Nectar robbing (consuming nectar from a perforated flower without pollinating) generally negatively affects plant fecundity, and plants exhibit multiple mechanisms in defence. The colour of bird-pollinated flowers that are prone to robbing by bees may consequently be explained by the bee-avoidance hypothesis, proposing that bird-pollinated flowers are less visible to bees than bee-pollinated flowers. This visual avoidance mechanism may co-occur with mechanical avoidance mechanisms; however, flower conspicuousness has not yet been related to nectar robbing rates across multiple species using a combination of comparative methods and field observations. To test the bee-avoidance hypothesis, we quantified conspicuousness and flower visitor rates in multiple congeneric ornithophilous and entomophilous Erica species from South Africa. Visual modelling and Phylogenetic Generalised Least Squares analyses of spectral reflectance of 33 ornithophilous and 31 entomophilous species revealed that bird-pollinated flowers were less conspicuous (relative to their leaf background) to honeybees than insect-pollinated flowers, but equally conspicuous to birds, supporting the bee-avoidance hypothesis. However, natural nectar robbing rates in 27 populations of bird-pollinated species were unrelated to conspicuousness to bees, and pollination rates were unrelated to conspicuousness to birds. Finally, the conspicuousness of bird-pollinated flowers to bees and birds was unrelated to mechanical traits associated with nectar robbing (variation in corolla length, corolla aperture diameter, sepal size, or corolla stickiness). These results suggest that nectar-robbing bees may have influenced the evolution of flower colour of sunbird-pollinated species, although the maintenance of this trait appears decoupled from current robbing pressure.

Study species

To obtain spectral reflectance and morphological measurements, we collected flowers from 33 ornithophilous (i.e., with bird-pollination traits) and 31 entomophilous (i.e., with insect-pollination traits) Erica species occurring in the Fynbos, South Africa, from both natural habitats and from Kirstenbosch National Botanical Garden (Cape Town) (Fig 1, Table S1). All species are shrubs with bisexual flowers that last several days. We sampled species randomly with respect to phylogeny and colour. For 18 species, we sampled multiple populations, including different colour morphotypes (from different sites, except for two sites that contained multiple morphs of the same species), for a total of 108 populations. This research was approved by South African National Parks (CRC/2016-2017/020—2016/V1) and CapeNature (AAA043-0008-0056).

We classified species into pollination syndrome categories based on flower size and shape, but not colour (Rebelo & Siegfried, 1985). We measured morphological traits with digital callipers (to two decimal places) (Fig S1) on 3–28 flowers per population (limited by the number of available flowers), and calculated means per population. We measured the outer diameter of the corolla aperture (where corolla tips separate); corolla length from the aperture to the receptacle; pistil length from the receptacle to the stigma; and stamen length from the receptacle to the anther tip. Sepal surface area was quantified as the product of sepal length and width (at the widest point). Flowers with short corollas, or long corollas with narrow corolla apertures (< 2 mm), were classified as entomophilous, and flowers with long corollas and large apertures as ornithophilous (Fig 1). Sampled species showed bimodal distributions in corolla length, aperture diameter, pistil length, stamen length, and sepal surface area, all of which differed significantly between assigned syndromes, supporting these classifications (Fig S2). Another exertion length also differed significantly, ly but was not as distinctly different between the syndromes and therefore was not included in syndrome classification. Published pollinator observations were available for 57 populations, and for the remaining populations, we predicted syndromes based on flower morphological traits. A Principal Component Analysis using corolla, pistil, and stamen length (n = 81 populations) revealed that the morphology of species with predicted syndromes overlapped completely with those species with confirmed syndromes of the same type (Fig S3). Since pollinator data are lacking for most Erica species (van der Niet, 2020), syndrome categories could not be refined further. Our sampled ornithophilous species were dominated by pink and red flowers, and entomophilous flowers by white and pink, but our sample also included purple, cream, green, yellow,w and orange (according to human vision; Fig 1I-J, Table S2, Fig S4). Blue flowers are absent in the genus (Rebelo & Siegfried, 1985).

Reflectance measurements

We measured the spectral reflectance of the corollas of at least three fresh flowers and two leaves per population. To quantify the most obvious signal, we measured the colour covering the largest area of the perianth in bi-coloured flowers and the sepals when the sepals covered two-thirds of the corolla. We took measurements indoors using a calibrated spectrophotometer (Ocean Optics Jaz models), with a QR200-7-UV-VIS reflectance probe and PX-2 Pulsed Xenon light source (Ocean Optics, Dunedin, FL) and a Spectralon 99% white standard (Labsphere, Congleton, UK), with the probe placed at a 45° angle and 5 mm distance from samples. We captured reflectance data with Spectrasuite software using an integration time of 30 milliseconds, average scans of 5, and a boxcar width of 5. Reflectance spectra of 300–700 nm were interpolated to 1 nm bins, smoothed (span = 0.2; Fig S5), and negative values were removed by adding the lowest reflectance value to all wavelengths (thus preserving the spectrum shape). All processing and modelling of reflectance data was conducted with the pavo 2.7.1 package (Maia et al., 2019) in the R environment (R Core Team, 2022).

