Floral colour structure in two Australian herbaceous communities: it depends on who is looking

Abstract

Background and Aims

Pollinator-mediated interactions between plant species may affect the composition of angiosperm communities. Floral colour signals should play a role in these interactions, but the role will arise from the visual perceptions and behavioural responses of multiple pollinators. Recent advances in the visual sciences can be used to inform our understanding of these perceptions and responses. We outline the application of appropriate visual principles to the analysis of the annual cycle of floral colour structure in two Australian herbaceous communities.

Methods

We used spectrographic measurements of petal reflectance to determine the location of flowers in a model of hymenopteran colour vision. These representations of colour perception were then translated to a behaviourally relevant metric of colour differences using empirically calibrated colour discrimination functions for four hymenopteran species. We then analysed the pattern of colour similarity in terms of this metric in samples of co-flowering plants over the course of a year. We used the same method to analyse the annual pattern of phylogenetic relatedness of co-flowering plants in order to compare colour structure and phylogenetic structure.

Key Results

Co-flowering communities at any given date seldom had colour assemblages significantly different from random. Non-random structure, both dispersion and clustering, occurred occasionally, but depended on which bee observer is considered. The degree of colour similarity was unrelated to phylogenetic similarity within a co-flowering community.

Conclusions

Perceived floral colour structure varied with the sensory capabilities of the observer. The lack of colour structure at most sample dates, particularly the rarity of strong dispersion, suggests that plants do not use chromatic signals primarily to enable bees to discriminate between co-flowering species. It is more likely that colours make plants detectable in a complex landscape.

INTRODUCTION

It has been recognized for decades that pollinators can affect the structure of flowering plant communities (Levin and Anderson, 1970; Sargent and Ackerly, 2008; de Jager et al., 2011; Pauw, 2013). The prerequisites for pollinator-mediated interactions between plant species – shared pollinators and pollination limitation of reproductive success – are common (Campbell, 1985; Burd, 1994, 1995; Vázquez and Aizen, 2004; Knight et al., 2005; Bosch et al., 2009; Carvalheiro et al., 2014) and many communities show evidence of structure in pollination-related traits. For example, competition for pollinator services among Stylidium species appears to limit the co-occurrence of floral morphologies that place pollen on the same parts of insect visitors (Armbruster et al., 1994), and seed-set in a native Lythrum species is reduced when an invasive congener is present (Brown et al., 2002). Flowering phenology in some communities can be attributed to competitive displacement among species (Waser, 1978; Ashton et al., 1988; Aizen and Vázquez, 2006). However, not all community structure is derived from competition. Co-occurrence of species can improve pollinator visitation rates and reproductive success (Waser and Real, 1979; Schemske, 1981; Laverty, 1992; Moeller, 2004; Ghazoul, 2006; Lázaro et al., 2009; Liao et al., 2011; Wolowski et al., 2017; Kemp et al., 2019), even for sympatric congeners (Thomson, 1978; Moeller, 2004). Laverty (1992), for example, found that mayapple, Podophyllum peltatum, received more frequent visits from bumble bees and had greater fruit-set in the presence of a lousewort, Pedicularis canadensis. Similar ‘magnet plants’ that enhance the pollination of co-flowering species might be common in plant communities (Schemske, 1981; Laverty, 1992; Pellegrino et al., 2008; Johnson et al., 2003; Peter and Johnson, 2008).

The role played by floral colour in community structure is poorly known, although it is one of the most important traits that mediate plant interactions with their pollinators. Among bees, for example, chromatic cues are known to have strong effects on floral detection (Giurfa et al., 1996; Spaethe et al., 2001; Bukovac et al., 2017a; Dyer et al., 2016a), discrimination (Giurfa, 2004; Dyer et al., 2008; Avarguès-Weber et al., 2010; Spaethe et al., 2014) and preference (Giurfa et al., 1995; Morawetz et al., 2013; Dyer et al., 2016b). Some evidence suggests that floral colour plays a role in plant species coexistence. For example, Texan species of Phlox undergo competitive colour displacement in sympatry: P. drummondii evolves flowers with greater red saturation than P. cuspidata due to selection on an allele in the anthocyanin synthesis pathway (Hopkins and Rausher, 2012).

