The seasonal dynamics of a High Arctic plant - visitor network: temporal observations and responses to delayed snow melt

Plant - visitor food webs provide important insights into species interactions, and more information about their seasonal dynamics is vital to understanding the resilience of species to external pressures. Studies of Arctic networks can also improve our understanding of species responses to the pressures of climate change. This study provides the first description of a plant – insect visitor network in Svalbard, a High Arctic archipelago already experiencing the consequences of climate change. A subset of the network was collected from experimental plots where the snow melt date was delayed with snow fences. The deep snow plots delayed flowering and we expected this to disrupt plant-visitor interactions compared to ambient snow conditions. However, the composition of flowers and insect visitors were similar between regimes, and the network tracked patterns of overall flowering phenology. Nevertheless, the deep snow significantly reduced the average overlap between flower availability and insect activity, reducing the probability of an interaction. We suggest that at a landscape scale, Arctic pollinators will benefit from patchy changes to snow melt that maintain heterogeneity in the timing of flowering but changes that increase homogeneity in snowmelt across the landscape may negatively impact some species.

5 2011). However, as network generalisation increases with latitude and is lower for islands (Olesen and Jordano, 2002), Svalbard provides an interesting trade-off location as an isolated, high latitude archipelago. We hypothesised (H1) that Svalbard would have a species-poor network, commensurate with the archipelago's low insect biodiversity. We also hypothesised (H2) to find similar temporal dynamics to other High Arctic networks, in that the network would exhibit a gradual build-up and sudden collapse. The latter characteristic occurs because the flowering season comes to an abrupt halt due to either the limited length of flowering for many species that have periodic phenology (Semenchuk et al., 2016), or the onset of cold autumn weather and significant reduction in day length (e.g., Pradal et al. 2009).
The current warming climate in the Arctic and other snow-covered areas plays a crucial role in plant-pollinator interactions, with phenological changes likely to affect the coupling between some plants and pollinators (Rasmussen et al., 2013, Kudo and Cooper, 2019. This prediction has not been explored in detail in the high Arctic as studies rarely cover the timespan required to establish connections between network dynamics and abiotic factors or experimentally explore the impacts of climate drivers (Burkle and Alarcón, 2011). Here, we adopt the latter approach using an experiment that artificially delays snow melt at the plot level leading to delayed flowering (Semenchuk et al., 2016). We previously showed that peak insect visits tended to move from flowers in ambient conditions to the later emerging flowers in deep snow plots, apparently prolonging the flowering season across the local landscape at the site level (Gillespie et al., 2016). However, here we focus on the static and dynamic network properties and the impact of snow melt delay on the likelihood of flowers and insects interacting. To date, only one other study has explored the impact of experimentally altered snow melt on plant-visitor networks, and found that advanced snow melt increased the degree of specialisation due to shifts in flower phenology (Hoiss et al.,6 2015). Advanced flowering may have reduced the availability of resources for insect-visitors by creating temporal mismatches (Memmott et al., 2007). As our delayed snow melt regime also only manipulates one half of the interaction, we similarly hypothesised (H3) that delays to flowering phenology would result in reduced floral resource availability for insects. The findings from this experiment provides important insights into the response and resilience of interacting species in snow covered landscapes to changes in the phenology of resource availability.

Study site and experimental design
The study was carried out in 2015 in Adventdalen (78ᵒ10'N, 16ᵒ06'E) on Svalbard, in an area approx. 2.5 km x 1.5 km, with mean July temperature of 5.9 °C and 190 mm annual precipitation, most of which falls as snow and is subsequently redistributed by strong winds.
In order to build separate networks to compare the effects of snowmelt timing we used fences to experimentally enhance snow depth by 1 m, resulting in c. 2 weeks delay in melt ('Deep' 7 snow regime: snowmelt 11 June 2015), paired with adjacent ambient areas (approx. 30 cm snow depth, snowmelt 26 May 2015 'Ambient' regime). The site, climate, and experimental set-up has been described in detail elsewhere , Morgner et al., 2010, Semenchuk et al., 2013, and more details are available in the Supplementary material. For this part of the study we used five fence -ambient area pairs in the mesic meadow (Supplementary Table S1). At each fence location permanent vegetation plots (0.75 x 0.75 m) were used (two in each of the Deep and Ambient regimes, totaling 20 plots), to record flowering phenology and insect visitation (see below).

