The experiment was conducted at three sites (Freiburg, 48° 00′ 50.1″ N 7° 50′ 10.4″ E; Ebnet, 47° 58′ 59.9″ N 7° 55′ 21.9″ E; Kirchzarten, 47° 57′ 44.3″ N 7° 56′ 36.7″ E) close to the city of Freiburg, Germany, in July and August 2018. All sites had a high density of Cichorieae (either Crepis capillaris or Leontodon hispidus) in flower at this time. These herbs belong to the Asteraceae family and have yellow flowers arranged in composite flower heads (capitula) that were found to close in response to pollination but typically around noon (Fründ et al. 2011). Given their high abundance and their open flower morphology (short corollas, easily accessible pollen), Cichorieae are visited by a wide range of pollinator species (Fründ et al. 2011; Maia et al. 2019). The Freiburg site was a dry grassland on public green space not mown before July or August. This site was dominated by Crepis capillaris as our focal species, with Thymus vulgaris, Daucus carota, Pimpinella saxifraga, and Achillea millefolium as other abundant species in flower (see Supplementary Table 2 for full species list). The two other sites (Ebnet and Kirchzarten) were hay meadows mown once a year in June or July and afterwards dominated by Leontodon hispidus as our focal species. Other abundant plant species in flower at these sites were Centaurea jacea, Knautia arvensis, and Leucanthemum vulgare. At two sites, there were very low numbers of one or two other species of the Cichorieae (Hypochoeris radicata at Freiburg and Ebnet and Leontodon autumnalis at Freiburg, Supplementary Table 2) that showed similar flower closure and visitation patterns as the two focal species.
At each site we carried out two experimental runs, which consisted of a reference day without application of a treatment and a consecutive treatment day during which we manipulated afternoon availability of Cichorieae flowers by excluding pollinators in the morning (Supplementary Figs. 1, 2, 3). We established ten 2 m × 1 m plots at each site, and assured that each plot contained a similar and relatively high number of flower heads of the focal Cichorieae species. Sites were separated by a minimum distance of 3 km, while plots within each site were arranged as five pairs of a control and a treatment plot (1–5 m distance between paired plots and 5–30 m distance between pairs).
On reference days, pollinator visits to all plant species were recorded on all plots during four sampling rounds (two in the morning, two in the afternoon). After all sampling was finished for the reference day, half of the plots were covered by pollinator-exclusion cages constructed from insect protection nets (RANTAI Hobbynetze, TYP S48), which had a height of approximately 1 m (Supplementary Fig. 2). The pollinator-exclusion treatment was applied until noon of the subsequent day (treatment days). On this second day, the five control plots without pollinator-exclusion cages were again sampled two times in the morning and two times in the afternoon, while the treatment plots were sampled two times in the afternoon after the pollinator-exclusion cages had been removed. The exact time of cage removal was adjusted based on the time when most Cichorieae flower heads in the control plots had already closed and thus was accomplished between 11:00 and 12:30. This 2-day sampling procedure (a “run”) was carried out twice at each site. Experimental runs at the same site were at least 5 days apart and plots were assigned to a different treatment in both runs to avoid any influence of plot identity.
During each of the four sampling rounds, two observers recorded pollinator visits to flowers within plots, resulting in two separate data sets. The first observer recorded each pollinator visiting a flower for 6 min per plot. These 6 min were pure observation time excluding the time necessary to catch and label insects. Individuals were identified in the field, or caught with a sweep net, killed with ethyl acetate and stored in dry tubes for later identification. All captured insect specimens were identified to species or morphospecies level with the help of taxonomists (see Acknowledgements). This procedure revealed the identities of pollinator species visiting the different plant species, but may have underestimated visitation rates as the individuals caught were likely to have visited more than one flower per plot. Thus, during each sampling round each plot was additionally observed by a second observer for 5 min to assess the total number of flower visits, including multiple visits by the same pollinator individual (here pollinators were not caught and, in most cases, not identified to species-level). For individuals with a very high number of visits we estimated the number of visits in the field or assigned a plausible number in retrospect and, if possible, based on observations of the same species (Supplementary Table 1). Although only pollinators that touched the reproductive organs of flowers were recorded, we highlight that observed pollinator visits may not equally contribute to the reproduction of plants (King et al. 2013). During each sampling round we also counted the numbers of open Cichorieae flower heads in each plot to assess whether the treatment had been effective. Flower heads were considered to be open if their diameter reached at least 50% of the diameter of a fully opened flower head.
