Does Landscape Structure Affect the Presence of Woodland Specialist Pollinators in Farmland? Implications for Agri-Environment Scheme Design

Abstract

Some pollinator species found in agricultural areas are strongly dependent on surrounding areas of natural or semi-natural habitats to nest and/or forage. Landscape structure has been shown to influence pollinator communities and understanding how landscape structure affects farmland pollinators can improve Agri-Environment Schemes (AES). This study explored how landscape metrics affect the presence of pollinators associated with woody vegetation in farmland in the Republic of Ireland. Two study regions were selected, and pollinators were collected using pan traps placed in farm linear features. Hoverfly and bee species were selected based on their body size and association with woody vegetation. Relevant landscape structure metrics were extracted from around each trap and used to develop explanatory models for the abundance of pollinators. The total abundance of target species was relatively low but correlated with three explanatory variables: the connectivity of the linear feature to woodlands; the distance from the trap to the closest woodland; and edge density. Hoverfly and bee abundance data, when analysed separately, showed significant differences within regions. Results seem to indicate that incentivising the connectivity of farm linear features to surrounding woodland patches and increasing optimal habitat availability in agricultural landscapes could benefit woodland specialists. This information is helpful to improving AES design.

Introduction

Pollination is an important ecosystem service underpinning agricultural production and sustainability (MEA 2005; Winfree et al. 2008; IPBES 2016). According to Klein et al. (2007), 75% of worldwide food crop species benefit from animal pollination and almost 90% of wild flowering plants species benefit, directly or indirectly, from the transfer of pollen by animals (Ollerton et al. 2011). In temperate regions, bees (Hymenoptera), hoverflies and other anthophilous flies (Diptera), butterflies (Lepidoptera) and beetles (Coleoptera) are the most important pollinators (Vanbergen and the Insect Pollinators Initiative 2013; Rader et al. 2016).

Substantial losses of pollinators have been reported in many regions of the globe and for many European countries it is well documented (e.g. Potts et al. 2010). Pollinator declines are especially high in regions characterised by intensive agriculture (Chagnon 2008; Connelly et al. 2015; Kovács-Hostyánszki et al. 2017). Agricultural crops and wildlife (e.g. wild plant species) dependent on pollinating insects are therefore at risk (Kearns, Inouye and Waser 1998; Biesmeijer et al. 2006; Betts et al. 2019). Abundance and diversity of pollinators influence the delivery of pollination services (Winfree et al. 2007, 2008). A more species-rich community of pollinators has an increased diversity of functional traits, which can also be beneficial for production (Dainese et al. 2019; Pfeiffer et al. 2019; Woodcock et al. 2019) and provides insurance for future changes in cropping systems (Yachi and Loreau 1999).

Landscape structure and diversity are important factors for maintaining pollinator density and species richness within farming areas (e.g. Bianchi et al. 2006; Hendrickx et al. 2007; Kennedy et al. 2013). Landscape structure is defined as the pattern of a landscape, which is determined by the number of different types of use (composition), and also by its configuration, i.e. the size, shape, arrangement and distribution of individual landscape elements (adapted from Walz 2011 and Mitchell et al. 2013). Landscape metrics include the density and area of landscape elements/habitats, heterogeneity, evenness, fragmentation and connectivity (Riitters et al. 1995; Walz 2011).

Pollinator species found in agricultural areas can be divided into two main groups: a) species that nest and forage in farm habitats (grasslands, orchards, embankments, hedgerows, etc.) and b) species that are dependent on surrounding areas of natural or semi-natural habitats (woodlands, wetlands, etc.) to nest and/or forage.

The spatial organisation of habitats, such as their fragmentation or connectivity in a landscape, also seem to influence survival and dispersal capacity of many pollinating insects (Hadley and Betts, 2012; Viana et al. 2012; Boscolo et al. 2017; but see Fahrig 2017). For example, patch isolation can influence species richness and density of butterflies (Öckinger and Smith 2006; Herrault et al. 2015); distance from natural habitats containing suitable nesting sites can affect pollinator richness and abundance (Kremen et al. 2004; Ricketts et al. 2008; Jauker et al. 2009); connected hedgerows facilitate the movement of bumblebees (Cranmer et al. 2011); and the connectivity of these linear elements to forest patches promote the abundance of hoverflies in farmland (Haenke et al. 2014). In turn, the degree of isolation of natural areas has a negative effect on this group of pollinators (Herrault et al. 2015) and functional connectivity (see for example Taylor et al. (2006) for definition) has been reported to affect species richness and abundance of flower-visiting bees (Boscolo et al. 2017).

