The silver-spotted skipper butterfly Hesperia comma is a habitat specialist; in Britain, it is confined to calcareous grassland, laying eggs on a single host plant species, sheep’s fescue grass Festuca ovina (Thomas et al. 1986). Only short tufts (<10 cm) of F. ovina are selected for oviposition, restricting the butterfly to sites with intermediate to high levels of grazing (Thomas et al. 1986). In H. comma, both fecundity and egg-laying microhabitat availability are temperature-dependent: females show increased egg-laying rates at higher temperatures, and select warm microclimates, next to patches of bare ground, for oviposition (Davies et al. 2006).
Hesperia comma reaches its northern range limit in Britain (Fig. 1), where historically, populations have been confined to hotter south-facing slopes (Thomas et al. 1986). However, over the past 30 years, warming summer temperatures have broadened the range of microhabitats suitable for egg-laying (Davies et al. 2006) and permitted the colonization of cooler north-facing habitats (Thomas et al. 2001). This, together with more widespread grazing from rabbits and livestock (Thomas & Jones 1993), has increased the availability of suitable breeding habitat, catalysing a range expansion from fewer than 70 populations in 1982 to over 250 by 2000 (Davies et al. 2005; Wilson, Davies & Thomas 2010). However, the fragmented distribution of these habitat patches and the species’ limited dispersal abilities (Hill, Thomas & Lewis 1996) have constrained the rate of this expansion (Wilson, Davies & Thomas 2009, 2010).
Figure 1. Maps of recorded changes in Hesperia comma site occupancy between 2000 and 2009 across south-eastern Britain, with chalk geology shown in white.
Download figure to PowerPoint
In 2009, we conducted the fourth UK national survey of H. comma’s distribution (for previous surveys in 1982, 1991 and 2000 see Thomas et al. 1986; Thomas & Jones 1993; Davies et al. 2005). As H. comma is univoltine, this represents nine generations since the previous survey (Davies et al. 2005). The search was restricted to five main habitat networks in south-east England (in the counties of Kent, Sussex, Surrey, Hampshire, and the Chiltern hills), which encompass the majority of H. comma populations in Britain (Davies et al. 2005). As in previous surveys, suitable habitat patches were defined as any unimproved chalk grassland containing more than an estimated 5% cover of F. ovina plants <10 cm tall, and neighbouring patches were defined as separate if their nearest perimeter points were divided by at least 25 m of unsuitable grassland, or a woodland or scrub barrier. All habitat patches meeting these criteria within a 30 km radius of known populations were surveyed; this radius is considered sufficient to detect all new populations because it exceeds the maximum recorded colonization distance over an 18-year period (1982–2000; Davies et al. 2005), and because the species was not recorded between 2001 and 2008 more than 13 km from populations known in 2000 by the British ‘Butterflies for the New Millennium’ recording scheme (R. Fox, pers. comm.).
Once located, the perimeter of each patch was mapped using a hand-held Global Positioning System (GPS; accuracy <±10 m) and later digitized using ArcMap software (ESRI 2009). Habitat characteristics that could affect H. comma occupancy were recorded, including the total area of the patch in hectares, the percentage cover of bare ground and vegetation <10 cm tall and proportional host plant cover in vegetation <10 cm (Table 1). Patches were then searched for the presence of H. comma, based on either observation of adults or timed egg searches.