We tested the effects of treatment, time of day, and sampling day on the number of Cichorieae flower heads as well as on the number of pollinator visits (counted by the second observer) to both Cichorieae and other plants with three generalized linear mixed models. As we did not have morning data for treatment plots on treatment days, we used three different subsets of the full data set for these models: (i) to show the effect of our treatment, we used only afternoon data and tested the effects of treatment, day, and the interaction of both on flower head abundance and number of visits. (ii) To show the effect of time of day, we used only control data and tested the effects of time of the day, sampling day, and the interaction of both on the three response variables. (iii) To show that there was no systematic plot bias between treatment and control plots, we used only reference day data and tested the effects of time of the day, treatment, and the interaction of both on the three response variables. In all cases, data were modeled on a negative binomial distribution using the “nbinom2” family in the glmmTMB package ver. 184.108.40.206 (Brooks et al. 2017) in R version 4.0.2 (R Core Team 2020).
Construction of full-day networks
For each sampling day (6 reference days and 6 treatment days), we constructed two full-day plant–pollinator interaction networks, one for the control and one for the treatment. For constructing these networks, we considered only interaction data for which the species (or morphospecies) was identified (= data sampled by the first observer). Each network was based on interaction data sampled during 120 min pure observation time (4 rounds × 5 plots × 6 min). Control networks were based on all interactions observed in the five control plots across the four sampling rounds per day. Treatment networks were based on the interactions observed in control plots during the morning (on treatment days treatment plots were covered by nets during the morning), combined with all interactions observed in treatment plots during the afternoon. For treatment days, treatment networks thus represent an artificial scenario where Cichorieae flowers are available throughout the day. The “treatment” networks for reference days were constructed in the same way as for treatment days (combining morning data from control plots and afternoon data from treatment plots) to allow comparisons.
Timing flexibility of pollinators
To test whether flexibility in timing depends on the degree of specialization of pollinators, we classified pollinator species with > 5 observations as Cichorieae specialists if those species visited Cichorieae in > 90% of cases. We performed a paired t-test to test whether the contribution of these specialists to the observed visits on Cichorieae differs between morning and afternoon in treatment networks.
Temporal network dynamics
To describe the diel dynamics of plant–pollinator interactions underlying the full-day networks, we constructed morning and afternoon sub-networks for each day and calculated four measures of dissimilarity between the two sub-networks: plant turnover, pollinator turnover, link turnover, and link rewiring. These four measures of temporal dynamics were calculated (using the bipartite package ver. 2.14 in R: Dormann et al. 2009; Fründ 2021) as quantitative Jaccard dissimilarities between sub-networks, which were standardized to proportions beforehand. Plant and pollinator turnover reflect temporal dissimilarities between communities derived from marginal totals. Link turnover (betaWN) is the total dissimilarity between the two sets of interactions. Link rewiring refers to switching of interaction partners over time among temporally co‐occurring species (Poisot et al. 2012) and thus only considers the dissimilarity of interactions among species shared between morning and afternoon sub-networks. We focused on differences between morning and afternoon as flower heads of Cichorieae usually close around noon.
Pollinator sharing and switching
To explore the influence of our treatment on pollinator sharing and switching among Cichorieae and other plants, we assigned the visits observed in the afternoon to three categories: (a) pollinators that had visited Cichorieae in the morning; (b) pollinators that had visited other plants in the morning, and (c) pollinators that were not observed in the morning (due to pollinator timing or undersampling). For each pollinator species, we first assessed the relative frequency of Cichorieae and other plant species, respectively, among its morning visits to define the degree (proportion) to which it belonged to category (a) and (b), respectively. Second, we multiplied the afternoon visits for each pollinator species by the proportion it belonged to (a) and (b), respectively, to assign afternoon visits to these categories. If a pollinator species was only observed in the afternoon, we assigned all its visits to category (c), i.e., pollinators entering the network in the afternoon. Afternoon visits assigned to each of the three categories were then summed across pollinator species separately for Cichorieae and other plants.