Insect sampling
Flower-visiting insect sampling was undertaken in the summer of 2015 and has mostly been described elsewhere (Gillespie et al., 2016). We give a brief overview here together with details not included in that paper. A full description can be found in the Supplementary material. Between 17 June and 6 September 2015, there were 22 suitable days for sampling flower visitors. Each day, flowers were observed in the 20 permanent plots for 10-minute intervals and whenever an insect landed on a flower it was caught and the plant species was noted. The plots of the study site varied in flower density and species composition, with some consisting of large flowers (e.g. D. octopetala) and others containing only small flowers (e.g. B. vivipara). In many cases, the plots often accommodated a low number of flowering species and few flowers, so insect sampling was supplemented with additional observations outside the marked plots (but within the study area) in order to build-up an overall network of plantinsect interactions. On each day, the study area, including the heath and other Deep areas, was searched for flowering plants of all species not included in the plots. Where possible, two individuals of each additional (non-plot) flowering species were randomly selected and observed for 10 minutes. This sampling methodology is likely to have led to oversampling of 8 the more common species found in plots (see Supplementary Table S2 for a list of species and the number of flowers observed on each date). However, we have retained all records in our analysis due to the low number of observations resulting from sub-sampling. Under the cool Adventdalen conditions, the insects were docile and easy to catch and, therefore, the vast majority of observed individuals were caught. On the rare occasion that individuals were missed, they were noted to a broad taxonomic level, but the addition of these data did not affect the results so they were omitted. We randomized the order of plot sampling each day to avoid possible confounding effects of sampling time. Insects were identified to Family level because good identification keys were lacking for many species. All data analyses were conducted using the R programming environment (v3.6.2, R Core Team, 2019). The data from all plant-insect observations were first pooled in order to construct a full-season static plant-visitor network for the Adventdalen study community. To increase the size of the dataset this included plant-insect visitation observations from Deep areas, which we assume represent the snow depth and melt of gulley areas in Adventdalen, as recent work has shown similarity in snowmelt timing of our Deep regime and nearby gullies (Moriana-Armendariz et al., in press). Network diagrams were produced using the bipartite package (Dormann et al., 2008). The structure of the network was compared to four other Arctic networks published previously (see Supplementary material for details). To study the seasonal build-up of the network, we separated the data into temporal windows or "time slices". To avoid arbitrarily determining the time slices and ensure that each one was populated with sufficient data (>2 days and >30 interactions), we used a method described by We compared the networks in experimental plots in two ways. First, we compared the data from Ambient against Deep to identify any differences in overall network structure and the appearance of phenological mismatches due to a delay in snow melt. This comparison was largely descriptive as plot level data were not sufficient to conduct statistical modelling.
Nevertheless, to explore whether the resulting pattern was different between the experimental regimes, we fit General Additive Models (GAM) with the mgcv package (Wood, 2017) with the quasi-poisson family to account for overdispersion in the response variables: the number of species and number of links. For both models, snow regime was a fixed effect, and Day of the Year (DOY; the day of the year of the survey as a number, where January 1 = day 1) was 10 fit with a penalized thin plate regression spline smoother (bs = "ts"). In order to test if the smoother differed between snow regimes, we treated the categorical variable as ordered, with Ambient as the reference level, and fit a "difference" smooth as a factor-smooth interaction using the by argument. In both cases, the models were checked for normality and homogeneity of variance in the residuals. After fitting the model, we used posterior inference with 10 000 simulations to calculate the maximum point on each curve and estimate the DOY (and 95% confidence intervals) when the networks reached peak number of links or species.
Finally, we examined the indirect effect of the snow regime on plant-visitor interactions via the overlap between plant and insect phenophases (H3). As a greater degree of overlap between plants and insect-visitors is more likely to lead to an interaction (Olesen et al., 2010), we tested whether delayed flowering led to reduced overlap in phenophases and subsequently a reduced likelihood of visitation. Here, the term phenophase refers to the period of flowering for plants, and the period of active flying for adult insects. In this part of the study, the phenophases were calculated using insect visitation data only. Therefore, flower onset and senescence dates were defined by the first and last dates a flower was observed in any plot, regardless of whether it was visited by an insect. Similarly, insect start and end dates were recorded as the date an insect was first or last caught.
To test H3, we used a "piecewise" path model, a method of piecing together the coefficients of several "sub-models", which can be built using a range of mixed model techniques (Lefcheck, 2016). We constructed two sub-models for the hypothesised path model. The first was a zero-inflated Generalised Linear Mixed model (GLMM) with poisson distributed errors. The length of overlapping periods between the phenophases of all plant-visitor species pairs found within the plots (sensu Olesen et al., 2011), hereafter termed "overlap", was the response variable, and snow regime was the explanatory variable. To account for differences in plant species availability and insect abundance, we included plant species and insect Family as random factors. The zero-inflated part of the sub-model included the same structure to predict excess zeros. The second sub-model was a GLMM with binomial distributed errors, using the binary link variable as response (1 = an insect family visited the plant species, 0 = absence of a link), snow regime and overlap as explanatory variables, and the same random factors as above. Both models were fit using the glmmTMB package (Brooks et al., 2017), and were assessed for normality and homogeneity of residuals and overdispersion using the DHARMa package (Hartig, 2020). Usually, after fitting valid sub-models, piecewise path models are assessed for goodness of fit using Shipley's test of directed separation (Shipley, 2000, Lefcheck, 2016, whereby "missing paths" are tested using the Fisher's C test statistic.
However, as our path model represents a saturated model, there are no missing paths to test and the resulting regression coefficients can be used to infer the significance and magnitude of paths in the model.