However, a potential weakness of much past work on floral colour is that it was based on human-observed colour categories or measurement of physical spectra without models of the perceptual mechanisms of pollinators (McEwen and Vamosi, 2010; Eaton et al., 2012). Recent developments in the modelling of animal colour vision offer new ways to account for the perceptual abilities of pollinators (Chittka and Menzel, 1992; Arnold et al., 2009a, b; Dyer et al., 2012; Shrestha et al., 2013; Garcia et al., 2017, van der Kooi et al., 2018), and indeed show that there are often significant differences in visual capabilities even between comparatively closely related hymenopteran bee species (Dyer et al., 2008; Morawetz and Spaethe, 2012; Bukovac et al., 2013; Garcia et al., 2017). Such differences fit with visual ecology principles that suggest that animals often have visual systems suited to collect nutrition most efficiently given physiological constraints such as body size and eye optics (Lythgoe, 1979; Land and Chittka, 2013). These new colorimetric approaches are increasingly being used to assess the composition of flowering plant communities, providing new insights into how the complexity of plant pollination operates.

Certain chromatic signals may be detected more easily than others by pollinator visual systems (Bukovac et al., 2017b), a constraint that could shape pollinator preferences and favour a limited range of floral colours (van der Kooi, 2018). Limited colour diversity could also be favoured if the collective advertising signals from multiple plant species attract increased pollinator activity to a patch, to the benefit of all species present (Brown and Kodric-Brown, 1979; Schemske, 1981; Laverty, 1992; Johnson et al., 2003; Liao et al., 2011; Kantsa et al., 2017, 2018). In a recent study of aster communities in South Africa, notable for the use of pollinator vision models and measurement of pollination networks, Kemp et al. (2019) found significant clustering of floral colour patterns, consistent with facilitation among species. In contrast, floral colours should be far more disparate in communities structured by competition, in which easily distinguishable floral colours would be favoured (Chittka, 1997). Pollinator-perceived floral colours were reported to be overdispersed among 244 mostly herbaceous species in mixed-usage habitat in central Japan (Makino and Yokoyama, 2015). Alternatively, coexistence may not depend strongly on pollination, so that communities would appear to be randomly assembled with respect to floral colour. Finally, we must consider a temporal dimension to floral colour structure: clustering, dispersive and neutral processes could play changing roles in the transient co-flowering assemblages that arise throughout an annual cycle of blooming in a community.

In the present work, we document the temporal structure of floral colour structure over the annual course of flowering in two species-rich herbaceous communities of temperate Australia. We compare the array of floral colours present at any given time with equivalent random averages sampled from the colours occurring at the site throughout the year. Colours that are more disparate than, more clustered than or consistent with the random expectations would be consistent with the influence of, respectively, competition, facilitation or neutral processes (weak pollinator-mediated interactions with little effect on species coexistence). We introduce an important innovation by quantifying colour structure according to the perceptual capabilities of four different hymenopteran visual systems in order to gauge the extent to which perceived colour structure might vary depending on the pollinators in question.

MATERIAL AND METHODS

Sites and data collection

Our data were collected at two natural communities in Victoria, Australia: Boomers Reserve (37°37′39″S, 145°15′21″E), a 90-ha site ~35 km north-east of Melbourne, and Baluk Willam Reserve (37°55′32″S, 145°20′45″E), a 68-ha site 40 km south-east of Melbourne. Both sites are Eucalyptus-dominated woodland with well-developed shrub and herb layers, managed to restrict non-native, invasive species (Fig. 1A–C). We set up 0.01-ha circular permanent quadrats at the sites using stratified-random placement to ensure sampling over the entire spatial extent of the reserves (43 quadrats at Boomers and 46 at Baluk Willam). The quadrat size is large relative to the size of the herbaceous plants sampled. All quadrats were sampled at 2-week intervals between March 2010 and August 2011. We thus obtained data on the temporal communities of co-flowering plants over slightly more than a complete annual cycle of flowering at the two sites. For each quadrat on each sampling date we counted the number of individuals and number of open flowers of every herbaceous species in flower, identified with the aid of Walsh and Entwisle (1994, 1996, 1999) and the expertise of the first author on plants of Victoria. Two orchid genera at our sites, Acianthus (two spp.) and Pterostylis (eight spp.), are dipteran-pollinated (Dafni and Bernhardt, 1990; Kuiter, 2016) and were not included in the analysis but are the subject of a parallel study (Shrestha et al., 2019). All the remaining species are pollinated by hymenopterans (Armstrong, 1979; Shrestha et al., 2013, and references therein; Kuiter, 2016). A list of species at each site is given in Supplementary Data Table S1 and five representative species are shown in Fig. 1D–H.

