Ethics Statement

We obtained the necessary permit for surveying butterflies within the EU Natura 2000 network from Progresul Silvic, the organization officially entrusted with the custody of the protected area by the Romanian government. The survey procedure was approved beforehand by the ethics committee of Leuphana University Lueneburg.

Data Availability Statement

All data underlying the findings reported in this study are available from the Dryad Digital Repository (http://doi.org/10.5061/dryad.97s1k).

Study area and experimental design

The study area covered approximately 7,000 km² in the lowlands of Southern Transylvania, Romania (Figure 1). We followed the notion of a natural experiment [36], with randomised site selection in pre-defined strata at two levels: study villages and survey sites within villages.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Location of the study area with investigated village catchments in Transylvania, Romania.

The small letters indicate the village catchments illustrated for predictions in Figure 4 (a = Cincu, b = Granari, c = Viscri).

https://doi.org/10.1371/journal.pone.0103256.g001

To select study villages, we first allocated each raster pixel of the study area to different “village catchments”. These were calculated using a cost-distance algorithm in ARCGIS with the village centre as the reference point and the slope and the distance to the next village as the cost variables. Information about village locations was extracted from CORINE land cover data 2006 (http://www.eea.europa.eu/data-and-maps/data#c12=corine+land+cover+version+13&b_start=0&c17=CLC2006), and slope was calculated from the digital elevation model ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer). Topographically based village catchments were used instead of administrative boundaries because administrative boundaries were only available at the commune level (typically 3–5 villages). However, we found that the resulting polygons accurently reflected historical land use responsibilities. Second, we stratified village catchments along a gradient of terrain ruggedness and according to their protection status under the EU Birds and Habitats Directives. Terrain ruggedness was calculated as the standard deviation of the altitude of the catchment, and we used quantiles to classify ruggedness as either low, medium or high. Protection status of the catchments was either unprotected, SCI (Site of Community Importance) or SPA (Special Protection Area). Third, we randomly chose 30 villages, covering all combination of ruggedness and protection status (Table S1).

To select survey sites, we stratified the agricultural area within these 30 villages according to CORINE land cover as grassland or arable land and excluded other land cover classes. Within these strata, we spanned two gradients that we assumed sensitive to change in the future as a result of structural simplification, namely woody vegetation cover and heterogeneity. We estimated woody vegetation cover in a circular one hectare area based on classified 10 m SPOT data (CNES, ISIS programme). To assess heterogeneity, we used the standard deviation of 2.5 m panchromatic SPOT data within a one hectare circle. We assigned each hectare of the agricultural landscape to a combination of three classes of woody vegetation cover by three classes of heterogeneity. We distinguished low (0–5%), medium (>5–15%) and high (>15%) woody vegetation cover and used the lower, middle and upper third of percentiles to classify heterogeneity. Within these combinations, we randomly selected replicates for each cross-combination (except for the combination of high heterogeneity and low woody vegetation cover, which did not exist (Table S2)). In total, we selected 120 circular 1 ha survey sites, with 60 in grassland and 60 in arable land, and an average of four survey sites per village catchment. Notably, sites in arable land in this context were consciously placed not to represent only arable fields specifically, but rather to capture the whole range of conditions within the mosaic of arable land [38], including field margins and fallow land.

Data collection

Environmental data (explanatory variables).

We followed a multi-scale approach and included explanatory variables that could potentially explain butterfly distribution at the local (1 ha), context (50 ha) and landscape scale (i.e. village catchments, ranging from 430 to 4963 ha). An overview of all variables included in the analysis is presented in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Definition of environmental variables used in the study at three different scales and method of obtaining those. Abbreviations are used in Figure 2 and Table 2.

https://doi.org/10.1371/journal.pone.0103256.t001

At the local scale, we collected data on vascular plant species richness in eight randomized quadrants (1×1 m). We used cumulative plant species richness per site as an explanatory variable. We also calculated indices for heatload (after [42]) and terrain wetness as a measure of potential soil wetness, and included heterogeneity assessed by the spectral variance of SPOT data (see Table 1 for details). We calculated percent woody vegetation cover at local and context scales, and used CORINE land cover to calculate percent forest at the landscape scale. For the context and landscape scales, we calculated the terrain ruggedness as the standard deviation of altitude. We also quantified compositional or configurational heterogeneity of the different land covers grassland, arable land and forest as provided by CORINE land cover data. At the context scale, our chosen heterogeneity measures (Simpson index of land cover diversity, edge density) were correlated (r = 0.76). Hence, we included only edge density as an explanatory variable (following [7]). At the landscape scale, we used both edge density and the Simpson index of diversity and added the amount of pasture and forest per village catchment, based on CORINE land cover data. Variables on compositional and configurational heterogeneity were calculated using FRAGSTATS v4.2 [43] and all other variables using ARCGIS 10.1 (ESRI Inc., Redland, CA).

