The Grands Causses (average altitude 800 m) are located south of the Massif Central (southern France) and include the Causse de Méjean, the Causse noir, the Larzac, and the Causse de Sauveterre. Only the latter three are part of the ‘Parc naturel régional des Grands Causses’, the Causse de Méjean being a part of the ‘Parc national des Cévennes’ [12,13]. The Grands Causses, sparsely populated, are wide, dry calcareous table lands cut by deep canyons (karstic system) characterized by harsh winters and subject to low temperatures and dry summers with strong thermal amplitude and a vegetation adapted to drought. They contain vast settlements of steppic lawns (causse) and woodland pines, in particular in their eastern and septentrional zones, while pubescent oak spontaneously colonizes the most exposed slopes.

Ant sampling was carried out in May 2005 around the village of Saint-Léons-en-Levezou (department of the Aveyron) by 35 people (the authors plus graduate students from the University of Toulouse-3). We sampled a causse (centre of the parcel: 44°13′N, 03°01′E; altitude: 850 m) characterized by shrub and herbaceous plant species, such as Amelanchier ovalis, Bromus erectus, Buxus sempervirens, and Rubus fructicosus, an oak grove (44°13′N, 02°58′E; altitude: 770 m) dominated by Quercus pubescens, and a pine forest of Pinus sylvestris (44°11′N, 03°00′E; altitude: 780 m). The distance between (1) the causse and the oak grove was 2.2 km, (2) between the causse and the pine forest, 4.5 km, and (3) between the oak grove and the pine forest, 4.3 km. The leaf-litter layer was thick in the oak grove and pine forest (up to 12 cm), but much more limited in the causse (up to 3 cm).

Based on recommendations presented in previous studies in the tropics [7,14], we used four sampling methods: Winkler extractors, pitfall traps, baiting and manual collection. Sampling was conducted in the centre of each of the three distinct vegetal formations on six adjacent transects 125 m long and 25 m wide (1.875 ha), each divided into five 25 m ×25 m parcels (30 parcels in total) where the four sampling methods were used. All methods were applied randomly in each parcel, leading to 30 replicates for each environment. The total time spent sampling with all methods was about 400 h (all-pooled).

For manual collection, four observers walked randomly around each parcel during 20 min (1 h 40 min of sampling effort per transect or 10 h per vegetal formation) and collected ants foraging on trees or on the ground, plus colonies nesting in dead wood or under rocks. Baits (one piece about 2 cm³ of tinned sardines; 30 baits per parcel) were placed on absorbent paper (20 cm×20 cm) on the ground for 3 h [3]. Pitfall traps (30 per parcel) consisted of plastic 75-mm-diameter cups (a size recommended by Bestelmeyer et al. (2000) [7] to optimize trap efficiency) filled to one-third of their height with slightly (5%) soapy water (neither attractive nor repulsive to ants). They were placed randomly, but more than 1 m from the edges of each parcel to avoid interference between parcels, and the vegetation around them was not cleared. The traps were removed 24 h later and the ants collected. Finally, we sifted the litter from a 1-m² quadrate in each parcel, and placed it in a numbered bag. In the laboratory, the sifted litter was placed into the inner mesh bag of a Winkler extractor. This bag was then introduced into the outer, funnel-shaped canvas bag whose end is connected to a cup containing 80° ethanol. The extractor was then hung in a well-aired environment, out of direct sunlight. In response to the increasing levels of dryness, the arthropods, including ants, crawl out of the inner mesh bag, and then fall into the ethanol [7]. All the resulting samples were first coded and then put into small plastic tubes filled with 80° ethanol. Later the ants were mounted in order to constitute a reference collection and identified according to Bolton [15,16]). Voucher specimens were deposited at Micropolis (Saint-Léons-en-Levezou, France), the ‘Laboratoire d’évolution et diversité biologique', University Toulouse-3 (France), and the ‘Laboratório de Mirmecologia’, Cocoa Research Center CEPEC/CEPLAC (Ilhéus, Bahia, Brazil).

Species richness (number of collected species) and species occurrences (number of times that a given species was collected in all samples using the same method) were compared between vegetal formations and sampling methods. To estimate and compare the efficiency of the sampling methods, rarefaction curves (cumulated numbers of species according to the number of occurrences for each species) were calculated for each vegetal formation, each method, and all methods combined. Because comparing sampling techniques on the basis of number of samples has little meaning (no two methods having the same probability of capture) and because the number of individuals is not a reliable parameter in ant studies (data can be distorted when ants are collected close to a nest, an ant trail or a site where ants recruit nest mates), we took into account only species occurrences. Matrices of species occurrences (presence/absence) according to the area where the ants were sampled and to sampling methods were treated using the EstimateS 7.1 software [17] with 500 randomizations of the sampling order without replacement (see [18]).

To sort the samples (i.e. type of environment and sampling methods) according to the ant species they host (see Appendix A), we used the Self-Organizing Map algorithm (SOM Toolbox version 2 for Matlab^®; http://www.cis.hut.fi/projects/somtoolbox/documentation/; see [19,20]). Combining clustering and functions, the SOM is a powerful tool used to visualize high-dimensional ecological data; more specifically, this technique is particularly relevant to patterning species assemblages using data on species occurrences. The SOM algorithm is an unsupervised learning procedure that transforms the multidimensional input data constituted here by 33 nodes (one per ant species) connected to the 273 samples into a two-dimensional map (shown as 80 hexagonal cells) subject to a topological constraint. The SOM plots the data so that similar samples are found together on the grid and, inversely, very different samples are far from each other. In this study, we used the ordination process detailed in [21] where the training was broken down into an ordering phase (the first 3000 iterations), and a fine-tuning phase (7000 iterations). At the end of the training, a species assemblage is known for each virtual site and the ‘best-matching unit’ (BMU) is determined for each sample that is set in the corresponding hexagon of the final SOM. To assess whether the map has been properly trained, the topographical error was used as a measure of map quality [20]. The intensity of the connection between input and output layers calculated during the learning process can be considered as the probability of occurrence of each ant species in the area concerned. It is represented on the SOM using a shaded scale, allowing us to analyse the effect of each variable (ant species) on the patterning input dataset (environments and sampling methods). At the end of the learning process, in order to detect the cluster boundaries on the map, the unified-matrix (U-matrix) approach [22] was applied. The U-matrix displays the distances between the virtual sites and provides a landscape formed with light plains separated by dark ravines. The sites in the plains are similar (for ant species assemblage) and clusters become clear.