Static Full Season network
Over the 22 days of sampling from 17 June to 21 August 2015, 19 plant species were observed receiving insect visitors (two species, Ranunculus nivalis L. and Saxifraga nivalis L., were observed but did not receive visitors) and 11 insect families were determined with a total of 685 insect specimens collected. The "static" plant-visitor network of Adventdalen 12 using data collected over the entire field season and sampling site is depicted by a bipartite network matrix (Fig 1). The network shows that eight of the 19 plants were more generalist in that they received visits from a range (5-6) of insect families. These plant species received over 68% of the visits by insects. The most visited plant species was D. octopetala (191 visits), followed by S. polaris (157 visits) and S. crassipes (93 visits). However, it should be noted that the number of flowers observed were not equally divided among the plant species (Supplementary Table S2). It is therefore likely that the number of visits to plants within permanent plots are overestimated compared to less common plants.
Of all the insect specimens collected, 79.4% were Diptera (true flies), with the Sciaridae, Chironomidae, Empididae and Muscidae responsible for the majority of interactions (Table   1). The remaining 20.6% of interactions were by Hymenoptera (wasps), with the vast majority As with networks described elsewhere in the Arctic (Table 2), the Adventdalen network is small relative to those in temperate regions. The system size and number of interactions are most similar to the networks described at a similar latitude at Alexandra Fjord, Canada, and the island of Uummannaq, Greenland, although those networks were established over much smaller spatial scales. Connectance indicates the level of generalisation in the network, with high percentages associated with a high number of generalist species, and this is quite similar across most Arctic networks. However, those with a higher level of taxonomic resolution have much lower connectance values, so the true value for Adventdalen is likely to be lower. The 13 mean interactions per insect or plant species (linkage level) are also indicators of generalisation at each trophic level, and the estimate for plants in Adventdalen is low at 3.9, and similar to the Uummannaq study.

Dynamic temporal network properties
The plant-visitor interactions observed in each of the seven temporal windows across the entire study site are illustrated using a network matrix diagram (Fig. 2). The presence of different links in the network windows demonstrates the temporal build-up in complexity of the web up to the third temporal window in the second week of July. After this peak the plantvisitor network gradually decreased in complexity, as more species left the network by senescence (flowers) or inactivity (insects). This structural assembly and disassembly can also be depicted by plotting the number of species and links involved in the network over time (Fig. 3a), showing a hump shaped relationship to the build-up of both variables. The cumulative number of species and links over the active period (Fig. 3b) displays an almost linear accumulation of both species and links during the peak complexity phase, before slowing down indicating that the period of greatest growth in the network is during 3rd and 4th temporal windows (July).