Analysis

We pooled all observed butterfly species and individuals from the four survey rounds for each survey site. First, we tested for differences in species richness and abundance between different levels of official protection by using Analysis of Variance (ANOVA). Second, we conducted a detrended correspondence analysis (DCA) to describe species composition and its relation to environmental variables. We used a permutation test to fit and test the correlation of environmental variables with the ordination.

Third, we used generalized linear mixed effects models (GLMMs) to assess effects of environmental variables on butterfly species richness. Beforehand, we tested the explanatory variables for collinearity (all r<0.7; Table S4; [44]). We log-transformed woody vegetation cover at local and context scales and heterogeneity at the local scale because these variables were highly skewed. All numerical explanatory variables were scaled to mean zero and unit variance. We included the variables listed in Table 1 to model species richness of butterflies. To test for a unimodal relationship in response to woody vegetation cover, we included a quadratic term of local woody vegetation cover. We furthermore expected that the effect of heterogeneity may differ between grassland and arable land and therefore included an interaction term between land cover type and heterogeneity. Grasslands are also interesting to look at separately because they are among the most species rich biotopes for butterflies in Europe [45]. We assessed the variance inflation factor (VIF) of the generalized linear model (GLM) and tested for spatial auto-correlation in the residuals. We included the village catchment as a random effect and corrected for overdispersion by adding an observation level random effect. We simplified the model by stepwise backwards selection retaining all variables with p<0.1. For GLMMs, significance levels are only approximations, hence many statisticians suggest using a significance level of p<0.1 [46].

Likewise, we modelled species richness of functional groups. To this end, we distinguished between species of low mobility (Bink’s mobility classes 1 and 2) and high mobility (Bink’s mobility classes 7, 8 and 9; [47]). Highly mobile species were Colias crocea, Pieris brassicae, Vanessa atalanta and Vanessa cardui. Low-mobility species were Brenthis daphne, Brenthis ino, Coenonympha glycerion, Cupido minimus, Euphydryas aurinia, Hamaeris lucina, Heteropterus morpheus, Lopinga achine, Melitaea britomartis, Melitaea diamina and Satyrium acaciae. As a third group we modelled the richness of grassland specialists, namely Euphydryas aurinia, Polyommatus coridon, Cyaniris semiargus, Lysandra bellargus, Phengaris arion, Cupido minimus and Erynnis tages [48].

We also modelled the abundance of individual species considered to be declining in Western and Northern Europe, but that are widespread or even increasing in Eastern Europe [48], [49], [50], [51]. We only used species that were common enough in the study area to obtain reliable models, namely Maniola jurtina, Coenonympha pamphilus, Polyommatus Icarus, Lycaena dispar and Glaucopsyche alexis. We performed all statistical analyses in R [52], using the packages MASS, ade4, vegan, gdata and lme4.

Both grassland and arable land have conservation value

Our findings revealed a high conservation value for butterflies of the small-scale farming system in the lowlands of Transylvania. Interestingly, butterfly species richness and abundance were similar in arable land and grassland. This is a surprising result and suggests a need to broaden the emphasis of conservation activities from grassland protection towards the maintenance of heterogeneous mosaic farmland, including cropland [38]. This is particularly important in the context of criticisms that the recent reform of the European Union’s Common Agricultural Policy, for example, falls far short of what is needed in terms of biodiversity conservation [53]. Throughout Europe, grasslands are considered most important for butterfly conservation (e.g. [8], [45]). Arable land, on the other hand, has received far less attention. In Western Europe, arable land has been found to support lower species richness and more homogenous butterfly communities than grassland [9], [54]. Our results indicate that this situation may be different in Eastern Europe, and that certain types of arable land can in fact support similar levels of butterfly diversity as grasslands. A possible explanation for the similar species richness in arable land and grassland in Transylvania may be spillover effects [23], which may be more likely in small-scale mosaics of land covers. The mosaic character of the landscape also could explain the strong overlap in butterfly communities between arable land and grassland.