Variations in the environment (e.g., vegetal formations) or the sampling method frequencies among the six clusters resulting from the SOM were analyzed using a generalized linear model [23] with a binomial error (proportion data, χ2 statistics) by using sample frequencies in each cluster as denominator vector (e.g., reference). First, models with the six clusters were fitted to the data (e.g., complete models). Second, simplified models (related to the environment or the sampling methods) grouping certain clusters were adjusted to the data and only those which were statistically similar to the complete model (χ2 test) were retained (e.g., simplified models).

All vegetal formations and methods pooled, we recorded 33 ant species representing 13 genera and three subfamilies: Dolichoderinae (Tapinoma), Formicinae (Camponotus, Formica, Lasius and Plagiolepis), and Myrmicinae (Aphaenogaster, Temnothorax, Messor, Myrmecina, Myrmica, Solenopsis, Stenamma, and Tetramorium) (Appendix A, Table 2). We noted 29 species in the causse, 22 in the oak grove, and 17 in the pine forest. All three vegetal formations were suitably sampled because the height of the curves tends to decrease (i.e., the number of new species falls as the number of samples increases; Fig. 1). Thus, for an equal sampling effort using identical sampling techniques, each vegetal formation seems to have been globally sampled with the same effectiveness.

Each sampling technique enabled us to collect 20 species or more (manual collection: 27; baits: 20; pitfall traps: 21; Winkler method: 25). If we compare the passive sampling methods (manual collection was not taken into account in the comparison because it is an active method), we note that none of them allowed us to collect all of the species, regardless of the study area. Because the curves are not stabilized, each technique resulted in the undersampling of the three vegetal formations, indicating that a greater sampling effort would have contributed to the addition of new species. The Winkler extractor best sampled the three vegetal formations, particularly in terms of species occurrences (Fig. 2), but was less efficacious on the causse, where there is little leaf-litter.

We tried to determine the optimal combination to best sample each vegetal formation by combining two, three or four sampling methods (Table 1). For the causse, the combination of only three methods (i.e., manual collection, Winkler extractors, and pitfall traps) allowed all of the ant species to be captured, highlighting the fact that baiting did not add anything in this case. For the oak grove and the pine forest, only the combination of the four methods allowed all ant species present to be captured. Consequently, overall the four methods are complementary, each contributing a component to the total data.

The 80 units the SOM trained with species presence/absence data had a topographical error of 0.0001. This map thus preserved the typology of the input data well, and was relevant for subsequent interpretation (clusters A–F, Fig. 3a and b). After training the SOM with species data, the U-matrix helped to tease out six clear clusters. Thus, samples were classified into six subsets according to ant assemblages; i.e. according to the different distribution patterns, which were characteristic of each species. To reveal the relationships between environments, sampling techniques, and ants, we subsequently introduced the environmental types and sampling techniques into the SOM previously trained with the 33 ant species (Fig. 4).

Among the most frequent species, one can distinguish those represented in only one cluster from those, more generalist, represented in two clusters or more (Appendix A; Fig. 3a and b), such as Myrmica sabuleti or M. ruginodis. The species represented in one cluster also characterize a vegetal formation (bio-indicators): Formica cunicularia and Tetramorium caespitum for the causse; F. gagates for the oak grove; and Lasius fuliginosus, Temnothorax luteus and Myrmecina graminicola for the pine forest. Other less frequent species are associated with these numerically dominant ants, thus constituting species assemblages more or less specific to each vegetal formation; their distribution is ‘influenced’ by the environment.

When the distribution of environments and sampling techniques was represented on the SOM previously trained with ant species (Fig. 4), all environments and techniques were present in the six clusters. The causse is mostly characterized by the cluster F (Tetramorium caespitum, the most frequent species), while clusters A and B (Myrmica ruginodis) characterize both the oak grove and the pine forest, and cluster C (Formica gagates) permits the oak grove to be differentiated from the pine forest. Concerning the sampling techniques, cluster A rather characterizes the Winkler extractors, the group constituted by clusters B, C and D rather characterizes pitfall traps, while baiting samples are dispersed throughout the entire chart, but are greater for cluster F. We therefore noted both an ‘environmental influence’ and a ‘sampling method influence’, so that the sampling can be biased and it is more correct to speak of ‘species assemblages’ than of ‘communities of species’ to characterize the set of ant species recorded in each vegetal formation.

Significant differences between the clusters represented in the SOM were found for the environment (χ172=252.4; P<10−5) and the sampling technique (χ232=248.67; P<10−5) (Fig. 5). Simplifying the models made it possible to reduce the number of factors for each variable. Five groups were obtained for the environment (χ142=250.5; P<10−5) by grouping two clusters (A and B), and only three groups for the sampling methods (A; B+C+D; E+F; χ112=233.2; P<10−5). These simplified models are not statistically different from the corresponding complete models (Δχ2:χ32=1.9; P>0.10 for the environment; Δχ2:χ122=15.51; P>0.10 for the methods).