Snow fence experiment
Eleven of the 19 plant species sampled in the static network were not present in Deep, and were sampled outside of areas manipulated by the snow fences, so observations from these plants have been excluded from the snow fence analysis. Similarly, it is difficult to compare networks built from regimes that are located so close together due to possible spatial nonindependence of observations, so these results should be reviewed with caution. tetragona with Chironomidae, and C. arctictim with Chironomidae and Empidae. It is not known if these are particularly important interactions for either species, and could also be an artefact of sampling error. There were also links missing from Ambient: C. arcticum with Sciaridae and Scathophagidae, and S. crassipes with Chironomidae.

Static Svalbard network
When viewed at a comparable taxonomic level, the Adventdalen network is similar to other Arctic networks in terms of the low number of vascular plants, and as expected has the lowest number of insect families involved. The complexity of the Adventdalen network in terms of number of interactions is most similar to that found on the Greenlandic island of Uummannaq (Lundgren and Olesen, 2005). While our family level connectance estimate was higher than at Uummannaq indicating more generalist members of the network, plant generalisation was at a similarly low level. These properties can be explained partly by the low level of biodiversity of High Arctic regions generally (Callaghan et al., 2004), and by the geographic patterns of network properties (Olesen and Jordano, 2002). While plant generalisation tends to increase with latitude and at lowland sites, specialisation tends to increase on islands due to low pollinator diversity. As insect species-poor islands at high latitudes, Svalbard, Uummannaq, and indeed Ellesmere Island appear to represent a tradeoff of these general patterns: the trend towards plant generalisation is mediated by the low pollinator diversity of Arctic island ecosystems. The Adventdalen network is also entirely made up of species from the Diptera and parasitoid Hymenoptera, unlike networks further south and elsewhere in the Arctic that have at least a few species of bee, moth and/or butterfly performing a pollination role. This is not surprising given the dominance of these Orders in the Svalbard insect fauna (Dipera: 62% of insect species, Hymenoptera: 13%; Gillespie et al., 2020), and supports previous observations of dipteran dominance of pollination at high latitudes (Elberling and Olesen, 1999, Lundgren and Olesen, 2005, Rasmussen et al., 2013, Robinson et al., 2018. Of the Diptera order, the Muscidae family is a particularly important group of pollinators in Arctic and alpine systems (Elberling and Olesen, 1999, Pont, 1993, Tiusanen et al., 2016, especially the genus Spilogona. In Svalbard, this is the only genus of muscid flies known to occur (Gillespie et al., 2020), and they are likely to be important pollinators because of their large body size and elongated proboscis providing access to closed flowers (Pont, 1993).
As with many other plant-visitor networks, the Adventdalen network displays a pattern of generalist plant and insect-visitor species and a number of less common species that are attached to the network with a low number of interactions (Bascompte and Jordano, 2007).
Generalists tend to occur in higher abundances with low population fluctuations relative to rarer species (Rasmussen et al., 2013). The most common plant species are likely to be more apparent and attractive to insect-visitors and other consumers (Gillespie et al., 2016, Cooper andWookey, 2003). We lack the data to determine whether the rarer species here are specialists, but true specialists are not common to Arctic networks (Olesen and Jordano, 2002), because high latitude communities tend to be subject to severe abiotic conditions and limiting resources and species exhibit wider niches, as also shown for high altitude networks (Hoiss et al., 2015). Studies of networks at high elevations have also shown that plants and pollinators are more generalist in terms of interacting partners than those at low elevations (Ramos-Jiliberto et al., 2010, Hoiss et al., 2015, resulting in networks with less specialization and structure. High generalization among insect visitors is, therefore, likely to be an important compensatory feature for such species poor networks as those reported here. Nevertheless, species poor pollinator networks occurring in highly seasonal systems are also subject to low stability and resilience in the face of perturbations such as sudden climatic changes (Encinas-Viso et al., 2012).