The effects of landscape structure vary depending on species mobility, foraging behaviour and habitat requirements (Steffan-Dewenter et al. 2002; Zurbuchen 2010; Popov et al. 2017). Thus, the response of organisms to landscape should be investigated at the appropriate scale (Eigenbrod 2016). For example, as pollinators' body size decreases, what would be considered a 'local-scale factor' for highly mobile species might be considered a 'landscape-scale factor' for low-mobility species. Studies report that small-bodied, solitary bee species are more bound to semi-natural habitats because they are less able to disperse, exhibiting shorter foraging distances (Brosi et al. 2008; Carrié et al. 2017). Therefore, what is considered a fragmented or connected habitat patch will depend on the flight distances that characterise the species (see Wright et al. 2015).

To support the abundance and richness of habitat specialists within agroecosystems, a better understanding of the impact of landscape structure is required (Popic et al. 2013). This understanding can be used to improve AES design (Senapathi et al. 2017) and support the objectives of the EU Biodiversity Strategy for 2030 (European Commission 2020a).

This study aims to understand the landscape-related factors (compositional and configurational) that influence the presence of woodland pollinator species in farmland in Ireland. The study focused on species that have lower (predicted) dispersal distances due to small/medium body sizes (Ball and Morris 2015; Falk 2015; Speight et al. 2016). Ireland is one of the least wooded countries in Europe (e.g. Bullock and Hawe 2014). However, the remaining small woodland reserves scattered across a landscape have the potential to deliver high levels of multiple ecosystem services (Valdés et al. 2019). Since these natural habitats have the potential to support many wild pollinators, they can provide complementary pollination services (Bodin et al. 2006; Garibaldi et al. 2011, 2013; Krishnan et al. 2020). We hypothesise that the abundance of woodland specialists will be higher in landscapes with a higher proportion of woody areas and with higher connectivity between farm and woody habitats.

Methodology

Study areas and farm selection

Two study areas in Ireland were selected using existing sub-catchments maps (shapefiles created by the Irish Environmental Protection Agency 2018). The study areas were located in contrasting regions (north-west and south-east of Ireland; see Figure 1), with climatic variations. The selected sub-catchment in the north-west (Co. Sligo) has an area of 145.5km², whilst the sub-catchment located in the south-east (Co. Wexford) has an area of 197.12km². Co. Sligo has slightly lower mean temperatures and higher rainfall than Co. Wexford (mean annual temperature and precipitation in Co. Sligo: 9.6 ºC and 1260.1mm; and in Wexford 9.8 ºC and 840.2mm; Met Éireann 2020). Pollinator data were collected from May to August 2019. Monthly temperature and precipitation for this period are shown in Table A1 in the Appendix.

In each sub-catchment, fifteen farm parcels were selected for sampling pollinators. A farm parcel consisted of two-five adjoining fields and all surrounding linear features. The selection of farm parcels was based on the average habitat quality score from all existing habitats in each farm parcel (i.e. grasslands, hedgerows, drainage ditches, stonewalls or heathlands), which were first scored individually using tailored scorecards developed by Rotchés-Ribalta et al. (2021). The scorecards comprised several criteria, such as vegetation structure, cover and abundance of positive plant species and poaching levels (see Rotchés-Ribalta et al. 2021 for a detailed description

Fig 1.

Location of the study areas (1 – Co. Sligo sub-catchment; 2- Co. Wexford sub-catchment) in the Republic of Ireland.

of habitat quality scoring). The average habitat quality score of each parcel was determined from the scores of each individual farm feature. In each study area, five farm parcels of low-quality score, five farm parcel of intermediate quality score and five farm parcels of high quality score were surveyed.

Pollinator traps had to be at least 1km apart from one another (to reduce the risk of spatial autocorrelation), resulting in one farm parcel in the Co. Sligo study region being excluded. One of the surveyed parcels in Co. Wexford was also excluded since it was located outside of the sub-catchment boundaries and functional connectivity metrics could not be determined. In total, 28 farm parcels were considered. In Co. Sligo, all selected farms were dedicated to livestock production (grasslands and semi-natural pastures, such as heathlands). In Co. Wexford, besides livestock production, five of the selected farm parcels had fields of cereals and other crops (e.g. barley, radish).