To compare the structure of full-day networks between control and treatment, we calculated plant generality, pollinator generality (Bersier et al. 2002), modularity Q (Beckett 2016), and network specialization H2’ (Blüthgen et al. 2006). These network indices were used to understand how temporal availability of Cichorieae flowers affects different aspects of specialization in networks. Weighted quantitative generality can be calculated for trophic levels (plants and pollinators) and reflects the mean effective number of interaction partners of the species in the focal group weighted by their marginal totals (Bersier et al. 2002). To understand which part of the network drives the overall change, plant generality was also calculated separately for Cichorieae only (Cichorieae generality) and for all plants excluding Cichorieae (non-Cichorieae generality). Modularity Q describes the degree of compartmentalization in a network and ranges from 0 (the network does not have more links within modules than expected by chance) to a maximum value of 1 (all links are within modules) (Olesen et al. 2007; Beckett 2016). Network specialization H2’ describes the degree of specialization among plants and pollinators within the network (Blüthgen et al. 2006) and ranges between 0 (extreme generalization) and 1 (extreme specialization). We expected that extended flower availability of Cichorieae in our treatment will accumulate more pollinators on Cichorieae, thereby increasing plant generality, in particular Cichorieae generality. We also expected mostly generalist pollinators to respond to extended availability of Cichorieae, thereby increasing pollinator generality. In addition to this overall expectation of more generalized networks in the treatment, we expected that extended availability of Cichorieae will reduce temporal niche differentiation and increase connectivity between network compartments, and thus reduce modularity and H2’.
Null model comparisons and permutation tests
To assess the significance (“non-randomness”) of the observed values of our response variables describing network dynamics and structure, we compared the mean of observed values across the six experimental runs against a corresponding null model. For each null model simulation, we constructed one random treatment and one random control network for each of the six runs, calculated the respective response variable and averaged over these six values to generate one null model mean. We repeated this procedure 1000 times to provide 95% confidence intervals for the mean of the null model values. This procedure was the same for all null model comparisons although the null models differed slightly between response variables.
The null model for the measures of temporal dynamics (“timing null model”) randomly shuffled interactions (visits) among morning and afternoon sub-networks but kept constant the frequency of interactions per sub-network as well as the frequency of each unique link per day. It thus tested whether the observed values were simply due to abundance differences between morning and afternoon or due to real differences in the timing of interactions.
To test for the significance of network structure and elucidate the influence of species timing on network structure, we compared the observed network indices against two null models. These null models randomized interactions within networks but accounted for (i) the abundance per plant and pollinator species (“structure null model”) and (ii) the abundance per plant and pollinator species per time of the day (“structure-per-time null model”). The structure null model was realized by applying the Patefield algorithm (function r2table called by function nullmodel in bipartite) to the full-day networks. The structure-per-time null model applied the Patefield algorithm separately to morning and afternoon sub-networks and afterwards summed both sub-networks to generate a randomized full-day network. Using the Patefield algorithm, which keeps row and column sums constant, allowed us to account for species abundances. The only difference between these two null models is whether or not network structure is constrained by species timing.
To test whether response variables differed between treatment and control networks, we performed permutation tests that compared the observed difference between treatment and control against the difference between two sets of random combinations of study plots irrespective of their true treatment. We first generated all 252 possible combinations of five plots from our ten study plots, and subsequently drew one of these 252 plot combinations for each sampling day. All plots within a drawn combination were used to construct randomized treatment networks, while the remaining plots were used to construct randomized control networks. Note that we again used the morning data of the five true control plots for both randomized treatment and randomized control networks. We finally calculated the difference between these randomized control and treatment networks for each sampling day and, separately for reference and treatment days, took the mean across all six runs. We repeated this procedure 5000 times and calculated the 95% confidence interval for the mean difference between networks of randomized plot combinations. An observed mean difference outside the confidence interval indicates a significant effect of the treatment.