Study species and sites
Weigela hortensis (Caprifoliaceae) is a deciduous shrub that occurs in mountainous areas in Japan. From May to June, it produces many pale rose, bell-shaped flowers with 25–40 mm-long corolla tubes. Flowers are self-incompatible and typically last for 4–5 days (Suzuki and Ohashi 2014). Our preliminary survey showed that the number of ovules per flower was 69 ± 13 (mean ± s.d., n = 39).
The three study sites lie at different altitudes on Mt. Izumigatake, northern Honshu, Japan: Yoshinodaira (38°23′15″ N, 140°43′00″ E, 518 m a.s.l.), Otaira (38°24′51″ N, 140°43′14″ E, 650 m a.s.l.) and Taiwa (38°25′24″ N, 140°42′35″ E, 812 m a.s.l.). On Mt. Izumigatake, W. hortensis flowers are mainly visited by bumble bees (Bombus, Apidae), small bees (including Andrenidae and Halictidae), hunch-back flies (Oligoneura spp., Acroceridae) and hoverflies (Syrphidae) (Hasegawa et al. 2023). There are variations in assemblages of flower visitors along the altitudinal gradient; small bees are the predominant visitors at low-altitude sites whereas hunch-back flies increase in relative abundance with increasing altitude (Hasegawa et al. 2023).
Pollination effectiveness of flower visitors
From 24 May to 22 June 2022, we performed a field experiment to examine whether flower visits by different taxa of insects resulted in seed production at Yoshinodaira and Taiwa. At each site, 21 W. hortensis plants were used in the experiment. On each of 1–3 branches of each plant, 2–7 buds were selected and all anthers in the buds were removed using forceps. This emasculation prevented pollen import by visitors from being influenced by the presence of pollen grains produced by the receiving flower. The branches were then bagged to exclude insects until visit observations were made. After flowering of the emasculated buds, the bags were removed to allow insects to visit the flowers for about 60 min. During this period, insect visits to flowers on each branch were recorded using a digital video camera (HDR-CX420, HDR-CX680, Sony, Japan; GZ-R400, GZ-RX690, JVCKENWOOD, Japan). We marked the emasculated flowers and measured their corolla tube lengths. After 60 min of open pollination, the branches were bagged again until the corollas abscised. Visitor observations were conducted for three days at each site. In total, visitors to 155 flowers on 33 branches (of 21 plants) at Yoshinodaira and 137 flowers on 36 branches (of 21 plants) at Taiwa were recorded. After fruit maturation, the marked fruits were collected and all seeds in the fruits were counted.
From the video recordings, visitors that contacted stigmas of the emasculated flowers were counted and classified into the following four groups: bumble bees, small bees, hoverflies and hunch-back flies. Small bees possibly included Andrenidae, Halictidae and Apidae although their families were rarely identifiable on the video recordings. Female bumble bees and small bees have pollen transport structures on their hind legs called corbiculae and scopae, respectively (Thorp 1979). Bumble bee and small bee groups were subdivided according to whether the corbiculae or scopae contacted stigmas, as determined by the video recordings. Since the scopa of a small bee is located around the hind legs, we determined that the inside surfaces of scopae contacted stigmas when small bees climbed on the stigmas.
Morphology and conspecific percentages of pollen carried by insects
From 31 May to 11 June 2021, we sampled visitors to W. hortensis flowers at Otaira and Taiwa to compare morphology of W. hortensis pollen carried by visitors among pollinator taxa and among body parts within taxa. Bees and flies were captured immediately after leaving W. hortensis flowers using plastic vials or a sweep net. Insects captured using a sweep net were then placed into plastic vials. The insects held individually in vials were immediately chilled on ice, transported to the laboratory, preserved at − 20℃ for one day, and dried with silica gel at room temperature for more than one month. The insects were sent for taxonomic identification by experts after the pollen studies described below.
Pollen was sampled from the dorsal thoraxes of both bees and flies and from the corbiculae/scopae on the tibiae of the bees’ hind legs. The dorsal thorax of corbiculate bees is known as one of the ‘safe sites’, where pollen is less likely to be groomed away than it is from other body parts (Koch et al. 2017; Tong and Huang 2018). For analysis with scanning electron microscopy, pollen on each body part of an insect was gently removed with a conductive carbon double-sided tape (5 mm width, Nisshin-EM, Japan) affixed to a cylinder specimen mount (10 mm diameter, Nisshin-EM, Japan). Because corbicular pollen, mixed with regurgitated nectar (Michener 1999), agglutinated more firmly when dried, it was split in two using forceps, and the section was gently pressed on a double-sided tape. Pollen on specimen mounts was then coated with platinum on an ion sputter coater (E-1045, Hitachi High-Tech, Japan) and observed using a scanning electron microscope (S-3400N, Hitachi High-Tech, Japan) at 3.0 kV. For each specimen mount, approximately eight W. hortensis pollen grains were randomly selected and photographed at 1000x. Then, the diameters and spine lengths of pollen grains in the images were measured using a program developed in Mathematica 11.1 (Wolfram Research 2017). The diameter of a pollen grain was defined as the diameter of the inscribed circle inside its outline. The spine length was defined as the mean length of the five longest spines forming a part of the outline of the pollen grain. In addition, to compare pollen traits between pollen on insects and in flowers, we reanalyzed images of pollen grains sampled from flowers (Hasegawa et al. 2023). These images were collected in 2020 at the same sites used in this study. Pollen was sampled from five fresh flowers of each of ten individual plants at each site, and six to ten pollen grains from each flower were photographed (Hasegawa et al. 2023).