Fig. 1.

Examples of habitat and representative herbaceous species at the study sites. (A–C) Ground-layer vegetation with herbaceous species in flower, (D) Gompholobium huegelii, (E) Goodenia blackiana, (F) Burchardia umbellata, (G) Wahlenbergia gloriosa and (H) Glossodia major.

Floral spectra, colour distance and perceptual difference

The methods for representing colour as pollinators see it are complex. We present an outline of our procedures in some detail for the benefit of plant biologists who wish to understand how the findings of vision science can be incorporated into pollination research. However, readers should recognize that a large body of specialist literature should be consulted in order to apply current understanding of animal vision to ecological problems (Kemp et al., 2015).

Our goal was to represent the difference between two floral colours as a continuous metric that could be used in quantitative analysis of community colour structure. There are three fundamental steps that yield the metric we sought. First, we recorded the reflectance spectra of flowers of individual species. Second, these spectra were mapped to loci in a geometric representation of hymenopteran vision, a representation known as a colour space. Finally, Euclidian distances between floral colours in the colour space models were translated to metrics of effective perceptual discrimination by hymenopteran pollinators. This final step accounted not only for a range of behavioural reactions among pollinator species to a given colour difference, but also for different discrimination by a single pollinator in different regions of colour space.

Measurement of floral reflectance is the most straightforward of the three steps. We followed procedures that we had used in previous work (Dyer et al., 2012; Shrestha et al., 2013). In brief, reflectance spectra for wavelengths between 300 and 700 nm were recorded with an Ocean Optics spectrophotometer (Dunedin, FL, USA) using a PX-2 pulsed xenon light source and a pellet of dry BaSO₄ as a UV-reflecting white standard. Spectra from three flowers were averaged to obtain a single spectrum for each species. For species with multiple floral colours (including colours with a UV component), the predominant colour occupying the greatest petal area was measured. Examples of floral spectra for the five species in Fig. 1D–H are shown in Fig. 2A.

Fig. 2.

(a) Examples of raw floral reflectance spectra from the species shown in Fig. 1(D–H). Species abbreviations: Bu = Burchardia umbellata, Gb = Goodenia blackiana, Gh = Gompholobium huegelii, Gm = Glossodia major, and Wg = Wahlenbergia gloriosa. (B) Colour loci in bee hexagon colour space corresponding to the spectra in A. Distances between Gb and Gh, between Gm and Wg, and between Gm and Bu are labelled, respectively, by Roman numerals I, II and III. (C) Behavioural discrimination functions for Trigona fuscipenis, from Garcia et al. (2017). The curves show the probability that trained bees correctly choose a rewarding colour as a function of the Euclidian distance between the colours in a hexagonal colour model. Probabilities range from 0.5 (random choice) to 1 (consistent choice of a rewarding colour), and were used as the functional measure of perceptual distance in our analysis of community colour structure (see Methods). Discrimination functions for choices between colours in portions of the colour space corresponding to human-yellow or human-blue are shown as yellow and blue curves, respectively.

Next, the reflectance spectrum for each species was translated to a position in a hexagonal colour-space model of hymenopteran chromatic perception (e.g. Fig. 2B). The model calculates how a particular floral reflectance spectrum would excite the three photoreceptor classes in the hymenopteran eye (peak sensitivity in the UV, blue and green light wavelengths) and how these photoreceptor signals are mixed by neural processing (Chittka, 1992; Dyer et al., 2011). Three of the hexagon vertices represent maximal excitation of the three photoreceptor classes, while interior points of the hexagon represent colour mixtures from the subsequent neural processing. Colour similarity is represented by proximity in the hexagonal plane. For example, Burchardia umbellata, which has flowers that appear white to human vision (Fig. 1F), and Wahlenbergia gloriosa with human-blue flowers (Fig. 1G) are located near each other in the hexagonal model (Fig. 2B). To bees, the similarity between the two species in peak reflectance at short wavelengths around 450 nm outweighs the substantial difference in reflectance at longer wavelengths where hymenopteran vision is less sensitive (Fig. 2A). The hexagonal colour space model of hymenopteran vision has been used widely in ecological and evolutionary studies of floral colour (Chittka et al., 1994; Arnold et al., 2009a, b; Dyer et al., 2012; Shrestha et al., 2014; Ohashi et al., 2015; Kantsa et al., 2017; Kemp et al., 2019). Calculations for this model were based on peak sensitivity values for photoreceptors of three hymenopteran species, Apis mellifera, Bombus terrestris and Trigona spinipes (see Garcia et al., 2017, for all parameter values), sensitivity curves calculated with a template described by Stavenga et al. (1993), a standard illuminant representing noon daylight at temperate latitudes when corrected for photo flux (Judd et al., 1964), and adaptation to a leaf-green background (Bukovac et al., 2017a). Colour loci for all plant species in our samples are given in Supplementary Data Table S1.

