Backyard buzz: human population density modifies the value of vegetation cover for insect pollinators in a subtropical city

This study was conducted during November 2017 in Brisbane, Australia, in an area of approximately 20 × 30 km of suburban residential neighbourhoods varying in human population density and extent of green spaces (Fig. 1). We identified potential survey sites (private backyards) using social media and the staff newsletter at the University of Queensland. We received approximately 200 applications from Brisbane residents. From these, we selected 45 backyards (of 40 detached houses and 5 multi-family complexes) along gradients of vegetation cover and human population density per sealed surface, as detailed below.

We obtained data on human population density and land cover from Brisbane City Council: the 2011 population census mesh blocks and 2014 land cover map. The latter had been processed into a raster of 100 × 100 m grid cell size, containing sealed surfaces, and three vegetation classes based on LiDAR data (Caynes et al. 2016): tree canopy cover (≥ 2 m height), shrub cover (< 2 m), and grass cover. We combined all vegetation classes using R version 3.2.5 (R Core Team 2015) and the Raster package (Hijmans 2016). The polygon shape layer on human population density was rasterized in R into a 100 × 100 m grid raster, aligned with the vegetation raster. The census mesh block polygons vary in size and shape. We therefore divided the human population within a polygon between all 100 × 100 m raster cells that it intersected with, in proportion to the fraction of the polygon area each cell made up. We calculated human population density per sealed (grey) surface area in each grid-cell (Hahs and McDonnell 2006). This measure is more related to urban form and how intensively the built land is used than human population density per se.

To select focal backyards along urban gradients, we used these raster layers to perform moving window analyses to calculate mean vegetation cover and population density in each 1 × 1 km window (10 × 10 pixels) over an area including all candidate backyards. The output was visually inspected and data points (focal pixels) outside the 10^th and 90^th percentiles on each axis (population and vegetation) were removed to avoid selecting backyards in areas where the two gradients were highly correlated (Pasher et al. 2013). This allowed us to include both vegetation and human density gradients in the same statistical models. We divided the remaining data points into four categories: 1) low vegetation cover and low human density, 2) low vegetation and high human density, 3) high vegetation and low human density, and 4) high vegetation and high human density. The categories were exported as tiff-files and used to draw maps in ArcGIS (Esri). Based on these maps and the position of the candidate backyards, we selected 45 sites to cover a wide variation in vegetation cover and human population density (Fig. 2). Selected backyards were on average situated 1584 ± 1035 m (mean ± std) from the closest other backyard in the study.

To obtain the urban gradients for analyses of pollinator response variables, we used the extract function in the Raster package (Hijmans 2016) to extract vegetation cover, cover of sealed surface, and human population numbers within 100 m and 500 m radii of each selected backyard centre coordinate. Vegetation data was categorised as: tree canopy cover, and grass and shrub cover combined, because of the potentially contrasting effects these categories may have on pollinators (Hall et al. 2019).

Each backyard was surveyed for pollinators three times in early summer, 3rd to 25th November 2017, between 08:00 – 13.00 on days with favourable weather for flying insects (wind < 5 m/s, temperature between 22 and 34 ^o C). To avoid bias, the three surveys per backyard were rotated to be either early, mid, or late in the day, within the set sampling window (8:00 – 13:00). Each survey involved slowly walking along two 50 m long 1 m wide transects for a total of 10 min, while collecting pollinators with a standard insect aerial sweep net. In cases where we were unable to capture specimens, we instead noted them to the closest identifiable genus or morpho-group. Eleven small backyards (< 100 m²) only had space for a single 50 m transect, which was surveyed for 5 min. This difference in sampling effort was accounted for in analyses, as detailed below. To index the attractiveness to pollinators of individual transects, we estimated the number of open insect-pollinated flower units in each transect. A flower unit is an easy countable entity that requires a pollinator to fly between them (Szigeti et al. 2016). The flower unit counted varied by taxa, e.g., individual flowers for Rosaceae, inflorescences for Asteraceae, racemes for Fabaceae, umbels for Apiaceae and flowering stalks for small-flowered Lamiaceae.

