Cognitive abilities evolve within the context of ecological communities. Honeybees and bumblebees have become model systems for cognitive ecology; but pollination is performed by a diverse group of insects under similar pressures to forage efficiently in a mixed floral community. We studied the colour learning abilities of six species of Hymenoptera (two eusocial bumblebees, a cuckoo bumblebee, two wasps, and a leaf-cutter bee) within the context of an island plant community. We used records of insect visits to flowers in the field to determine the index of specialization of each species in the island plant-pollinator network, and measured the spectral reflectance of the flowers they visit. Species with higher specialization indices in our plant-pollinator network made a larger proportion of correct choices in a colour learning task than more generalist species. The more generalist species also visited a group of flowers more similar to each other in Hymenopteran colour vision space. These results indicate that better colour learning abilities may enable insects to forage on plants of different colours, whereas more generalist insects are visiting flowers that are similar in colour, and therefore are less reliant on repeated colour learning to forage efficiently.

(a) Plant-pollinator network

This study was conducted at the Bowdoin Scientific Station on Kent Island, an 80-hectare island in New Brunswick, Canada. Kent Island is 32 km to the closest mainland in Cutler, Maine, 5km from the closest larger island of Whitehead NB, and 8km to the larger island of Grand Manan, NB. In June and July of 2019, 2022, and 2023 one to two researchers collected insects when they were observed visiting flowering plants across the island. Across all three collection years insects were identified to the lowest taxonomic group possible using keys and by posting photos on iNaturalist. Tissues from a subset of 95 insects from 2022 were submitted to the Canadian Center for DNA barcoding for identification using the CO1 barcoding region (BOLD Project "Identifying Insect Flower Visitors in Plant Pollinator Network on Kent Island, New Brunswick Canada").

We used the ‘bipartite’ package [28] in R to visualize the plant-pollinator network and calculate indices of specialization of the insect species. To assess how representative a network is from a given year we compared the network data between years with a Jaccard similarity index for interactions using the ‘vegdist’ command in the R package ‘vegan’ [29]. We evaluated sampling completeness with a Chao 2 estimator [30, 31], using the ‘specpool’ command in the ‘vegan’ package. To determine specialization indices for pollinator species in the network (d’) [32] we used the ‘dfun’ command in the ‘bipartite’ package.

(b) Floral Reflectance Spectra

In 2022 and 2023 we measured the reflectance spectra of collected flowers using an Ocean Optics Flame miniature spectrometer with a DH-2000 BAL UV–VIS-NIR light source and PTFE diffuse reflectance standard (Ocean Optics, Orlando FL USA). For each plant species we measured floral reflectance from three different representative specimens and used the average reflectance of each wavelength from the three specimens for further analysis. We plotted our stimuli in the hexagonal colour vision space of the honeybee Apis mellifera [33] using the R package ‘pavo 2’ [34]. We used the honeybee spectral sensitivities for our visual model as while there are published spectral sensitivies for some bumblebee species [35, 36], they do not include the species we tested, and there are not spectral sensitivities published for our wasps or leaf-cutter bees. Hymenopteran colour vision is considered quite conserved [37-39], so we felt the most appropriate system to use was the well-studied honeybee. To compare how similar the two stimuli used in our colour learning experiment were to real flower colours, we created vectors of pairwise chromaticity distances between a test stimulus and the flowers in the community and compared the vectors using a two-sample t-test. We used a linear model to examine the relationship between the d’ (index of specialization) of the species we tested in our behaviour experiment and the pairwise chromaticity distances between flowers they visited in our network.

(c) Colour Learning Experiment

We were interested in how abilities to associate colours with rewards may relate to the range of flower colours insects are visiting in the field. In 2022 and 2023, we tested associative colour learning abilities of 6 hymenopteran species: Bombus flavidus (N = 15), Bombus sandersoni (N = 87), Bombus ternarius (N = 19), Dolichovespula arenaria (N = 40), Dolichovespula norvegicoides (N = 44), and Megachile melanophaea (N = 13). Insects collected on flowers in the field were returned to the laboratory and transferred into clean 50 mL tubes with a modified cap into which we inserted transparency paper with two holes ~1.5 cm apart. Insects were allowed to acclimate to the new tubes for 30-90 minutes before testing. Colour learning was tested using Free-Moving Proboscis Extension Response (FMPER) [40] following established testing protocols [41]. Insects were trained to associate either blue or yellow strips of paper (conditioned stimulus, CS) with a 50% sucrose solution reward (unconditioned stimulus, US). Papers were Neenah Astrobrights “Sunburst Yellow” and “Blast-Off Blue” (Neenah Paper, Atlanta, Georgia, USA). Treatments were balanced such that for each species an equal number of individuals received blue as the rewarded CS (CS+) and yellow as the CS+.  Each insect received two training trials and five choice trials. For the training trials we dipped the designated rewarded colour strip in 50% sucrose solution (CS+) and inserted it into one side of the cap (we alternated the offered side between individuals). We allowed insects to drink from the paper for a maximum of 3 seconds and then removed the strip and inserted the alternative colour strip which had been dipped in water (CS-) inserted into the opposite side of the cap. Insects were allowed to drink the water from the CS- for three seconds and then it was removed, or if insects did not drink we removed the strip after the insect touched it with an antenna. For the second pre-training trial we switched the sides of the presentation of the two stimuli. There was a ~10 minute interval between each training trial and between training trials and the choice trials.

After the two training trials we gave each insect five subsequent rewarded choice trials. For choice trials we simultaneously inserted the CS+ (with sucrose) and CS- (with water) and the first strip the insect touched with its antennae or proboscis was recorded as its choice. The insect was allowed to drink for 3 seconds before the strip was removed. The other strip was left in the tube until the insect drank from it for 3 seconds or touched it with its antennae. We repeated this for a total of five times, alternating the colour presentation sides each time with 10 minutes between each trial. To prevent re-testing the same insect we placed a dot of fluorescent pink acrylic paint (Sargent Art, Hazleton PA USA) on the thorax of each insect and released them back into the field. We analyzed the performance of insects using binomial generalized linear mixed-effect models (GLMM) in the ‘lme4’ package [42] in R with a choice of the CS+ as 1 and a choice of the CS- as 0. Treatment (whether the CS+ was blue or yellow), insect species, the interaction between treatment and insect species, trial number, and the interaction between trial number and species were included as fixed effects and individual insect ID’s were included as a random effect.