Table 1. Explanatory variables and the analyses they are used in: survival (S), colonization (C), and/or population density (P)
|Variable name||Symbol||Description||Spatial scale||Analyses|
|Area <10 cm||AR||Effective area (ha) of turf <10 cm tall||Local||S, C|
|Host plant cover||HO||Proportional coverage (%) of Festuca ovina in turf <10 cm||Local||S, C, P|
|Bare ground cover||BG||Proportional coverage (%) of bare ground in patch||Local||S, C, P|
|Solar index||IN||Incident solar radiation at midday in mid-August||Local||S, C, P|
|Macroclimate||MC||Mean August maximum temperature over 9-year period||Local||S, C, P|
|Direct connectivity||CD||Connectivity to occupied sites||Landscape||S, C|
|Indirect connectivity||CI||Connectivity to initially unoccupied sites, weighted by connectivity to occupied sites||Landscape||S, C|
|<10 cm cover||TE||Proportional coverage (%) of turf <10 cm tall in patch||Local||P|
|Population age||AGE||Categorical variable indicating whether population was founded after 2000 (‘young’) or prior to 2000 (‘old’)||Local||P|
For patches visited in favourable weather during the flight period, adult densities were estimated using transect walks (Pollard & Yates 1993). Using weekly H. comma counts from UK Butterfly Monitoring Scheme (BMS) transects, we calculated the proportion of peak abundance on the day each transect was walked. We then divided observed abundance on survey transects by proportion of regional peak to estimate peak density at each site, setting a minimum proportion of 20% to avoid excessively large density adjustments.
The area of each patch (ha) was calculated using digitized patch perimeters in ArcGIS. By multiplying patch area by proportion of the patch with vegetation <10 cm tall, an estimate of effective breeding area was obtained (Table 1).
To assess landscape-scale gradients in macroclimate during the adult flight season, we calculated the mean August daily maximum temperature for each site for two separate periods (1982–1991 and 2000–2009), using the UKCP09 5 km resolution gridded observation data set from the UK Meteorological Office (Perry & Hollis 2005). To assess patch-scale differences in microclimate, we calculated incoming solar radiation as a function of aspect and slope: we applied the ‘hillshade’ function to a 5-m resolution digital terrain model in ArcMap (Intermap Technologies 2007; vertical accuracy ±60 cm), using solar azimuth of 180° and altitude 60° (equivalent to the maximum solar radiation in south-eastern UK during mid-August), and extracted the median solar index for each patch using the spatial analyst tool (ESRI 2009).
As a measure of potential immigration into each patch, we calculated functional connectivity using Hanski’s connectivity index:
Where i is the focal patch and j all other patches, which have area Aj and are separated from i by distance dij (Hanski 1999). Here, Aj is effective area <10 cm (ha) of patch j, and dij edge-to-edge distances between patches i and j (km), both based on 2009 data. α (a negative exponential dispersal kernel) and b (a scaling function for patch emigration) are estimated from a previous study (Wilson, Davies & Thomas 2010). In the original formulation, pj denotes the presence or absence of the butterfly at site j and is here used in two different ways. ‘Direct connectivity’ calculates connectivity to occupied patches (pj = 1), discounting unoccupied patches (pj = 0). ‘Indirect connectivity’ calculates connectivity to unoccupied patches, weighting those patches by their direct connectivity scores (i.e. taking area of patch j as AjSj). Therefore, whilst direct connectivity is a measure of the probability of colonization in a single dispersal event from currently occupied patches, indirect connectivity is a measure of the probability of colonization in two dispersal events using currently unoccupied habitat as a ‘stepping stone’. Both connectivity measures were calculated for occupancy patterns in 3 years: 1982, 2000 and 2009.
Analyses were conducted in two stages. First, we examined whether establishment, survival and density of H. comma populations from 2000 to 2009 could be explained by local and regional factors (Table 1). Second, we used the best models to predict the probabilities of survival and colonization across all patches for two time periods (1982–1991 and 2000–2009). All models were fitted using generalized linear modelling (Crawley 2007) in r 2.12.2 (R Development Core Team 2011).
Survival models considered existing populations in 2000, with a binary response of ‘survival’ or ‘extinction’ by 2009; colonization models considered unoccupied patches in 2000, with the response as ‘colonized’ or ‘uncolonized’. In addition to present information from the 2009 survey, colonization analyses included temporary colonizations based on additional information from surveys conducted by the authors in 2001 and 2002, and records from 2001 to 2008 from Butterfly Conservation, the organization which coordinates the UK butterfly distribution monitoring scheme. We considered the butterfly as absent from any site not surveyed in 2000 that was more than 5 km from an occupied patch. Population density analyses focussed on density data from the 2009 survey.
