landscapes like in the Mediterranean region where heavily disturbed habitats coexist with stable habitats (Thompson, 2005; Blondel et al., 2010). Although atlas data have been used in widely different contexts and were proved to have many applications in ecology and conservation (Donald and Fuller, 1998 ; Muratet et al., 2009), whether one use atlas data to track biotic homogenization of species assemblages has never been proposed. Here, we use atlas data on plants for a Mediterranean region in order to investigate whether and how such data can be used to obtain a spatial distribution of CSI at the landscape level. To do so, i) we calculated SSI for each species; and ii) we investigated whether and how the species distribution was dependent on their SSI. At the assemblage level, we tested iii) the relationship between CSI and species richness and iv) the relationships between these two metrics and landscape disturbance.

METHODS 

The study region This study was carried out in the Languedoc Roussillon region (27 376 Km²) in southern France, which encompasses most of the Mediterranean region west of the Rhône valley (Fig. 1). The main landscape types which occur in this region are coastal landscapes with lagoons, marshes, cliffs and dunes, lowland garrigues. These habitats are often included in mosaic landscapes with cultivated areas, vast areas of vines, extensive upland limestone plateau areas, and hilly or mountainous landscapes on granite and schist in the southern tip of the Massif Central and the south-eastern Pyrenees. In the last 50 years profound modifications to the landscape of the region have occurred. First, extensive and rapid urbanization has occurred around towns and villages across the lowland plains and in conjunction with massive proliferation of coastal tourist resorts. Second, human population decline in rural areas has been accompanied by the abandonment of vineyards and grazing activity in many areas, changes which have set the scene for rapid natural reforestation of fields (IFEN, 2003) Species data Plant species data were compiled by the staff of the Conservatoire Botanique National Méditerranéen de Porquerolles which is in charge of the regional database of all plant species present in this region. It represents 3 792 species and a total of 551 488 occurrences data. Data were collected between the years 1611 and 2009 and come from all the naturalist inventories, herbariums, bibliography, reports and atlas available in the region. Data are georeferenced at three levels of spatial resolution: “municipality” corresponds to plant localities informed at the municipal level. “Locality” (Lieu-dit) corresponds to plant localities informed at the landscape/site level. “Point” corresponds to plant localities informed at the precise local GPS level. For the purpose of this study, we used only vascular plant species data collected since the year 1980. Indeed, earlier records are very sparse and with very vague localization. We obtained a database of 3 667 species for 420 659 occurrences data (6310 at the municipal level, 175 316 at the locality level and 239 033 at point level). All data were then combined and aggregated in a systematic grid of 5×5 km to define a standardized spatial scale resolution (Fig. 1). This grid was the best compromise between fine variation in species composition and number of data in each square. However, we also tested the robustness of our results to changes in the scale of aggregation with 2×2 km and 10×10 km grids. Chapitre 6 – Mesurer et cartographier la specialisation des communautés 140 Figure 1 – The study region and the distribution of the plant occurrences data (in front of the map are the grid cells of 5×5 km used for the aggregation). Measuring species and community specialization index First, an algorithm adapted from Fridley et al. (2007) was used to measure a Species Specialization Index (SSI) for each plant species. In brief (see Fridley 2007 for details and scripts), for each given species a random combination of sites (50 sites) in which this given species occurs is selected. Thus, the species occurring in less than 50 sites is eliminated from Chapitre 6 – Mesurer et cartographier la specialisation des communautés 141 the analysis. Among the 3 667 species, 1090 fulfilled this constraint. A similarity index is then calculated to reflect the degree of between sites species turnover in the 50 sites. The same operation is realized for 100 combinations of 50 sites for each species and the overall SSI of the given species is calculated as the average of the 100 similarity values obtained. Note that in this approach, similarity is always calculated for each species from a fixed number of sites (50). Thus, specialization for rare and common species is derived from combinations of species assemblages of equal sizes. Because they are expected to co-occur with similar species, specialists should have higher similarity values than generalists. Ecologists have used a large number of different measures of community similarity (also called beta-diversity or turnover) with different properties and meaning (Koleff et al., 2003). Here, we measured similarity using the average of pairwise ßsim calculated among sites (Baselga, 2010). For two sites, ßsim is a dissimilarity (or turnover) index given by ßsim = max(b,c)/(a+max(b,c)) where a is the number of species common to both sites, b is the number of species that occur in the first site but not in the second and c is the number of species that occur in the second site but not in the first. ßsim is, by construction independent from species richness (Baselga, 2010). We then used 1- ßsim to measure similarity. Note that other traditional way of measuring similarity between plots (e.g., in partitioning diversity in local, regional and among-site components) could also be used. However, an often common though unwanted property of these similarity metrics is to be correlated to species richness. In Fridley’s algorithm these other metrics tend to be highly sensitive to species from speciespoor habitats that have strongly skewed richness distributions (Manthey and Fridley, 2009). Once a SSI is obtained for each species, then a Community Specialization Index (CSI) is calculated for each square of 5×5 km as the average of the SSI values belonging to species present in this square (Devictor et al., 2008). CSI is therefore higher for species assemblages with higher number of specialist species (i.e. with high SSI) but is, by construction, independent from species richness. Measuring landscape disturbance We used an indicator of disturbance based on the compilation, for each site, of three kinds of human disturbance: road, urbanization and agriculture. Note that we use disturbance as a generic term without a priori on the negative or positive impact on species: some of these artificial landscape modifications can be positive for some species and negative for others. For road and urbanization, we used the “road” and “built-up” layers coming from the BD Chapitre 6 – Mesurer et cartographier la specialisation des communautés 142 TOPO®/RGE GIS database (IGN Institut Géographique National). For agriculture, we used the arable land, mixed agriculture, permanent cultures layers coming from the Corine Land Cover database. For each site, we calculated the proportion of disturbance elements within the site. Then a disturbance indicator was calculated for each site as the mean value of the normalized value (from 0 to 1) for each proportion. Data analysis We first focused on the characteristics of more or less specialized species. We tested whether the size of the species distribution was related to the species’ SSI (as expected from the Brown hypothesis) using linear regression. Similarly, we tested whether species with high or low SSI tended to be located in poor or rich assemblages. Second, we studied the relationship between species richness and CSI using General Linear Mixed Models (GLMM) accounting for spatial autocorrelation. In this model, CSI was considered as the dependent variable, and richness as a covariate. The spatial structure was added to this model as a random effect. The best spatial model (exponential) was selected after the study of semi-variograms as well as corresponding range and nugget (Pinheiro and Bates, 2000). Similar models were used to test the potential effect of landscape disturbance on species richness and CSI. We further looked for possible non-linearity of these relationships using quadratic terms or General Additive Models (GAM). GAMs are extensions of classical linear models in which the predictor is not restricted to be linear but is the sum of smoothing functions chosen in many different function types but constrained by a fixed degree of freedom (Wood and Augustin, 2002). We finally mapped the landscape disturbance, the CSI and the species richness of each grid. In order to discuss the spatial distribution of such parameters, we used the map of the biogeographic ecoregions provided by (Quezel et al., 2004)