Visual modelling

To illustrate how flower visitors likely view colour distances, we modelled the reflectance data as colour loci points in the tetrahedral visual space for birds (Goldsmith, 1990) and hehexagonalisual space for bees (hyperbolic-transformed quantum catches) (Chittka et al., 1992) using the visual models described below (Fig.). 1K- L.

Two complementary approaches (conspicuousness and visual discrimination abilities) were used to assess how bees and birds perceive flower colours. Conspicuousness quantifies how easily an object can be distinguished from its background by comparing an object’s reflectance with its background reflectance. This is modelled for a taxon using the known maximum absorbance values of its photoreceptors. Visual discrimination determines if a colour is adapted to a visual system by comparing its reflectance spectrum to a taxon’s discrimination optima (Shrestha, Dyer, & Burd, 2013). We compared the conspicuousness and discriminability of ornithophilous and entomophilous flowers by calculating both metrics at the population level.

Conspicuousness

We quantified flower conspicuousness as chromatic and achromatic distance (in Just Noticeable Differences, JNDs (Vorobyev & Osorio, 1998)) using the Receptor Noise-Limited Model. Chromatic contrast quantifies the colour difference, and achromatic contrast quantifies the brightness difference between a flower and its background. We modelled the visual systems of ultraviolet-sensitive birds (maximum photoreceptor absorbance at 372, 456, 544, and 609 nm (Endler & Mielke, 2005))and honeybees (Apis mellifera; maximum absorbance at 344, 43,  and 544 nm). Since visual systems within Hymenoptera are relatively conserved (Peitsch et al., 1992), the Apis mellifera visual system is appropriate for testing our predictions.

Using the vismodel function, we applied the visual models with the logarithm of the quantum catches, the standard daylight function (D65) as illuminant, scaled by 10,000 for bright illumination, and a typical “green” background (Stoddard & Prum, 2008). For chromatic contrasts (noise-weighted Euclidean distances, using the coldist function), we applied the more conservative Weber fraction of 0.1 (Maier & Bowmaker, 1993) and the neural noise model, which is appropriate for bright daylight conditions. In bird vision, we specified the photoreceptor ratio of (1, 2, 2, 4) and for achromatic contrasts we applied the Weber fraction 0.34 (Olsson et al., 2018) and double cone spectral sensitivity function of Blue Tit (Hart et al., 2000), and in bee vision, we applied a photoreceptor ratio of (1:1:1) and achromatic contrast as the difference between flower and background stimuli in the green photoreceptor with a Weber fraction 0.16. We calculated chromatic contrast for each Erica population (n = 101) by comparing each flower’s corolla reflectance to its own leaf reflectance and then averaging across flowers (3-28 flowers) within each population. For 13 populations without leaf reflectance data, we used average reflectance from leaves of 37 species (representing both pollination syndromes), giving 100 populations for this analysis.

Colour discrimination abilities of pollinators

Although using visual systems is a powerful way of analysing reflectance spectra, this method makes assumptions regarding perception. We therefore also used inflection point analyses, which do not rely on physiological data (Shrestha, Dyer, & Burd, 2013). This approach assesses the overlap between the discrimination abilities of a specific visual system and the inflection points (rapid changes of slope) in a reflectance spectrum. Flower colours are thought to be adapted to a pollinator’s discrimination ability when their inflection points are closely matched to the hue discrimination maxima. The hue discrimination maxima of birds and bees differ: 416, 489 a, nd 557 nm for ultraviolet-sensitive (UVS) birds and 400 and 500 nm for hymenopteran insects. We first averaged reflectance spectra within populations and smoothed them (span = 0.2), before identifying wavelengths at which major inflection points occurred (> 10% change in reflectance within 50 nm) using the Spectra-MP program (Dorin et al., 2020). We then calculated the absolute distance between each inflection point and its closest visual optimum. For each spectrum, we calculated mean absolute deviation (MAD) as the mean distance for all inflection points in the spectrum. We also calculated minimum absolute deviation (minAD) for each discrimination optimum: the distance to its closest inflection point (Fig S6). Of 104 populations, 33 had no major inflection points in their reflectance spectra and were therefore excluded. These included diverse colours, and 27 (ca. 82%) were ornithophilous. This left 71 populations (37 ornithophilous, 34 entomophilous) for this analysis.