We mapped backyard features to characterise backyards into design types. We overlaid a 2 × 2 m grid onto aerial photos of backyards obtained from NearMap (freely available to researchers at University of Queensland, 2017). During field visits, we noted the proportion cover of the following features in each grid cell: short cut lawn, tall grass, shrubs, tree-trunks, tree crown-cover, leaf-litter, open soil, vegetable and berry patches, flowerbeds, gravel or stones, and sealed surfaces (e.g., wooden deck, pool, concrete paths) and calculated the proportion of each per backyard. All features, except for tree crown-cover were mutually exclusive and thus summed to one. In addition, for each backyard we estimated its size (m²), total number of open flower units, and counted the number of eucalypt trees as a proxy for the number of native trees. We counted the number of flowering trees within a 50 m radius of backyards as a measure of locally available large flower resources. The most common trees in flower were several species of Myrtaceae, mainly Eucalyptus spp., and the exotic Jacaranda mimosifolia.

All statistical analyses were performed in R version 3.6.1. To evaluate potential effects of contrasting gardening styles on pollinators we used hierarchical cluster analysis on the mapped garden features, to obtain groups of backyard types. We included: lawn cover, tree crown-cover, shrub cover, backyard size, and flower density. Leaf-litter, tall grass, open soil, vegetable and berry patches, and flowerbeds were combined into “other features”, as each of these were missing or constituted only a small fraction of the majority of backyards. We first calculated a dissimilarity matrix using function daisy() in package Cluster (Maechler et al. 2020), then used hierarchical clustering with function hclust() to obtain groups. We selected three “backyard design type” groups based on visual inspection of the dendrogram and barplot, and interpretability of the resulting groups. The latter was checked by modelling how groups related to each of the backyard features included in the clustering using linear models, assigning each garden feature as the response variable and backyard design type as the fixed effect. We visually checked barplots of model residuals for equal variance among groups. The resulting three (3) groups were later included in pollinator models.

The backyard design types differed significantly in flower density (F_2,42 = 6.35; P = 0.0039), tree canopy cover (F_2,42 = 6.378; P = 0.0038), proportion lawn (F_2,42 = 50.46; P < 0.0001), shrubs (F_2,42 = 13.75; P < 0.0001), and “other features” (F_2,42 = 14.15; P < 0.0001), with a non-significant tendency for backyard size to differ between types (F_2,42 = 2.86; P = 0.069). Type 1 was characterised by lower cover of flowers, lawn and shrubs, and higher cover of “other features” (e.g., open soil, flowerbeds, vegetable patches, leaf litter and tall grass) than types 2 and 3. Type 2 was characterised by mixed vegetation with high cover of both flowers, lawn, trees, and shrubs. Type 3 was characterised by very high cover of lawn, but low cover of flowers, trees, shrubs, and “other features”, with a non-significant tendency to be of larger size than the other types.

We tested whether local backyard features differed along the urban gradients. First, we tested if backyard design types were related to the three gradients at 100 m and 500 m scales, using a non-parametric Kruskal–Wallis test, with design type as the response and gradient as the explanatory variable. Second, we used GLMs to check if the proxies for local native vegetation (number of eucalypt trees in backyards) and large flower resources (number of flowering trees within 50 m), were related to the urban gradients. Because models were over-dispersed, we used a negative binomial distribution with glm.nb from package MASS (Venables and Ripley 2002). We assigned the number of eucalypts, or flowering trees, as the response variable, and one of the urban gradients (low vegetation, tree cover and human population density at 100 or 500 m) as an explanatory variable in each model.

Collected insect samples were determined to species or morpho-species level by experts at the University of Queensland, the Australian Museum (Sydney) and Queensland Museum (Brisbane). Pollinator observations for the two transects and three survey rounds were aggregated into one observation per backyard. We calculated the mean flower abundance along the surveyed transects per backyard over the three surveys. We calculated the abundance of all observed and collected eusocial bee species (A. mellifera and Tetragonula spp.), non-eusocial bee species, and hoverflies. There were two reasons for analysing eusocial and non-eusocial species separately. First, because the former occurs both as feral wild colonies and managed hived colonies in the study region, their abundance is partly directly governed by humans. Second, because eusocial bees can communicate within the colony to recruit foragers to distant but rewarding foraging patches (Beekman and Ratnieks 2000), they may respond differently to urbanisation (Wenzel et al. 2019).