Pollinator sampling

Insect samples were collected at approximately three-weekly intervals from May to August 2019 (five sampling rounds in total) using sets of coloured pan traps (Stanley et al. 2013; Power et al. 2016). Each set of pan traps set consisted of three fluorescent, pre-painted plastic bowls (375ml capacity, 135mm diameter), one of each colour (blue, white and yellow) (Moreira et al. 2016). In each parcel, three sets of pan traps were installed (nine bowls in total) along a single farm linear feature. Each set of pan traps was placed 15m apart from each other and at least 10m away from the end of the linear feature. The pan traps were partly filled with water with a drop of soap to break the surface tension. Traps were collected after 48 hours and all captured specimens per site were pooled and stored in 70% ethanol for later identification (Stanley et al. 2013). Pollinating insects (solitary bees, bumble bees, hoverflies) were subsequently identified to species level, using Ball and Morris (2015) and Stubbs and Falk (2002) for

Fig 2.

(A): 'Optimal habitat' map for Co. Sligo; (B): 'Optimal habitat' map for Co. Wexford. The outlined circles represent the 500m buffers around the centre of the linear feature where the traps were placed. (C): Example of a buffer of 500 m from the pan trap locations, with optimal habitat shown in black.

Syrphidae; Falk (2015) and Else and Edwards (2018) for bees. The abundance of each species per parcel was calculated as the sum of the individuals found in the traps over the five sampling rounds.

From the insects collected, a subset of species was selected, based on two criteria:

1. 1). Body size/mobility: only species with body sizes less than or equal to 10mm (as defined by Moquet et al. 2018) were selected. Species with body length above this threshold were considered highly mobile and therefore excluded from the present analysis. Body size is reported to be an important factor for hoverfly and bee dispersal (Greenleaf et al. 2007; see for example Aguirre-Gutiérrez et al. 2016 and Steffan-Dewenter and Tscharntke 1999), thus mobility was inferred taking this information into consideration. Body size length (average between females and males) was determined for each species based on Speight et al. (2016) for hoverflies and on Falk (2015) for bees.

2. 2). Woodland specialists: all the species fitting the previous criteria were further classified as 'habitat generalists' or 'habitat specialists' by consulting existing literature (Ball and Morris 2015 for hoverflies; Falk 2015 for bees). In the end, eight hoverfly species (Cheilosia albitarsis (Meigen), Dasysyrphus albostriatus (Fallén), Dasysyrphus venustus (Meigen), Ferdinandea cuprea (Scopoli), Melanostoma scalare (Fabricius), Meliscaeva auricollis (Meigen), Meliscaeva cinctella (Zetterstedt) and Platycheirus scutatus (Meigen)) plus two wild bee species (Andrena fucata Smith and Andrena lapponica (Zetterstedt)), reported to have preference for woodland habitats, were considered in this analysis.

Landscape metrics

Complete habitat maps were developed (remotely) for each sub-catchment on ArcGIS.10.5 (ESRI, 2016) using Google Earth imagery. All areal and linear habitats were digitised and two distinct vector shapefiles for each sub-catchment were generated. We followed Fossitt (2000) level 2 classification for habitat classification and used Smith et al. (2011) for determining areal features and linear features minimum mapping sizes. The maps generated for each sub-catchment are available with corresponding Fossitt codes (linear habitats Co. Wexford: doi:10.13140/RG.2.2.31314.79044; linear habitats Co. Sligo: doi:10.13140/RG.2.2.12020.99201; areal habitats Co. Wexford: doi:10.13140/RG.2.2.32153.65128; areal habitats Co. Sligo: doi:10.13140/RG.2.2.20409.60005) and the list of the all habitats and corresponding Fossitt (2000) code can be seen in Table A3 (Appendix). The percentage of each areal and linear habitats (at level 2 of Fossitt 2000) was calculated in each sub-catchment, and the density of linear features was also determined for both regions by dividing the total length of linear feature by the total area of the sub-catchment (results in Figure A1 (Appendix)).

A 500m buffer (in ArcGIS.10.5) was created around each trap (from the middle of the linear feature where the traps were placed) to extract relevant landscape composition and landscape configuration metrics. This choice of spatial extent allowed us to avoid spatial auto-correlation and is in line with other studies (e.g. Joshi et al. 2016).

Some of the landscape metrics were extracted from the maps with all habitat types (see Table A3), whilst others were obtained from 'optimal habitat' maps – these maps were generated for each sub-catchment by only retaining areas of woody vegetation using the select by attributes function in ArcGIS. Therefore, two separate maps were generated where: background = non-woody vegetation areas and foreground = woody vegetation areas. See Figure 2 (A) and (B) with the

Table 1.