We assumed a binomial error distribution for survival and colonization models. Population density models assumed a Tweedie error distribution, a compound poisson-gamma distribution family suited to analysing positive continuous data with exact zeroes; parameters were estimated using the ‘tweedie’ package (Dunn 2010).
Explanatory factors included local and landscape variables (Table 1). Local variables comprised solar index and habitat characteristics (Table 1; see also Survey methods). A squared bare ground term was included because intermediate levels of bare ground are likely to be optimal for this species. We used habitat data gathered in 2000 in survival analyses, but because these data were not available for all unoccupied patches in 2000, we used 2009 habitat data in colonization analyses. Landscape variables included macroclimate, and direct and indirect connectivity (Table 1).
Population density analyses considered only local variables from 2009 and macroclimate (Table 1). Density analyses used proportion of turf <10 cm instead of effective area <10 cm, because the former investigates how density increases with breeding habitat independently of patch size. An additional variable, ‘age’, was also included, which indicated whether a population was present in 2000 (‘old’) or had been founded after 2000 (‘young’).
Data were checked for spatial autocorrelation (Beale et al. 2010) using Mantel tests with the ‘vegan’ package (Oksanen et al. 2011). There was no evidence that population density estimates were spatially correlated (r = −0·043, P > 0·05). We expected some spatial autocorrelation in survival and colonization data because of dispersal between neighbouring patches, as explicitly modelled by connectivity terms. Therefore, to test for autocorrelation over and above that accounted for by these covariates, we conducted mantel tests of the residuals from models including solely the two connectivity measures. Although there was no significant autocorrelation in colonization model residuals (r = −0·0080, P > 0·05), there was some sign of autocorrelation remaining in survival residuals (r = 0·24, P < 0·01). Consequently, we fitted mixed models including a random effect of ‘grid square’, classifying patches into spatial blocks of (i) 5 km and (ii) 10 km squares, using the ‘lme4’ package (Bates & Maechler 2010). However, top model sets were similar and effect sizes were within standard error bounds of those in the original models, indicating that spatial autocorrelation was not seriously affecting model parameters.
For each response variable, we fitted all possible combinations of linear terms, with no interactions, and used the Akaike Information Criterion, adjusted for small sample size (AICc; Burnham & Anderson 2002) to rank models. To obtain our best model sets, we selected models that were within six AICc units of the top-ranked model (Richards 2005), excluding models with simpler, higher-ranking nested variants (Richards 2008). This procedure guards against the selection of over-parameterized models whilst maintaining a high probability of selecting the true best model (Richards 2008).
To further explore which variables best explained survival and colonization, and to examine how variance was partitioned among them, we conducted hierarchical partitioning (Mac Nally 2002), implemented with the ‘hier.part’ package (Mac Nally & Walsh 2004). Hierarchical partitioning investigates the average change in fit (here, log-likelihood) between equivalent models with and without a given variable X, to assess the explanatory power of X independently of other terms (Ix). By randomizing the data set 1000 times and recalculating Ix, we estimated the probability of obtaining a value of Ix equal to or greater than that observed by chance, allowing us to assign statistical significance to the explanatory power of each variable (Mac Nally 2002). In addition, the average effect of other variables on the explanatory power of X (‘joint’ effects, JX) can be calculated; their direction indicates whether other variables are acting additively (positive), increasing the variation explained by X, or suppressively (negative), sharing variation with X (Mac Nally & Walsh 2004).
Following model selection, we applied our best model sets to predict the probabilities of colonization and survival across the metapopulations. Predictions were based on distribution data from either 1982 or 2009; in both cases, local variables (Table 1) were based on 2009 data because habitat information was not available for all patches in 1982. We used model-averaging (Burnham & Anderson 2002) implemented by the ‘AICcmodavg’ package (Mazerolle 2010) to obtain a single predicted survival and colonization probability for each patch. Doing so incorporates model selection uncertainty whilst weighting the influence of each model by the strength of its supporting evidence (Burnham & Anderson 2002).
To identify how survival and colonization limitations varied across the British distribution of H. comma, we classified each patch into one of four categories, based on whether they were primarily limited by survival, colonization, both (‘marginal’ habitat), or neither (‘supported’ habitat).