Estimates of bee species richness were based on all collected specimens and those easily identified to genus in the field without collection, i.e., A. mellifera, Tetragonula spp., Amegilla spp., Thyreus nitidulus, and Xylocopa (Koptortosoma) sp. The 26 specimens observed but not identified were left out of this estimate. Because sampling effort differed between backyards, with either one (50 m) or two (100 m) transects, we predicted the effective number of bee species (Hill number = 0) using the iNEXT() function (Hsieh et al. 2016). We combined eusocial and non-eusocial species to improve sample size, and thus robustness of the estimate. We did not analyse species richness for hoverflies, because only 82 specimens were caught and identified to species, mostly of a single species (Simosyrphus grandicornis, 57 of 82 specimens) (Online resource: Tables 3 and 4). Bee species were categorised in terms of their flower relationship, nesting habitat, body, and tongue length, according to Michener (2007), the PaDIL database, and measurements in the lab (Online resource: Table 4).

We modelled the effects of local and landscape predictors on pollinator response variables using GLMs. Full models were pollinator response variable = garden design type + (log)transect flower abundance + proportion tree cover + proportion grass and shrubs + human population density + (tree cover × human density) + (grass and shrubs × human density). In models of pollinator abundance, we included (log) transect survey area (50 or 100 m²) as an offset to account for unequal sampling efforts. The GLMs of pollinator abundance were over-dispersed, and we therefore used a negative binomial distribution with glm.nb from package MASS (Venables and Ripley 2002). Bee species richness (logged) was modelled assuming a Gaussian distribution as iNEXT predicted species richness are not integers. We ran one model for each combination of pollinator response variables (bee species richness, eusocial bee abundance, non-eusocial bee abundance, hoverfly abundance) and spatial scale (100 m and 500 m), i.e., eight models in total. Non-significant interactions were removed (backward selection), and models re-run to obtain final results for component factors.

All fixed factors were scaled prior to statistical analyses to facilitate interpretation of effect estimates, using function scale to centre values and divide them by their standard deviations. We calculated Variance Inflation Factors (VIF) to assess potential collinearity between fixed factors. We assessed over-dispersion and outliers using the package Dharma, function testResiduals() (Hartig 2020) and residual plots. We checked for spatial autocorrelation among residuals using Moran’s I. All VIFs of included variables were < 1.7, i.e., low enough to allow inclusion in the same model (Zuur et al. 2009). There was no autocorrelation in any model (Moran’s I, all P > 0.14).

There were no significant associations between the three backyard design types and the urban gradients at either spatial scale (all P > 0.09). The number of eucalypt trees inside backyards was not related to urban gradients at the 100 m scale (all P > 0.10). However, at the 500 m scale, there was a positive association between the number of eucalypts and tree canopy cover (slope estimate = 0.6522, Z = 2.367, P = 0.0180) and a non-significant tendency for a negative association to human density (est = -0.6475, Z = -1.873, P = 0.0611). The number of flowering trees within 50 m of backyards was positively related to tree canopy cover (est = 0.18806, Z = 0.09023, P = 0.0371), and negatively related to human population density (est = -0.21041, Z = -1.961, P = 0.0499) at the 100 m scale. There were no significant associations between flowering trees and gradients at the 500 m scale (all P > 0.15).

Species richness of bees was positively related to local transect flower abundance in models at both spatial scales (Table 2; Fig. 3). Neither backyard design types, nor any of the urban gradients at either spatial scale, showed significant associations with bee species richness.