List of habitats considered for generating the 'optimal habitat' maps (Figure 2), with respective code as in Fossitt (2000).

generated maps for both Co. Sligo and Co. Wexford and Figure 2 (C) with an illustration of a 500m buffer of 'optimal habitat'. The habitats assumed to be optimal for the selected hoverfly and bee species (and thus used to generate the 'optimal habitat' maps) included native woodland, mixed habitat containing native woodland, and scrub (Table 1). Hedgerows and treelines were not considered in the 'optimal habitat' maps, but were considered as a potential explanatory variable (i.e. total length of hedges in a 500m buffer) and thus tested as potential suitable habitat. Conifer plantations were also excluded as an optimal habitat in this study since in Ireland the management is very intensive. A study conducted by Gittings et al. (2006) showed that hoverflies richness in conifer plantations is more related to the existence of open areas, broad-leaved trees, shrubs and wet habitat features that surround the plantation than to the habitat created by the conifers per se. Consequently, conifer plantations were considered an intensively managed habitat.

All the landscape metrics were divided into landscape composition metrics and landscape configuration metrics. The compositional metrics calculated for each selected farm parcel were the following: 1) total area of optimal habitat in the 500m buffer; 2) area of intensive production in the 500m buffer; 3) area of semi-natural pastures (heathlands, peatlands and/or mosaics of heathland/semi-natural grasslands etc.) in the 500m buffer; 4) area of semi-natural grasslands in the 500m buffer; 5) total length of hedgerows in the 500m buffer; 6) area of the closest patch of optimal habitat (from the traps); and 7) greatest patch area in the 500m buffer. The configurational metrics generated were: 1) distance from the traps to optimal habitat; 2) connectivity of linear feature where the traps were placed to optimal habitat (i.e. if it was physically linked to an optimal habitat area); 3) production fields average size; 4) edge length in the 500m buffer; 5) edge density in the 500m buffer; 6) Splitting Index S (SPLI) for the 500m buffer; 7) Landscape Division Index (LDI) for the 500m buffer; 8) Patch Cohesion Index for the 500m buffer; and 9) functional connectivity metrics – comprised of seven indices (dPA, dF, dFWA, dPC, dpCcintra, dpCflux, dPCcon) and determined considering the total extent of the 'optimal habitat' maps for each sub-catchment. All variables are explained in detail in Table A2 (Appendix).

Beside these landscape structure variables we considered three additional variables: two variables related to the habitat quality (determined by Rotchés-Ribalta et al. 2021) of the parcel where the traps were placed – habitat quality of fields and habitat quality of linear features; and a categorical variable, Region, so we could account potential larger scale landscape differences between the two study regions.

Statistical analyses

Regression models were conducted to investigate the response of pollinator abundance data to the landscape metrics selected. Collinearity amongst the landscape metrics was assessed using a Spearman correlation analysis. When the correlation between two explanatory variables was greater than |r| ≥ 0.7 (Dormann et al. 2013), we decided to run the models, initially, with the variable that provided a more direct explanation (e.g. Area of the closest woodland patch instead of the connectivity index dPC). The list of variables introduced in the first iteration of the models, corresponding correlated variables and descriptive statistics are shown in Table 2.

Spatial autocorrelation of woodland specialist abundance was investigated via Moran's I for both sub-catchments combined and separately. Since the results of this analysis showed that the spatial distribution of the pollinator abundance is likely to be a

Table 2.

Descriptive statistics of the explanatory variables introduced in the models and correspondent correlated variables (|r| ≥ 0.70).

result of random spatial processes (Moran's I for both sub-catchments: Moran´s I = -0.086, p >0.05; for Co. Sligo farms' data: Moran's I = -0.072, p > 0.05; Co. Wexford farms' data: Moran's I = -0.016, p >0.05) it was not necessary to apply any corrections to the subsequent models.

To determine the relationship between our explanatory variables and the abundance of woodland pollinators, we first attempted to use Generalized Linear Models (GLMs) with a Poisson response distribution and a log-link function, appropriate for use with count data. A modified version of the stepAIC() function from the R package MASS (Venables and Ripley 2002) was used for model selection. This modified function uses the AICc (Akaike's corrected information criterion, applicable for small sample sizes) rather than AIC to select the predictor to add at each step (Read et al. 2018). The model selection was set for both 'backward' and 'forward' selection. An additional backward elimination approach to find the most parsimonious model was also applied: the least significant effect is removed when it does not meet the significance level of p-value <0.05 (Hong and Mitchell 2007). For the evaluation of the models' regression fit we considered the models' AIC. We compared different model iterations by replacing the significant variables with those with which they were correlated, to understand if the model would improve its explanatory power. When we obtained an apparently final model, where all the explanatory variables/parameters were significant (p-value < 0.05), we analysed the model's robustness by: a) testing the model for significance by performing a comparison (ANOVA) between the obtained model and a model where the response variables were fitted against 1; b) testing for overdispersion (by dividing the model deviance by the residual's degrees of freedom) (Thomas and the Guidebook Team 2017); and c) checking the adjusted-R² value. When overdispersion values above 2 were observed, we developed new models and repeated the process, but this time using a Negative Binomial response distribution (log-link function).