DISCUSSION

Using atlas data only, we were able to distribute species along a continuous gradient of specialization. An important point however is that using this approach, “specialization” is not equivalent to niche breadth as generally estimated using species response to environmental conditions. Instead, this approach gives an operational and quantitative metric for the “interaction milieu” proposed by (McGill et al., 2006). Indeed, what is quantified for each species is what is “seen” in terms of identity of other co-occuring species. This approach can also be viewed as a measure of specialization that integrates both the Grinnelian niche (i.e., the niche taken as species response to environmental gradients) and the Eltonian niche (i.e., the niche taken as species’ impacts on other species) (Devictor et al., 2010). However, the ecological meaning of this interaction-milieu is clearly scale-dependent. While plant species interacts at very local scales, the co-occurrence patterns at larger scales should be more influenced by regional and dispersal processes. In our study, specialization reflects neither fine and species interactions nor specialization for well-defined habitats. Rather, SSI reflects the tendency of species to occur in different landscapes of different species composition. Chapitre 6 – Mesurer et cartographier la specialisation des communautés 146 In other words, a species can be considered as specialist at a landscape level although the same species could be a generalist for habitats within landscapes. This raises the question of the scale-dependency of specialization (Devictor et al., 2010). The fact that the SSI was robust to change in the scale considered (2×2 km, 5×5 km or 10×10 km) can be explained by the poor landscape variation between such scales. In our case, species whose distribution overlaps some heterogeneous landscapes (here characterized at the 5x5km plot level) will have a low SSI score according to Fridley’s algorithm (those species will tend to co-occur with different species). On the contrary, species localized into areas of low between-plot variations in species composition will have high SSI values. Therefore, although the SSI and CSI cannot be interpreted as reflecting fine-scale specialization, these metrics can be used to reflect the variation in between-plot species composition. Here further investigation would be necessary to examine the spatial relationship between CSI and beta-diversity (McKnight et al., 2007). Note that, at such scale, we cannot not say if the SSI reflects also a specialization process due to the landscape as an environmental filter. We found that more generalist species tended to be more widely distributed but localized in poorer assemblages than specialists. When specialization is measured as species niche breadth, negative relationships between specialization and species ranges have been documented for many taxonomic groups at very different scale (Gaston, 2003). Here we found that species occurring in different landscapes (i.e. with low SSI) are also those with larger distributions. More interestingly, we also found that specialist species tend to occur in richer assemblages. By construction, species with high SSI are those occurring in areas with more homogeneous landscapes but our results suggest that, in addition, these landscapes tend to have high species richness. At the species assemblage level, the relationship between CSI and species richness was also positive. These patterns do not result from the way specialization is measured but reflect an emergent property of species with low and high SSI. These results suggest a non-random spatial distribution of specialist and generalist species across the study region so that richer assemblages were those that concentrated more specialist species and inversely poorer assemblages were those that concentrated more generalist species. In fact, considering the meaning of our specialization index, this non random distribution can be explained by the fact that species occurring in the same ecoregion (e.g. Pyrenean) will tend to have the same SSI. Indeed, the comparison of figure 4c and figure 4d visually shows Chapitre 6 – Mesurer et cartographier la specialisation des communautés 147 that there is a relationship between spatial distribution of the CSI values and ecoregion delimitation provided by (Quezel et al., 2004) (although this spatial congruence remains to be tested statistically). The non-random spatial distribution of the CSI value mirrors this nonrandom spatial distribution of plant species. Hence, our approach allows to characterize these ecoregions, known as coherent in their species composition and ecological functioning, in terms of both the local richness of the landscapes they are composed of, and the variation in between-landscape species composition. As an example, according to the figure 4, we can deduce that in the zone “x”, situated in the lowland plain, the local species richness is generally high but the species turnover is low. On the contrary, the zone “y”, situated in the Pyreneans, is composed of relatively lower local species richness but with higher spatial variation in terms of species composition.