The final step in representing colour was to translate Euclidian distances between floral colours in the hexagonal model into a metric that reflects the behaviourally demonstrated ability of different bee species to discriminate colours. The behavioural tests measure the probability that a trained bee will choose a rewarding colour over an unrewarding one as a function of the Euclidian colour distance between the two options (e.g. Fig. 2C). These responses probably reflect the particular ecological needs of a species relative to its physiological constraints (Bukovac et al., 2013). For example, foraging bees might benefit from ignoring some degree of colour difference among rewarding target flowers (Garcia et al., 2018). To date, appropriate behavioural data exist only for Apis mellifera (Dyer and Neumeyer, 2005), Bombus terrestris (Dyer et al., 2008), Trigona cf. fuscipenis and Tetragonula carbonaria (Spaethe et al., 2014). We characterized the behavioural responses of these four species using sigmoidal functions proposed for bee vision (von Helversen, 1972) and recently shown to model accurately their discrimination abilities (Garcia et al., 2017). The Apis, Bombus and Trigona response curves were used with colour model loci based on Apis, Bombus and Trigona photoreceptor sensitivities. The Tetragonula response curve was used in conjunction with the Trigona photoreceptor sensitivities, as these two genera belong to the same apid tribe, Meliponini, and represent neotropical and Australasian branches (Rasmussen and Cameron, 2007). These four species are not pollinators (at least not native, evolutionarily relevant pollinators) at our two sites, although Te. carbonaria is a native pollinator elsewhere in Australia (Heard, 1994, 1999; Nieh, 2004). However, these four species are currently the only source of appropriate data. At present, therefore, our ability to understand the diversity of pollinator visual systems, even within just the Hymenoptera, is limited. Nonetheless, extrapolation from the four species of this study to pollinator communities of greater diversity would imply an even greater variety of visual capacities and hence greater diversity of potential pollinator-mediated selective effects on floral colour. Thus, we explore the consequences of different visual systems using available data while awaiting empirical characterization of the colour vision of other pollinators.

Interestingly, the discrimination functions are species-specific and differ for different regions within colour space (Garcia et al., 2017). This latter feature is not unusual. Primate colour acuity also varies across colour space: primate photoreceptors are well tuned for discriminating fruit colours amongst foliage but provide somewhat poorer discrimination for other colours (Sumner and Mollon, 2000). In our analysis we used the functional response of each pollinator species to sets of both human-blue and human-yellow stimuli (Garcia et al., 2017). The functional responses of Tr. fuscipenis, for example, involve a more sharply defined behavioural reaction to ‘blue’ than to ‘yellow’ colour differences [Fig. 2C; see Garcia et al. (2017) for the functional responses of the other three species]. Three Euclidian colour distances in Fig. 2B are labelled I, II and III. These same distances are indicated on the horizontal axis of Fig. 2C. The distances are translated through the response curves to yield a behaviourally effective metric on the vertical axis that we label ‘perceptual distance’. Note, for example, that Tr. fuscipenis seems to judge the perceptual distance between Glossodia major and Burchardia umbellata to be very similar to the perceptual distance between G. major and Wahlenbergia gloriosa, even though the corresponding Euclidian distances (II and III) in the hexagonal colour space differ by a factor of approximately three. The same Euclidian distances would yield more clearly distinct perceptual distances if they involved pairs of ‘yellow’ colours (translating distances II and III to perceptual distances using the yellow behavioural response curve in Fig. 2C).

These methods allow us to calculate a behaviourally effective perceptual distance between any two floral colours for a given hymenopteran observer. The floral colours in a community, as perceived by each hymenopteran observer, can be summarized by the matrix of perceptual distances between all pairs of species in flower at the same time. We used the perceptual distances corresponding to both blue and yellow discrimination functions for each pollinator across all pairs of flowering species in order to illustrate the sensitivity of community colour structure not only to different observers, but also to perceptual differences within colour space by a single kind of observer.