Given the different results obtained (higher abundance of hoverflies in Co. Sligo and of bees in Co. Wexford – see Table 3 in Results section), we developed separate models for each taxonomic group and followed the same model development process and the above-mentioned selection criteria for each.

Models' results validation

Because of the small number of samples considered in this study (28 sampling points) and low abundance of selected pollinators, we tested if the significant relationships obtained between explanatory variables and woodland specialist abundance were a 'chance effect', due to possible overfitting.

Thus, results of the models were validated by fitting the emergent significant explanatory variables (via GLMs) to:

1. a). The abundance of randomly selected species from the full dataset of pollinating insects collated. In particular, we generated a new response variable by randomly selecting a group of ten species from the total list of pollinators – eight hoverfly and two bee species (from all species sampled). We were careful to exclude the woodland specialist species selected before conducting the random selection of species;

2. b). Randomly generated variables with the same distribution (Negative binomial, Poisson) and mean ('null models') as the woodland specialist response variables. The random response variables were obtained using the function rpois(N = sample size; m = mean) or rnbinom(N = sample size; m = mean). These R functions generate random numerical lists with a Poisson and negative binomial distribution, respectively. Fifteen 'response variables' were created: five different 'response variables' simulating all abundance data combined (hoverflies and bees); five 'response variables' simulating hoverfly abundance data; and five 'response variables' simulating bee abundance data (following Farine (2017)).

This was a strategy utilised to see whether the relationships obtained for the woodland specialists were spurious or not, by testing if the same variables would emerge as significant for other species (non-woodland specialists) and for randomly generated 'response variables'. Consequently, the results of these null models are not discussed in detail. All statistical analyses were conducted in R version 3.5.2 (R Development Core Team 2013).

Results

Abundance of woodland specialists (hoverflies and bees)

A total of 86 woodland specialists from the 10 target species were captured (see Table 3), with 56 hoverflies and 30 bees. This value represented c 2% of the total abundance of pollinators captured (more than 3,500 bees and hoverflies captured in both sub-catchments). The majority (36) of woodland hoverflies captured were M. scalare, mostly in Sligo (24 individuals), whilst the most abundant bee species was A. fucata, particularly in Co. Wexford (16 individuals).

The best model obtained for the abundance of woodland specialists included the Distance to the closest optimal habitat, the Connectivity of linear feature to optimal habitat and Edge density (of woody areas in a 500m buffer) as significant variables (see Table 4 (A) for model results). The Connectivity of linear feature to optimal habitat had a positive relationship with the abundance of woodland specialists, whilst Edge density and Distance to the closest optimal habitat had a negative relationship with the abundance of these pollinators.

When comparing this model with the model fitted to the abundance of ten randomly selected species (i.e. the null models), none of these three variables had a significant correlation with their abundance. Results of the null models (random

Table 3.

List of selected pollinators (with correspondent abundance and number of occurrences) with body size ≤10mm (average between males and females) and reported to have preference for woodland habitats.

negative binomial 1, 2, 3, 4 and 5) reveal that for only one of the models (random negative binomial 2), Edge density emerges as a significant explanatory variable, but this time with a positive correlation with the randomly generated variable. However, this model has a higher AIC value and a lower-adjusted R². The fact that the variables never emerge as significant for any other of the tested null models, and

Table 4.

(A): Results of the model obtained for the abundance of woodland specialists (hoverflies and bees) with respective results for goodness of fit for the abundance of 10 randomly selected species and for the random negative binomial 1,2,3,4 and 5; (B): Results of the model obtained for the abundance of hoverflies with respective results for goodness of fit for the abundance of 8 randomly selected hoverflyspecies and for the random negative binomial 1,2,3,4 and 5; (C): Results of the model obtained for the abundance of bees with respective results for goodness of fit for the abundance of 2 randomly selected beespecies and for the random POISSON 1, 2,3,4 and 5. Significance levels: '***' 0.001 '**' 0.01 '*' 0.05.

for the abundance of ten randomly selected species, indicates that the model results for the abundance of woodland specialist are likely to be more than a 'chance effect'.

