where a is the semi-major axis, or half the length or width (whichever was longest) of a flower’s receptacle or capitulum, bis the semi-minor axis, or half the length or width (whichever was shortest) of a flower’s receptacle or capitulum, and h is the height of a flower or inflorescence (Fig 2c and Appendix Table A2).
To determine which measurement of floral dimensions was the best proxy for FR amount, literature searches for daily nectar sugar mass (µg/day) and pollen volume (in µl/flower) were conducted for all flowering species encountered; these measurements have been previously used to assess FR available to pollinators (Hicks et al. 2016, Baude et al. 2016). Literature sources that provided counts of pollen grains per flower and volumes of individual pollen grains were used to calculate an estimate of pollen volume per flower for species for which we could not find measurements of total pollen volume. Nectar sugar mass was obtained for 46 species and pollen volume for 33 species of the 96 encountered (see Appendix Tables A3–A4 for full species lists). Pearson correlations between nectar sugar mass or pollen volume and the length, width, height, surface area, and volume measurements of each species (all variables log-transformed to approximate normal distributions) were used to determine which floral dimension could best estimate the amount of FR.
For all bee genera other than Peponapis , the abundance of FR in the landscape surrounding each sampling location was calculated by determining the mean FR value per flower of each species and multiplying this value by the count of each flower in a quadrat. In the genusPeponapis , pollen is collected exclusively from Cucurbitaspp. (Hurd et al. 1974). Therefore, in models of Peponapisvisits, the abundance of FR in the landscape surrounding each sampling location was calculated from the mean FR value per squash (Cucurbita spp.) or cucumber (Cucumis sativus ) flower, since 99.7% of all visits observed were to squash and 0.3% were to cucumber. While the other bee genera we considered include some oligolectic (pollen-specialist) species in our study area, they are not uniformly specialized on a single plant taxon, so all rewarding plant taxa were included in calculations of FR for these bees. The mean abundance of FR per 1 m2 was then calculated across quadrats for each transect, and the median of the transect-level values was calculated for each land type during each time period. This number was then multiplied by the total area of each land type within 250 m, 500 m, and 750 m around a sampling location to obtain an estimate of the total FR at a given spatial scale during a given time period.
Statistical analysis
All statistical analyses were performed in R version 3.6.1 (R Core Team 2019). Analyses of bee visitation rate per transect were conducted on bee genera that were present in at least four out of 27 sites during a given time period. Spatial autocorrelation among sites in the number of visits by each genus in a given time period was assessed using Moran’s I (Paradis et al. 2004). Visits from Apis mellifera were found to be spatially autocorrelated across all time periods (p = 0.02, Moran’s I |observed – expected| = 0.14), likely due to the presence of hives on certain farms, so were not analyzed.
Generalized linear mixed models were run with a zero-inflated negative binomial distribution and log link function, using the glmmTMB package (Brooks et al. 2017). Models treated the total number of bee visits observed within a transect as the response variable and were run separately for each genus; all models included a log offset to account for varying lengths of observation time based on transect sizes, and the crossed random effects of time period and site. For genera that were present in at least four sites in only a subset of the time periods, only models including data for those time periods were run. The model for the null hypothesis was of the form