Colour structure and phylogenetic structure at individual sample dates

Although pollinator-mediated interactions may involve time lags (Waser and Real, 1979), we expect the strongest interactions to involve co-flowering species. We therefore used the community of plants in bloom at each sample date during the 14 months of observation as the units of analysis. Colour structure in these sample units was quantified by a nearest-neighbour criterion. The nearest-neighbour distance for any focal flower is the smallest perceptual distance between its colour and another floral colour present at the same time. This criterion reflects the likelihood that a focal species will be most affected by the most similar heterospecific flowers in the community, rather than by the average similarity it has to all other community members. The mean of the nearest-neighbour distances calculated across all blooming species in a sample constituted the metric of community colour structure. We calculated this metric for each hymenopteran observer and for perceptual distances that rely on both yellow and blue discrimination functions, for each sample date. Calculations were carried out by the ses.mntd function in the R package picante, using floral abundance weighting so that the metric of community structure accounts for variable flower number among the component taxa (Kembel et al., 2010). In reference to the R function, we label the metric MNTD_colour (mean nearest-taxon distance).

MNTD_colour values are meaningful only in reference to the values at other sample dates or to values expected at random. A distribution of values expected at random (which can act as a null distribution for significance testing of the empirical MNTD_colour at a given date) was constructed by drawing species (and their floral abundances) at random from the complete species pool at each site across the entire 14-month sample. That is, random samples were constructed from species that actually occur at the site but that do not necessarily bloom together, so that random assemblages of floral colour were drawn from the complete annual pool of colours. For each empirical MNTD_colour metric, we constructed 999 equivalent random communities and calculated their MNTD_colour values. A comparison of the empirical value to the null distribution of random values provides a test of statistical significance. These tests were made with the ses.mntd function of picante using the ‘taxa.labels’ randomization method and abundance weighting of flower number. More details on the function are available in Kembel et al. (2010).

We used a similar mean nearest-neighbour criterion to characterize phylogenetic structure within co-flowering communities. The procedure is identical to that for MNTD_colour except that the inter-taxon distance in these calculations was not the colour-based perceptual difference but phylogenetic distance (the sum of branch lengths in the phylogenetic tree of the sampled species) (Webb et al., 2002). We used the angiosperm family-level topology of Soltis et al. (2011) as a scaffold to construct a phylogenetic tree for the species at our two sites. Sub-familial topology was added to it from a number of published sources (Table S2). We dated internal nodes in the tree with the maximum likelihood dates of Wikström et al. (2001). Nexus code for the resulting phylogenetic tree is given in Supplementary Data Appendix S1. We used the ses.mntd function of picante to calculate the metric of phylogenetic structure, which we label MNTD_phylo. Construction of null distributions and significance testing were conducted analogously to the procedures for MNTD_colour.

Although we believe that the nearest-neighbour criterion is more informative regarding colour structure in a community than a measure of average perceptual distance among community members, we calculated and tested the significance of such average colour and phylogenetic metrics for each sample date, hymenopteran observer, and behavioural discrimination curves for both yellow and blue discrimination. We used the picante function ses.mpd for these calculations, and we refer to the metrics of structure as MPD_colour and MPD_phylo. Results for these metrics are reported in the Supplementary Data Figs S1 and S2. They do not lead to conclusions substantively different from those drawn from MNTD_colour and MNTD_phylo.

RESULTS

Patterns of abundance

Flowering at the two sites is strongly concentrated in the spring (October to December in south-eastern Australia). The temporal pattern of species richness of blooming plants and total flower number followed a similar progression at the two sites (Fig. 3). Shannon and Simpson diversity indices also reflected a peak flowering in spring (Supplementary Data Fig. S3). Most species in the samples were uncommon and few were abundant. At Boomers Reserve, 25 species never displayed more than 20 open flowers in the sampling quadrants at any given time, while two species had more than 2000 flowers open on at least one sample date. At Baluk Willam reserve, 34 species displayed fewer than 20 flowers at a time, and the most abundant species had a maximum of 880 open flowers.

Fig. 3.