Abundance of non-eusocial bees was significantly related to an interaction between human population density and tree canopy cover at the 100 m scale (Table 2), such that the association was positive at low levels of human population density but turned negative at high human densities (Fig. 4a). The proportion of grass and shrub cover at the 100 m scale was not significantly associated with bee abundance (Table 2). Non-eusocial bee abundance was positively associated with flower abundance in surveyed transects (Table 2; Fig. 4b). The three backyard design types differed significantly (Table 2), such that mixed vegetation (Tukey post hoc: Z = 2.781, P = 0.0149; mean = 4.38, 95% CI: 3.64–5.27) and lawn dominated backyards (Z = 3.118, P = 0.0051; mean = 4.78, CI: 3.90–5.86) had higher bee abundance than yards with an overall low vegetation cover and dominated by “other features” (mean = 2.91, CI: 2.34–3.62).

At the 500 m scale, the abundance of non-eusocial bee species was negatively related to cover of grass and shrubs (Fig. 4c), while a significant interaction showed that the effect of tree cover again was moderated by human population density (Table 2). As for the 100 m scale, bee abundnace was positively associated to tree cover at low population levels, while the association turned negative at high levels of human density. Again, flower abundance in transects was postively related to bee abundance, and the effect of backyard design groups was significant (Table 2; Fig. 5), with more bees observed in mixed vegetation (Tukey post hoc: Z = 2.625, P = 0.0235; mean = 4.60, 95% CI: 3.71–5.71) and lawn dominated (Z = 2.837, P = 0.0127; mean = 4.62, CI: 3.73–5.72) backyards than the “other features” type (mean = 2.82, CI: 2.27–3.50).

Abundance of eusocial bees was significantly related to an interaction between human population density and grass and shrub cover at the 100 m scale, (Table 2). Eusocial bee abundance was negatively related to grass and shrub cover at low population levels, but the association turned positive at high population levels, (Fig. 6a). The abundance of eusocial bee species was negatively associated with tree cover within 100 m (Fig. 6b). Bee abundance was strongly positively associated with flower abundance in the surveyed transect, while the backyard design types did not differ in eusocial bee abundance (Table 2).

At the 500 m scale, the abundance of eusocial bees was positively related to human population density, and negatively related to proportion grass and shrub cover (Table 2; Fig. 6c, d). Again, flower abundance in transects had a strong positive effect on bee abundance, while backyard types did not differ (Table 2).

The abundance of hoverflies was positively related to local cover of both grass and shrubs, and tree canopy, and negatively related to human population density at both spatial scales (Table 2; Fig. 7). Hoverfly abundance was positively associated with flower abundance in the surveyed transect, while backyard types did not differ significantly (Table 2).

The positive association found between bee species richness and availability of flowers in surveyed transects is in line with previous research, e.g., reviewed by Winfree et al. (2011) and Wenzel et al. (2019). Thus, habitat improvements in individual backyards may benefit urban bees, at least the generalist species that dominated our sample. Contrary to expectations, we found no effects of urban gradients on bee species richness, which has been shown by a study in Perth (Western Australia), where species richness in backyards and remnant bushland decreased with proportion of built space and distance to nearest bushland (Prendergast et al. 2022b). This may be a result of our relatively low sample size, and consequently that only a small proportion of the species pool was sampled (see below). There can be a high species turnover of bees between different type of urban green spaces, such as backyards, parks, and semi-natural sites (Martins et al. 2017; Banaszak-Cibicka and Żmihorski 2020; Prendergast et al. 2022b). As we only sampled backyard habitats, we may subsequently have missed the part of the bee community dependent on natural or semi-natural habitat. Alternatively, our proxies for habitat quality (the urban gradients) did not capture true habitat quality for the larger bee community. However, the fact that we found significant relationships between urban gradients and pollinator abundance suggests that we used relevant proxies for bee habitat quality.