When we analysed the relation between the explanatory variables (Table 2) and the abundance of woodland hoverflies (GLM with negative binomial distribution), we verified that Region and the Connectivity of linear feature to optimal habitat emerged as significant, with less abundance of these pollinators found in Co. Wexford and a higher abundance found in the farm-linear features connected to optimal habitat (i.e woody areas). When comparing this model with the model fitted to the abundance of eight randomly selected hoverfly species, we can see that this time there was a higher abundance of these randomly selected species in Co. Wexford and that there was no relation between these species abundance and the Connectivity of linear feature to optimal habitat. Additionally, results of the null models (random negative binomial 1, 2, 3, 4 and 5) reveal that neither of the two significant variables ever emerged as significant for the randomly generated response variables. Thus, the model results for the abundance of hoverfly woodland specialists are likely to be more than a 'chance effect' (all results shown in Table 4 (B).

For bees, from all variables inputted in the model, only Region emerged as significant in the GLM (with a Poisson distribution). In contrast to what was observed for hoverflies, a higher abundance of the two woodland bee species selected was sampled in Co. Wexford. This variable did not emerge as related to the abundance of the two randomly selected bee species, or to any of the randomly generated Poisson variables (Poisson 1, 2, 3, 4, and 5), which indicates that this variable is likely to have a 'real' significant effect on the abundance of woodland bee species (results shown in Table 4 (C)).

Since Region was a significant variable in both hoverfly and bee models, we also examined whether there were significant differences between sub-catchments, via Kruskal-Wallis tests, for all the explanatory variables considered in the model initially (e.g. between the total area of optimal habitats within the 500m buffers defined around each trap; between the distance to the closest optimal habitats, etc.). We did not find significant differences between regions for any of the variables (Kruskal-Wallis p > 0.050), except for the area of semi-natural pastures, which was higher in the 500m buffers surrounding the Co. Sligo farm parcel. Results from this analysis are presented in the Table A4 (Appendix). Whilst they are not discussed in detail, they enable better interpretation of the model results for the abundance of the two groups of woodland pollinators.

Discussion

This study aimed to explore how landscape structure is related to the presence of low-mobility woodland pollinators in Irish farmland. Few explanatory variables explained the patterns observed; however, it was possible to verify that some of the selected landscape metrics had a significant relationship with the abundance of woodland pollinators.

The abundance of woodland pollinators collected was relatively low, with two species (one hoverfly: M. scalare; and one bee, A. fucata) dominating. This might be related to the fact that these two species are less exclusively bounded to woodland areas and are considered as frequently occurring (Ball and Morris 2015; Falk 2015). Yet, the abundance of these two species is also not very high, considering the number of traps placed in the farms, the amount and variability of farms sampled and the sampling period distribution. The low abundance can be seen as a result in itself and is not necessarily surprising, due to an overall low cover of native woodland in the Irish landscape (Bullock and Hawe 2014). In this study, we also verified a low percentage of area of woody habitats in the 500m buffers of our sampled farms (5.5% of the total buffer area on average) with most of the land-use dedicated to food and fibre production. In fact, most of the area in both study regions, at sub-catchment level, comprises improved grasslands (almost 44% in Co. Wexford and 54% in Co. Sligo) (see Figure A1 in the Appendix).

It could be possible that the low abundance of woody vegetation areas is compensated by the presence of hedgerows. Yet, the total length of hedgerows in the surrounding 500m from the traps did not correlate with the woodland specialist abundance. This might indicate that these linear farmland habitats can function as corridors and provide resources (Cranmer, McCollin and Ollerton 2011), but their ecological role does not seem to be equivalent to that of areas of woody vegetation. This could be the result of management actions that reduce hedgerow ecological value (Byrne and del Barco-Trillo 2019).