Temporal pattern of species richness and flower abundance at (A) Boomers Reserve and (B) Baluk Willam Reserve. Months are indicated by initial letters starting with March 2010. Solid line represents the number of species in bloom (scale on left vertical axis, green colour); dotted line represents total number of open flowers (scale on right vertical axis, orange colour).

Colour structure

The most common floral colours occurred in the bee-blue–green portion of the hexagonal colour space (Fig. 4) where innate colour preferences exist in honeybees (Giurfa et al., 1995; Morawetz et al., 2013), bumblebees (Gumbert, 2000; Raine and Chittka, 2007) and Te. carbonaria (Dyer et al., 2016b). This pattern is similar to the colour distribution in the floras of Israel (Chittka et al., 1994), Australia (Dyer et al., 2012) and alpine environments (Bischoff et al., 2013; Shrestha et al., 2014), where hymenopteran pollination is important (Fig. 5). Floral colour was not obviously associated with abundance: neither rare nor abundant species tended to be spatially segregated in colour space (Fig. 4).

Fig. 4.

Floral reflectance spectra mapped in the hexagonal colour-space model of hymenopteran (Apis melifera) vision. (A) Boomers Reserve, (B) Baluk Willam Reserve. Each symbol represents a single species. Symbol size is proportional to the logarithm of flower abundance at peak flowering for the species.

Fig. 5.

Frequency distributions of floral colours in 10° sections of the hexagonal colour space (inset). Distributions for the species in the two sites reported here (red and black symbols and lines) can be compared to colour distributions from species across the Australian continent (green symbols and line; data from Dyer et al., 2012), along sub-tropical and sub-alpine transects of the Nepalese Himalayas (orange symbols and line; data from Shrestha et al., 2014) and Israel (blue symbols and line; data from Chittka et al., 1994). The dotted rectangle encloses the bee-blue and bee-blue–green sectors of the colour space (30°–90°, see inset) where honeybees (Giurfa et al., 1995), bumblebees (Raine and Chittka, 2005) and the stingless bee Tetragonula carbonaria (Dyer et al., 2016b) all demonstrate their strongest innate colour preference.

Colour structure in the biweekly samples (Fig. 6) had four notable features:

Fig. 6.

Mean nearest-neighbour distances (MNTD_colour) within co-flowering assemblages at Boomers Reserve (left column) and Baluk Willam Reserve (right column). Months are indicated on the horizontal axis by initial letters starting with March 2010. Values on the vertical axis are the difference between observed MNTD_colour and the mean from 999 randomizations. Thus, negative values represent greater colour clustering and positive values represent greater colour disparity than expected at random; filled symbols indicate a value significantly different from zero (P < 0.05). Blue dotted line and yellow solid line represent analyses based on empirically determined functions for discrimination of colours in two different parts of the bee colour space (Garcia et al., 2017).

• (1) Most samples throughout the annual cycle at either site had mean nearest-neighbour perceptual distances (MNTD_color) that were not significantly different from null expectations based on random assembly from the complete annual pool of colours. Therefore, there is little evidence that floral colour similarity plays a dominant role in structuring the coexistence and blooming schedules of the species at our sites.

• (2) There were exceptions at occasional dates of significant colour clustering (e.g. September and October, Fig. 6A, H) or significant colour dispersion (e.g. December and January, Fig. 6F, H). However, these exceptions did not occur simultaneously in the two sites, indicating that communities separated by only 40 km and sharing about two-thirds of their species need not share colour structures.

• (3) The annual pattern of colour structure differed among the visual systems and perceptual responses of the four hymenopteran species that we modelled. Apis melifera vision would perceive a significantly clustered colour structure at Boomers Reserve in September and February–March (Fig. 6A) and significantly dispersed colours at Baluk Willam Reserve in December–January (Fig. 6B). In contrast, B. terrestris vision, which does not discriminate colour differences as finely as that of A. melifera, would perceive no significant departures from random structure (Fig. 6C, D). Colour structure as seen by the perceptual systems of Te. carbonaria and Tr. cf. fuscipenis fluctuated more strongly than would occur for an A. melifera or B. terrestris observer, although most sample dates did not show statistically significant departures from random expectations (Fig.6E–H). When significant clustering or dispersion under did occur under Te. carbonaria and Tr. cf. fuscipenis vision, it occurred at sample dates different from those of significant clustering or dispersion for A. melifera or B. terrestris vision.