We found a positive association between non-eusocial bee abundance and tree cover at low levels of human population density, but a negative one at high human densities. This effect was consistent across spatial scales. Our sample of non-eusocial bees was dominated by generalist species, mainly Amegilla spp., but did include a few specialists, as well as a mix of cavity and ground nesting species. Foraging and reproduction of generalist solitary bee species have been shown to respond to local availability of single flowering species (Persson et al. 2018). A recent study found that the number of provisioned brood cells of non-eusocial bees in urban parks increased with the proportion of native plants locally (Sexton et al. 2021), and it has been suggested that flowering phenology of managed and non-native plants may not match the activity period of native pollinators (Harrison and Winfree 2015). Assuming that more ornamental and intensively managed vegetation is found in more densely populated areas (Kowarik 2008; Aronson et al. 2017), it is therefore possible that the present result is caused by a reduction in quality of trees as foraging resources, and possibly also as nesting habitat, for non-eusocial bees along the human density gradient. Although not directly tested here, this is supported by the negative relation found between human density at the block scale and the number of (native and non-native) flowering trees in and near backyards, as well as the tendency for lower numbers of (native) eucalypt trees in backyards in more densely populated neighbourhoods. The negative associations between non-eusocial bee abundance and grass and shrub cover at the neighbourhood (500 m) scale further suggests that non-eusocial bee species in the Brisbane region of Australia do not prefer open urban habitats. In addition, the presence of pesticides, both by direct application (Larson et al. 2013) and through purchased ornamentals (Lentola et al. 2017), may reduce habitat quality in backyards for bees, and non-eusocial species may be particularly negatively affected (Rundlöf et al. 2015).

In contrast to non-eusocial species, eusocial bees (A. mellifera and Tetragonula spp.) were negatively associated with tree cover at the local block scale. In addition, we found that grass and shrub cover had a negative association with bee abundance at low human population density, but a positive one at high population density. At the larger neighbourhood scale, there was a positive association of bee abundance with human density, and a negative association with grass and shrub cover. These results indicate that, at local spatial scales, eusocial generalist species can capitalise on highly human modified green spaces, which include exotic ornamental shrubs and herbaceous flowers (Threlfall et al. 2015; Kaluza et al. 2016). In fact, urbanisation may even favour exotic bee species (Makinson et al. 2017; Wilson and Jamieson 2019), in our case the European honeybee. The affinity to humans among eusocial bees could partly be related to domesticated hives being kept in backyards. Urban beekeeping, using both honeybees and stingless bees, is indeed an increasingly common hobby in Australia (Heard 2016).

However, the negative effect of grass and shrub cover at larger spatial scales suggests that residential neighbourhoods dominated by such vegetation do not provide enough resources for eusocial bees, possibly because of a lack of habitat variation and complementation (Dunning et al. 1992) over the extended activity period of these perennial species. In addition to flower resources, stingless bees require resins collected from wounded trees (Kaluza et al. 2016), which is more likely found in mature trees (Roulston and Goodell 2010). The larger scale used here (500 m) corresponds well to stingless bee foraging ranges (Kaluza et al. 2016; Smith et al. 2017), making lack of habitat complementarity a possible explanation.

As expected, hoverfly abundance was positively associated with vegetation cover (both of trees, grass, and shrubs) and negatively associated to human population density, at both examined spatial scales. This suggests that increased vegetation cover partly can offset negative effects of dense urban areas on hoverflies. Only the adult life stage of hoverflies forage from flowers, while larvae are, for example, predators on aphids, decomposers, or live in aquatic environments. Most gardens likely lack a variety of these features due to management, including pesticide use, design, and plant choice. The dominant species, S. grandicornis, is a habitat generalist whose larvae feed on aphids and soft caterpillars and is one of Australia’s most common hoverflies (Soleyman-Nezhadiyan and Laughlin 1998). The abundance of tree dwelling invertebrates, including aphids and caterpillars, can be reduced in urban non-native (compared to native) vegetation (Jensen et al. 2022), possibly due to lack of co-evolution (Ehrlich and Raven 1964), with potential negative bottom-up effects on predatory insects. Combined, this could explain the overall low hoverfly sample size. As hoverflies can benefit from remnant natural or semi-natural habitats in both urban (Dylewski et al. 2019) and rural (Verboven et al. 2014) areas, the general positive effect of vegetation cover may thus result from higher availability of such resources at landscape scales, and subsequent spill-over of individuals into backyards. Interestingly, the results indicate that not even generalist aphidofagous hoverflies are common in densely populated blocks and neighbourhoods. Our results thus corroborate studies showing that hoverflies are more sensitive to urbanisation than bees (e.g. Bates et al. 2011; Verboven et al. 2014; Persson et al. 2020). As hoverflies are important pollinators (Rader et al. 2016) and contribute to other ecosystem functions (pest insect control, nutrient cycling), loss of hoverflies may lead to shifts in several ecosystem services in urbanised landscapes.