Landscape configurational metrics have a greater influence than selected compositional metrics

We found a positive relationship between the abundance of woodland specialists, particularly hoverflies, and connectivity of the linear feature (where the traps were placed) to a woodland patch, which is in line with other studies (e.g. Ricketts et al. (2006) for different pollinator communities; Van Geert et al. (2010) for pollen dispersal; Haenke et al. (2014) and Jovičić et al. (2017) for hoverflies; and Schirmel et al. (2016) for predatory carabids). Thus, besides retaining and/or creating linear features to increase farmland biodiversity (e.g. Sullivan et al. 2017), it may also be important to develop a network of connected linear elements to areas of existing or new semi-natural habitats. This can potentially result in an increase in pollination services in farmland (Cranmer, McCollin and Ollerton 2011; Delattre, Vernon and Burel 2013). However, connectivity of woodland patches to woody areas was not related to the abundance of selected bee species. In other studies, e.g. Krewenka et al. (2011), grass strips acted as sinks rather than corridors for bees, and connectivity was not considered important. This result might support the idea that species and/or taxonomic groups respond differently to landscape structure (e.g. Kennedy et al. 2013; Moquet et al. 2018). In fact, specific functional traits among pollinator groups might explain distributional variability (Aguirre-Gutiérrez et al. 2016). Yet, because of the low abundance values when we separated both groups, the observed patterns for the separate regression models can be considered less reliable.

An increase in distance from natural habitats has been reported to affect pollinator richness and abundance in agricultural areas in other studies, since fewer pollinator species are able to either forage or nest in agricultural areas (Jauker et al. 2009; Garibaldi et al. 2011; Bailey et al. 2014). Similarly, we observed that the abundance of woodland pollinators in our farms decreased with increasing distance to optimal habitat patches. Moquet et al. (2018) also showed that distance to forests areas was the main factor related to hoverfly composition in their study, but this variable did not emerge as significant when we considered hoverfly and bee abundance data separately, again, possibly due to the low abundance values.

The functional connectivity metrics determined did not emerge as significant. However, these functional connectivity metrics are only potential (based on theoretical flight distances and theoretical ideal habitats) and not actual (Calabrese and Fagan 2004). In fact, foraging distances might vary with the environmental conditions of the other landscape units (e.g. physical resistance of the different habitats to flight, in Ricketts, 2001). Other studies have reported that the quality of different land-use units may change pollinator flux among patches, and that they do not perceive landscapes as binary habitat-non-habitat systems (Abdel Moniem et al. 2013; Slancarová et al. 2014). This influences how functional connectivity metrics are related to pollinator movement (Slancarová et al. 2014). Thus, we cannot conclude that there is no effect of these variables per se: we can only state that we did not find any relationship based on our assumptions, in particular the flight distances.

A negative relationship between the combined abundance of woodland specialists and edge density (compositional metric) was observed (but not for hoverfly and bee abundance separately). This variable was correlated with total area of optimal habitats and edge length in the 500m buffers in our study, yet, when we replaced edge density by either the total area of optimal habitats or edge length, the model lost explanatory power, which suggests variable effects of these three factors on the abundance of woodland specialists. Positive relationships between edge habitat in woodland patches and the abundance of pollinators have been reported by Popov et al. (2017) and Roberts et al. (2017). Nonetheless, we believe that our results are not contradictory to this observation: we sampled fewer woodland pollinators in farms when there was a high edge density of optimal habitat in the surrounding 500m, possibly because these species dwell more in the edge than in the surrounding agricultural fields (where we sampled) due to higher ratios of feeding resources and nesting sites available. In fact, a great number of the species that were selected in this study are associated with edge habitats of woodlands and forest areas (see Table 3). Thus, this result might indicate that the abundance of woodland specialists is a function of high levels of edge habitats per total forest area.

Variables related to habitat quality for pollinators in the surrounding woodland patches and farmland habitats (flower resources, nesting sites, and vegetation structure) were not considered in this study, since our goal was to contribute to understanding the effects of landscape structure in the abundance of woodland specialists in farmland. However, this information could improve models' performance and could allow the disentangling of some of the results obtained, since habitat quality (for pollinators) of surrounding habitats can affect pollinator abundance (Krewenka et al. 2011; Kennedy et al. 2013).

Regional differences between hoverflies and bees

When hoverfly and bee abundance data were separated, most landscape metrics lost significance and the variable Region consistently emerged as significant. Whilst the abundance of woodland hoverflies was significantly lower in Co. Wexford, the abundance of woodland bees was lower in Co. Sligo. The existence of contradictory responses between these pollinator groups, and pollinator guilds in general, is reported in other studies (e.g. Jauker et al. 2009; Robert et al. 2017), thus it is not surprising. In the case of the bees, Region was the only significant predictor, but its real effect should be carefully considered since this model had the lowest adjusted-R² value.