• (4) The floral colour structure an observer perceives in a community may depend on what colours it must compare, that is, in which part of the colour space model the colours occur. Apis melifera and B. terrestris visual systems can often discriminate relatively fine colour differences, while Te. carbonaria and Tr. cf. fuscipenis visual systems may generalize flower spectral signals so that similar colours are not discriminated (Fig. 6A–D). For example, the perceptual distances among co-flowering species in the December–January samples at Baluk Willam Reserve would be considered significantly dispersed by Te. carbonaria if the distances involved human-yellow flowers but not if they were human-blue (Fig. 6F). Trigona cf. fuscipenis, however, would not consider the same set of flowers to be significantly dispersed if they were human-yellow but would if they were human-blue (Fig. 6H).

Colour structure measured by the overall mean perceptual difference among co-flowering species (Supplementary Data Fig. S1) differs only in detail from the mean nearest-neighbour distances shown in Fig. 6. The four conclusions given above would be drawn equally from the results in Fig. S1. Taken together, these data show that although floral colours in these communities seldom differ more than would be expected from random assembly, the colour structure is variable over an annual cycle and the pattern of variability depends upon the visual profiles of different bee observers.

Phylogenetic structure

Phylogenetic structure among co-flowering species was variable over time at each site, and the pattern differed between the two sites (Fig. 7). MNTD_phylo was generally low (co-flowering species were related) in winter and early spring (July−September) and rose (more phylogenetic disparity) from mid-spring to summer. Most of this variation was not significantly different from random expectations. Co-flowering assemblages were, however, significantly overdispersed during the mid-spring flowering peak (October) at Baluk Willam Reserve and in mid-summer (December−January) at Boomers Reserve (Fig. 7). There were no instances of significant phylogenetic clustering at any sample date at either site. Phylogenetic structure as indicated by mean pairwise distance (MPD_phylo) showed generally similar patterns (Supplementary Data Fig. S2). Significant phylogenetic over-dispersion occurred in mid-summer at Boomers reserves but never occurred at Baluk Willam Reserve, and no instance of significant phylogenetic clustering occurred at either site.

Fig. 7.

Mean nearest-taxon phylogenetic distance (MNTD_phylo) within co-flowering assemblages at Boomers Reserve (magenta diamonds, interrupted line) and Baluk Willam Reserve (brown circles, continuous line). Interpretation is as for Fig. 6 except phylogenetic distance rather than colour distance is involved. Filled symbols indicate a value significantly different from zero (P < 0.05).

A comparison of these annual patterns in phylogenetic structure with the annual patterns of colour structure indicates that the two were unrelated. Significant clustering in the colour metric did not correspond to significant phylogenetic clustering (which never occurred), and significant colour dispersion occurred on sample dates different from those of significant phylogenetic dispersion.

DISCUSSION

In the current study, we evaluated the structure of floral colours in two communities of herbaceous, native plants in Australia over a 14-month period. Common ideas on the ecology of flower colour have typically emphasized pollinator constancy as a probable mechanism driving competition for pollinator attention (Levin and Anderson, 1970; Chittka et al., 1999; Brown and Mitchell, 2001; Dyer and Chittka, 2004; Fishman and Hadany, 2015; van der Kooi et al., 2016), although co-flowering species have also been implicated in facilitating pollination within a community (Brooker et al., 2008; Liao et al., 2011; Kantsa et al., 2017; Kemp et al., 2019). Had competition for pollinator service been the main driver of colour signalling in these communities, we would have expected co-flowering species to distinguish themselves from each other with colours that were more dispersed than the average in randomly assembled samples. Had facilitation in pollinator attraction been an important determinant of coexistence, we would have expected more colour clustering than occurs at random. Our results suggest that neither process is a strong force shaping the assemblage of floral colours at our field sites (Fig. 6).

Moreover, our results show that perceived colour structure varies among bee species that have evolved different sensory capabilities (Fig. 6). Limited data on insect colour vision required us to use as visual models four bee species that were not native to our sites, but even this small sample of hymenopteran visual diversity illustrates the sensitivity of community colour structure to the perception of the observer, even among a group of insects with a fair degree of phylogenetic conservatism in their photoreceptor sensitivities (Briscoe and Chittka, 2001).