The sampling throughout November covered part of the peak season of insect pollinators in the region. Despite three survey rounds, we obtained a small sample compared to equal efforts in temperate urban study systems (pers. obs.). A likely reason is that bees mainly foraged from flowering trees and were therefore less attracted to the comparably low flower density (of mainly exotic species) in surveyed ground transects, resulting in low bee densities. Similar results were found in Sydney, Australia, where ground level visual surveys were unsuccessful in natural habitats dominated by flowering trees (Makinson et al. 2017). Our results therefore mainly give an indication on the availability of insect pollinators to herbaceous backyard plants and crops, and we cannot rule out that there are other patterns in bee diversity across Brisbane that we did not capture. Obtaining a larger sample with more species represented, for example through traps placed in flowering trees, may have allowed us to detect effects of urban gradients on species richness of both bees and hoverflies, and to look closer into which pollinator traits that may drive such relationships. However, we believe our main results (positive effects of tree cover for native non-eusocial bees, and high affinity to human habitats for eusocial generalist bees) would likely persist even if a larger part of the species pool were to be included in analyses. Including more pollinator taxa, such as Lepidopterans which are known to be sensitive to urbanisation (e.g., Chong et al. 2014; Theodorou et al. 2020), may also have rendered detectable effects on species richness.

The abundance of all pollinator taxa was positively associated with local flower abundances in surveyed transects. However, only non-eusocial bees differed in abundance between backyard design types, with higher abundances in the two types with more vegetation cover (both mixed vegetation and dominated by lawn). Non-eusocial bees, containing many small bodied species are more restricted in foraging range than the eusocial species (especially A. mellifera) (Greenleaf et al. 2007), and require nesting and foraging habitats at close distance. In contrast, eusocial bees are more efficient in using large but ephemeral flower resources because of recruitment of foragers into the same patch (Beekman and Ratnieks 2000), compared to foraging by a single female of non-eusocial species. Thus, backyard designs with high vegetation cover, and flowering trees and shrubs, can particularly benefit non-eusocial bees. If, on the other hand, the objective of a backyard owner is to attract large numbers of bees to pollinate garden crops, our results indicate that addition of flowering herbaceous ornamentals and reduced tree cover can increase the abundance of honeybees and stingless bees locally. Interestingly, the backyard design types were not related to any of the urban gradients, potentially because individual life-style identity is more important in shaping gardening preferences than surrounding neighbourhood features (Grove et al. 2006). This may open up for information campaigns, or so called “nudging”, to promote pollinator friendly design and management.

At the block and neighbourhood scales, an increased tree cover, including flowering species, will improve urban habitats both for non-eusocial bees and for hoverflies, and provide landscape scale habitat complementation for generalist eusocial bees. Because a large part of the Australian bee community will benefit from native flowering plants, especially trees within the Myrtaceae (Michener 1965; Threlfall et al. 2015), retaining or increasing such vegetation may be particularly important to promote bee diversity. Retaining remnant natural vegetation and green spaces with less intensive management (e.g., leaving dead plant material) may further benefit hoverflies by providing larval habitats, and subsequently facilitate spill-over into backyards. Urbanisation in Australia has occurred recently, compared to e.g., Europe, and remnant natural areas in and surrounding cities can harbour a high diversity of specialised and rare species (Hahs et al. 2009a, b; Ives et al. 2016). This makes urban conservation efforts particularly valuable. Urban areas are growing increasingly dense, with smaller backyards as a consequence (Sushinsky et al. 2017). It is therefore pivotal to both provide ample tree cover from public green spaces (Threlfall et al. 2017) and to promote biodiversity friendly gardening (Wenzel et al. 2019), in order to support urban biodiversity and important ecosystem service providers, such as bees and hoverflies.