Pollen carried by insects captured on W. hortensis flowers originated not only from W. hortensis flowers but also from heterospecific flowers. The percentages of W. hortensis pollen grains were estimated as one of the factors influencing pollination efficiency of insect visitors. For each specimen mount, three or four images of pollen grains were taken at 100x or 200x using a scanning electron microscope to include as many pollen grains as possible regardless of donor species. Up to 50 pollen grains for each specimen mount were classified according to whether they were W. hortensis or heterospecific. For specimen mounts with sparce pollen, we counted all W. hortensis and heterospecific pollen grains without taking images.
Statistical analyses
All statistical models were developed in PyMC3, a Python probabilistic programming framework for Bayesian parameter estimation (Salvatier et al. 2016). In all Bayesian statistical analyses described below, the No-U-Turn Sampler was used to generate four Markov chain Monte Carlo (MCMC) chains each with 10,000 iterations following a burn-in period of 10,000 iterations. The potential scale reduction factors (R-hat) were below 1.01 for all parameters, indicating convergence of the MCMC chains. Variance inflation factors in the models with multiple predictors were at most 1.25, suggesting that there were no problems of multicollinearity (Dormann et al. 2013). The fieldwork was done at multiple study sites, so we developed separate statistical models for each site.
We ran multiple regressions to test for the effect of visits by different visitor groups on seed production. We used the visit and seed data for flowers receiving at least one visit with a stigma contact. However, we excluded the data for flowers on branches that did not produce any fruits with seeds, as these branches may have lacked energy. As a result, the visit and seed data for 40 fruits (on 16 branches of 12 plants) at Yoshinodaira and 96 fruits (on 30 branches of 19 plants) at Taiwa were analyzed. The multiple regression models included the number of seeds per fruit as an outcome variable and the number of visits with stigma contacts by different visitor groups and standardized corolla tube length as predictors. For bees, visits with stigma contacts by bodies or by corbiculae/scopae were included as separate predictors. The models also included branches as a random factor to account for repeated measures. Although we observed insect visits to 1–3 branches on the same plants, plants were not held as a random factor because only one branch was used for eight out of 12 plants at Yoshinodaira and 12 out of 19 plants at Taiwa. A negative binomial sampling distribution was used with a log link function.
To compare pollen traits among pollen sources, we developed statistical models estimating the means and standard deviations of trait distributions of pollen grains from different pollen sources. The estimations for pollen sampled from insects were performed with a single model that included individual insects as a random factor. The models for pollen from flowers included plants and flowers as random factors. For pairwise comparisons between pollen sources, posterior distributions of differences between the means of trait distributions were estimated for all pairs of pollen sources using the MCMC samples. A gamma sampling distribution was used for both pollen grain spine length and pollen grain diameter.
To test for the effects of pollen traits on the probability of pollen collection from the bodies to the corbiculae/scopae of bees, we ran separate multiple logistic regressions by the bee group. We used only the pollen data where pollen was sampled from both the bodies and corbiculae/scopae of the same individual bees. The multiple regression models included whether pollen was collected or not as an outcome variable (1.0 for pollen collected into the corbiculae/scopae, and 0.0 for pollen remaining on the bodies), standardized spine length and diameter as predictors and the interaction of the predictors. We also incorporated the quadratic terms of both predictors in the models to distinguish whether the collection probability was highest at an extreme or intermediate pollen trait. When the estimate of a quadratic term was negative, we estimated an intermediate pollen trait value achieving the highest collection probability using the MCMC samples. The regression models also included individual bees as a random factor. A Bernoulli sampling distribution was used with a logit link function.
To compare W. hortensis pollen percentages among pollen sources, we developed statistical models estimating W. hortensis pollen percentages from different pollen sources. A beta-binomial sampling distribution, which allowed us to account for overdispersed count data (Gelman and Hill 2007), was applied to the numbers of W. hortensis and heterospecific pollen grains with a logit link function. The estimations for pollen sampled from pollen sources to be compared were performed with a single model. For pairwise comparisons between pollen sources, posterior distributions of differences between the W. hortensis pollen percentages were estimated for all pairs of pollen sources using the MCMC samples.