Given the predominance of random community colour structure, why do flowers tend to use a limited array of colour signals, drawn especially from the bee-blue–green portion of the colour hexagon (Fig. 4), colours which are known to be closely linked to the visual capabilities of bee pollinators (Chittka, 1992; Dyer et al., 2012; Shrestha et al., 2013)? One possibility is that floral colours within a community function more for easy detection by pollinators, rather than for discrimination among co-flowering species (Dyer et al., 2008; Streinzer et al., 2009; Bukovac et al., 2017b), at least at the spatial scale of herbaceous plants in forest understoreys. In this case, colour distance between species would be of less importance than how well chromatic cues differentiate a flower from its background (Bukovac et al., 2017a, b). Another possibility is that floral colours are adapted to innate pollinator colour preferences. Innate colour preferences are difficult to measure in behavioural experiments (Kemp et al., 2015; van der Kooi et al., 2018), but measurements have been made for the honeybee (Giurfa et al., 1995; Morawetz et al., 2013), the bumblebee (Gumbert, 2000; Raine and Chittka, 2007), and the Australian native bee Te. carbonaria (Dyer et al., 2016b). All three bee species show their strongest preference in the short-wavelength ‘blue’ region of the electromagnetic spectrum. The prevalence of bee-blue–green floral colours at our sites is similar to that found in colour distributions worldwide (Fig. 5), raising the possibility that our study may be representative of processes more globally. Of course, detection and innate preference might not be independent, because the ability to detect a visual signal could be associated with innate preference for the signal to enhance foraging efficiency (Lehrer et al., 1995; Mery and Kawecki, 2004). Indeed, Bukovac et al. (2017b) built multi-agent simulations to explore how well discrimination may explain flower evolution and found that there had to be other factors in addition to discrimination, such as the capacity of a colour signal to promote detection.

A recent study of pollinator attraction in Mediterranean scrubland communities used an expanded perspective on floral signalling and found that flower colour, measured in a colour hexagon, was often linked to key floral scents, but different species were not using colour or scent as alternative signals to promote flower constancy by bees (Kantsa et al., 2017). This also suggests that promoting pollinator visitation is of more value than maximizing identification through distinct signals.

Our analysis of community colour structure used a model of colour vision to translate physical reflectance spectra of flowers to colour loci that represent pollinator perception of colour similarity and dissimilarity. Such colour models are being increasingly used in studies of floral ecology (Arnold et al., 2009a, b; Dyer et al., 2012; Burd et al., 2014; Shrestha et al., 2014; Ohashi et al., 2015; Kantsa et al., 2017, 2018; Kemp et al., 2019). No previous community study, however, has represented colour differences in terms of the ability of bee pollinators to discriminate colours in choice tests, as we have done here using discrimination functions informed by the behaviour of bees (Garcia et al., 2017). Such behaviourally informed measures of perceptual colour distance should allow colour structure metrics to reflect more closely the pollination ecology within the community. Still needed are data on a much broader variety of pollinators in order to make these analytical advances more widely relevant. It is noteworthy that Kemp et al. (2019) were able to use A. melifera visual models in South African plant communities where A. melifera is a native and evolutionarily relevant pollinator. Our result show that inferences regarding the role of colour signalling in competition, facilitation, detection and innate preferences are likely to be quite sensitive to the details of pollinator vision. The conundrums of how plant–pollinator interactions operate depend very much on what observer is making colour judgements.

SUPPLEMENTARY DATA

Supplementary data are available online at https://academic.oup.com/aob and consist of the following. Table S1: List of species with their bee-vision colour space coordinates. Table S2: Phylogenetic studies used for sub-familial topology in the composite tree assembled for phylogenetic comparative analyses. Figure S1: Colour structure as indicated by the MPD_colour metric. Figure S2: Phylogenetic structure as indicated by the MPD_phylo metric. Figure S3: Shannon and Simpson diversity index. Appendix S1: Nexus code for the phylogenetic tree used in the analyses.

FUNDING

M.S. was supported by a Monash Graduate Scholarship (MGS), Monash International Postgraduate Research Scholarship (MIPRS), and Faculty of Sciences Postgraduate Publication Award (PPA) while collecting and analysing data. M.S. also acknowledges the School of Media and Communication, RMIT University for support and Associate Professoor Alan Dorin, Faculty of Information Technology, Monash University, for support while analysing data. A.G.D. was supported by Australian Research Council Discovery Projects grant DP160100161. M.B. was supported during the data analysis by the Sabbatical Scholars programme of the National Evolutionary Synthesis Center (NESCent), Durham, NC, USA.