Low-intensity landscapes provide important resources for butterflies

The fine-grained mosaic nature of arable land and the low-intensity nature of grassland in Southern Transylvania emphasize that low-intensity land use practices have major benefits for butterfly conservation. Semi-natural elements occur throughout the landscape, and are a likely reason why species richness is high throughout different land covers [22]. Furthermore, species richness of vascular plants can be high in field margins, which in turn may indicate high quality habitat for butterflies [55]. Consistent with the findings of Kumar, Simonson & Stohlgren [7], we found plant species richness strongly related to butterfly species richness. Currently, Transylvania contains some of the world’s most species rich areas for plants [56], which is partly linked to the low use of fertilizers [57]. Agricultural intensification, by contrast, would likely lead to increased use of fertilizers and hence reduced plant species richness [58], [59], [60]. Furthermore, intensification is typically associated with the use of fewer, high yielding crop varieties. Interestingly, many butterflies in Transylvania use the common crop Medicago sativa ssp. sativa (Alfalfa), a leguminous species that provides nectar and that we also observed to serve as a host plant for several butterfly species (e.g. Glaucopsyche alexis). Alfalfa is grown in small parcels, is primarily used as winter fodder for livestock, and may easily be lost as a result of intensification. However, high amounts of floral resources are critically important to maintain butterfly diversity. Similarly, woody vegetation offers important resources for butterflies, including shelter and space for thermoregulation [61]. At present, Transylvania contains many scattered trees and hedgerows, and we found that butterfly species richness responded positively to these structures at the local scale. By contrast, a large amount of woody vegetation at the landscape scale may lead to decreased species richness, probably due to a lack of open habitat.

Heterogeneity has a different effect in arable land than in grassland

We considered heterogeneity and woody vegetation cover at the local scale as two potentially important gradients describing the structure of the landscape. Interestingly, our results showed that the effect of local heterogeneity on species richness depended on land cover. In arable land, species richness increased with heterogeneity, supporting our hypothesis that small-scale farming benefits biodiversity by providing a range of different resources for butterflies. Notably, our land use class of “arable land” reflected the highly heterogeneous nature of traditional farmland, and included cropped areas as well as fallows and uncultivated field margins. These non-cropped areas are likely to be particularly important to maintain butterfly diversity in arable land. By contrast, in grassland, high heterogeneity was associated with reduced butterfly diversity. A possible explanation for this pattern is that heterogeneity of grassland may correspond to a higher degree of fragmentation of butterfly habitat, with likely negative consequences for species diversity [62]. Our study thus confirms that heterogeneity per se is not universally beneficial for species richness (see also [63]), although most work to date has focused on its positive effects (e.g. [64]).

The importance of considering multiple scales

To date, results from studies investigating multiple scales have been disparate and difficult to generalize [65]. We included three spatial scales in our study which we considered relevant for butterfly diversity and distribution. Our study revealed that all investigated scales affected butterfly community composition. Previous studies found local factors affecting butterfly species composition, with local heterogeneity in land cover being a good predictor for species composition in Canada [54], [64]. Butterfly species composition in Transylvania also showed a significant correlation with local factors, but was explained by heterogeneity and woody vegetation cover only at the two larger scales. Butterfly species richness also responded to variables at all different spatial scales, especially at the local scale, but also at the two larger scales (see also [9]). This suggests that local habitat conditions are particularly important, yet these cannot be considered in isolation from the surrounding landscape [33], [66].

Our models also showed that the different functional groups of butterflies were affected by variables from different spatial scales. For example, landscape heterogeneity appeared to benefit mobile species but not low-mobility species. Furthermore, we found that woody vegetation cover was related to species richness. Land abandonment induces natural succession, whereas intensification leads to loss of scattered woody vegetation, and both have negative effects on butterfly richness in the long term [67]. Both processes decrease structural heterogeneity, which is important for viable butterfly populations in agricultural landscapes. In our study, only Coenonympha pamphilus showed a unimodal relationship to local woody vegetation cover. For such low-mobile species, presence of woody vegetation is crucial for wind shield and thermoregulation. C. pamphilus is abundant in Transylvania, however its population state in other European countries is declining [68]. Habitat heterogeneity from different spatial scales, including the presence of woody vegetation, should be further investigated as possible key elements in landscapes to halt biodiversity loss in farmland.