Comparing the visibility of flowers

We compared the conspicuousness (chromatic and achromatic contrasts) of ornithophilous and entomophilous flowers in bee and bird vision separately. Similarly, we compared the discrimination ability (one model for each metric) between pollination syndromes. We used Phylogenetic Generalised Least Squares (PGLS) analyses to account for non-independence due to phylogenetic relatedness (pgls in the caper package (Orme et al., 2013) with Maximum Likelihood estimate of lambda). We constructed the phylogenetic hypothesis (Fig S7) by excluding taxa from a rate-smoothed chronogram of Erica (Pirie et al., 2024) inferred under maximum likelihood from a dataset extended from (Pirie et al., 2016). Five species could not be represented in the phylogeny due to sequence paralogy (Pirie et al., 2024), giving a final sample of 59 species. Not all of the pertinent subspecific variation was represented directly in the original phylogeny. Where there was evidence for species monophyly (i.e., from multiple samples), we included populations manually as additional taxa, arbitrarily resolving internal nodes with minimal length branches and adding terminal branch lengths consistent with those within other species in the phylogeny.

Nectar robbing and pollination rates

We sampled robbing and pollination rates in 11 bird-pollinated Erica communities occurring in undisturbed natural fynbos habitat during 2016–2017 (see Coetzee et al., 2021), for locations, seasons, and other details. All study communities were at least 5 km apart, and each included two or more co-flowering ornithophilous Erica species. At each site, we surveyed 3–28 flowers per plant and scored each flower as robbed, pollinated,d or unvisited. We identified robbed flowers as those with a slit or hole at the base of the corolla, and pollinated flowers by a disturbed anther ring (Geerts & Pauw, 2011). From these data, we calculated cumulative visit rates as the proportion of visited flowers (robbed/pollinated) out of the total flowers surveyed per plant (Davies et al., 2025). We then averaged this proportion across all plants in a population (8–40 plants per population).

We tested whether floral visitor rates were related to flower visibility at the population level using Linear Mixed Effect Models (LMM, using the me4 package) with site identity as a random intercept effect. For the nectar robbing rate, we included conspicuousness to bees (chromatic contrast), corolla length, and stickiness as predictor variables. We repeated this analysis, replacing conspicuousness with discriminability (MAD in bee vision, the metric that summarises information of the whole visual system). For pollination rate, we used the same approach, but with conspicuousness to birds, corolla length, and stickiness as predictors, and repeated the analysis with discriminability (MAD for bird vision) instead of conspicuousness. The final dataset comprised 27 populations, from 12 different species. Insufficient data were available for corolla aperture and sepal size, so these could not be included in the analysis. Power analyses showed these tests only had sufficient power to detect large effect sizes. Statistical model assumptions were checked with Q-Q plots and Residual vs Fitted plots. We report adjusted R² for each analysis.

Bee-avoidance traits

We identified four mechanical bee avoidance traits based on the literature and on the observation that they differed significantly between ornithophilous and entomophilous Erica flowers: corolla length, aperture diameter, sepal size (Fig S2), and stickiness. We classified corollas as sticky (any degree of viscosity detectable by touch) or non-sticky. Stickiness was almost exclusively found in ornithophilous species: only one entomophilous species was sticky, whereas 30 ornithophilous populations were sticky and 36 non-sticky. Although corolla hairiness and flower bracts could also function as mechanical defence traits, we did not include them because most populations had glabrous corolla surfaces, and because the bracts of Erica flowers are mostly smaller than the sepals.

Focusing on 26 ornithophilous species, we tested for associations between flower conspicuousness (for birds and bees separately) and four mechanical bee-avoidance traits separately: corolla length (n = 56 populations), corolla aperture diameter (n = 32), sepal surface area (n = 35), and corolla stickiness (n = 59). Sample sizes varied because not all morphological measurements were available for all populations. Erica halicacaba L. was removed as an outlier for analyses with corolla aperture, since its closed flowers require mechanical force by birds to open the corolla (Turner et al., 2012). We also tested for associations between achromatic contrast and mechanical traits. We conducted one PGLS analysis for each mechanical avoidance trait, with chromatic contrast asthe response variable and the mechanical avoidance trait as the predictor variable. All p-values were adjusted for false discovery rates following Benjamini & Hochberg (Benjamini & Hochberg, 1995). Statistical model assumptions were checked with Q-Q plots and Residual vs Fitted plots. We report adjusted R² for each analysis.