We could expect some underlying regional differences between the landscape variables introduced in the models (Table 2). However, no significant differences between most of the configurational and compositional metrics were found between the two regions, except for the fact that the amount of semi-natural pastures in the 500m buffers was higher in Co. Sligo. In fact, the percentage of semi-natural pastures represents around 20% of the total area in Co. Sligo (Figure A1, Appendix) and a much lower percentage in Co. Wexford (circa 5.7%). Yet, if the area of semi-natural pastures explained the contradictory pattern observed, then it would most likely have emerged as a significant predictor, which was not the case. Thus, it seems that other regional differences are affecting the abundance of hoverflies and bees and were not captured in this study.

Hence, we put forward two possible hypotheses: 1) the regional effect observed can be related to differences in quality of woodland patches for both pollinator groups (and therefore pollinator abundance in the woodland areas), such as the degree of wetness or flower resources available (see, for example, Fuller et al. 2017). Willaert (2019) found that the attractiveness of plant species differed largely between wild pollinator groups and that flower presence at the local scale affected pollinator species richness and abundance more than correlated landscape structure variables. However, we do not have information on pollinator abundance in the woodland areas or woodland patch quality from our study to test this hypothesis. Alternatively, 2) the low abundance data in the two taxonomic groups reduced the statistical performance of the models. In fact, when we randomly selected eight hoverfly species from the total sampled species list, we verified that Co. Wexford had a positive relationship with their abundance, indicating that not all hoverflies are negatively affected by existing regional differences. Furthermore, when we randomly selected two other bee species, we found that there was no difference between regions.

This difference in pattern of occurrence between bees and hoverflies could be related to differences in weather conditions between study regions (see Table A1 in the Appendix). However, we would then expect a higher abundance of both groups in Co. Wexford due to higher average temperatures and lower precipitation levels (Totland 1994; Lawson and Rands 2019); yet, the species of hoverfly considered were more abundant in Co. Sligo, which would imply that these species prefer lower temperatures and higher precipitation levels.

Although the apparently contradictory effect of region on the species selected might indicate that some parameter choices valid in one environment may not hold when other landscapes and environments are considered (Newman et al. 2019), this is a tentative suggestion given the low abundance of insects in our analysis. Therefore, as a tentative conclusion, we consider that Region emerging as explanatory variable might be a combination of a low sample size when the groups are separated, with some larger sub-catchment ecological differences that we could not disentangle (such as the quality of woodland areas and/or intensity of agricultural activities), which would have to be affecting the selected species of both groups and not the taxonomic groups per se.

Implications for aes design

Even if most of the agricultural production in both study regions is not significantly dependent on pollination (because it comprises mainly grasslands and cereal production), future climate change might lead to alterations in farming systems and the need for adapting to new weather conditions. In fact, climate change is projected to improve the conditions for growing crops in northern Europe (European Environment Agency 2019). Thus, the occurrence of a diverse pollinator community in farmlands, including woodland pollinators, can provide insurance for future production (Yachi and Loreau 1999; Winfree et al. 2007, 2008) and increase productivity of crops (Dainese et al. 2019; Pfeiffer et al. 2019; Woodcock et al. 2019). Furthermore, multispecies swards (a key strategy for reducing chemical fertiliser input and related greenhouse gas emissions) (Peyraud et al. 2010) comprise some species (e.g. clover) with seed set rates depending on pollination (e.g. Jing et al. 2021).

A diverse community of pollinators can also have positive effects on the ecological health of ecosystems, since pollinator richness can be important for the population dynamics of wild plant species (e.g. Ollerton et al. 2011; Albrecht et al. 2012). Hence, the presence of the woody areas scattered across the agricultural landscape (particularly when connected through farmland linear features) has the potential to maintain some wild pollinator populations, acting as a reservoir and source (Taki, Kevan, and Ascher 2007; Gaytán et al. 2020). These pollinators then travel to farmland to acquire floral resources and, thus, support pollination services (e.g. Carvalheiro et al. 2011; Valdés et al. 2019; Huais et al. 2020).

Despite the limitations and complexities of our results, they indicate that the safeguarding of pollination services by woodland specialists in farmlands might depend on the connectivity of linear features to surrounding woodland patches, and on maintaining and creating small woodland patches around or within the farms, thus reducing the distance from the fields to woodland areas. Semi-natural woodlands and scrublands occurring in farmlands are considered undervalued habitats and ineligible for AES in Ireland (Rotchés-Ribalta et al. 2021). Therefore, AES that aim to increase the abundance and diversity of pollinators in farmland (as highlighted in the recent Farm to Fork Strategy (European Commission 2020b)), should be incentivising: a) a decrease in distance from production fields to semi-natural areas, by increasing (and incentivising farmers to retain) the quantity of woodlands and scrublands within farmlands; and b) the creation and connection of farmland linear features to these areas.