Dynamic network
The dynamic aspects of the network displayed similar characteristics to those of previous studies with strong temporal dynamics in species number and linkage level (Pradal et al., 2009). In particular, the network shows rapid growth during the middle of the season in July, coinciding with peak flower production and insect visitation at our site (Gillespie et al., 2016), and insect emergence in Arctic regions generally (Olesen et al., 2008, Robinson et al., 2018, followed by a period of stasis where few new attachments are made in a system that is undergoing senescence and low foraging levels (Olesen et al., 2008). However, there are indications that this network does not exhibit the drastic collapse at the end of summer found in Greenlandic networks (Pradal et al., 2009). The Adventdalen network shows a similar pattern to the High Arctic network from Ellesmere Island described by Robinson et al. (2018).
Both networks exhibit a slow build-up of complexity as new species enter the network, and a less dramatic "collapse" following peak community flower expansion. This may be an artefact of the taxonomic level of identification or a lack of survey days in both studies, but may also be indicative of a species poor network operating under harsh weather conditions and low discrimination by insects in interaction partnerships. This is supported by the finding that, in the Adventdalen network, the main flower-visiting insect families have active adult phenophases that are relatively well spread out over the season, suggesting a slow decrease in turnover towards the end of the season, and a maximizing of phenological coupling between available mutualists (Encinas-Viso et al., 2012). These species may be adaptable to future changing plant phenologies, provided they are able to emerge from winter hibernation to coincide with flowering.

Response of networks to snow regimes
Our experimental data from the snow fences come with a caution. Generally, networks collected from replicated plots tend to be small and potentially biased in that they do not sample all the interactions present (Jordano, 2016). Similarly, the areas affected by the snow fence are relatively small compared to the total study area (Kenkel and Podani, 1991), and the majority of insects are likely to have emerged elsewhere. Therefore, the results of this study should be viewed from the perspective of the manipulation of the flowering phenology only (i.e. only one half of the interaction).
We found little difference in the structure of the static networks between Deep and Ambient (data not shown), but the temporal build-up of species and links involved in the network differed between regimes with a mean delay of 8 to 12 days, and this difference appeared to be relatively conserved across the season. While our Deep regime did not lead to a quantifiable temporal mismatch or a disruption in the interactions between insects and plants, delayed melting did shift the peak growth of the network and shortened the flowering season for these plots. This suggests that the two networks show the same gradual build-up and decline but implies a more dramatic collapse with delayed snow melt, delayed flowering and reduced flowering duration. This supports the point made by Pradal et al. (2009) that the collapse of the Arctic plant-visitor network may be due to a sudden absence of flowering plants.
Our prediction (H3) that delayed flowering would result in reduced availability of resources was confirmed by our findings, but it did not lead to temporal phenological mismatches. The path model suggests that there is potential for phenological mismatch between flowering plants and flower-visiting insects with enhanced snow, as short phenological overlap between species pairs is a key reason for unobserved or "forbidden" links in a network (Olesen et al., 2011). In the Deep, a reduction in phenophase overlap reflects fewer resources for insect visitors with consequences for the probability of visitation, and presumably pollination. This may be a contributing factor to the reduced cover of some species in the Deep regime since the snow fence experiment began , Semenchuk et al., 2013. If the 20 vegetation in Deep represents future Arctic conditions (for example, in the event of increased winter snowfall; Saha et al. 2006), this could have important ramifications for both plants and insects. Shorter flower seasons and declining insect visitor abundance have been linked to climate change elsewhere in the Arctic (Høye et al., 2013). Furthermore, the impact of snow depth on overlap and visitation may be an underestimate because we only manipulated one trophic level. By the time the Deep plots were in flower, the availability of abundant insect visitors had become established from feeding on flowers in Ambient and the surrounding landscape. If delayed snow melt was to occur over a wider homogenous area, the insects would be likely to delay their emergence at a similar rate to the flowers. This would both shorten the flowering season and reduce the amount of time available for insects to build-up the network and generate the periods of phenophase overlap required to ensure a high probability of interaction. Conversely, our experimental design is likely to represent a scenario of patchy distribution of deep and shallow snow depths across the landscape, leading to heterogeneity in the timing of flower availability. This may have the effect of extending the flowering season, and if insect-visitor emergence is unaffected, they are likely to benefit from an increased temporal availability of resources. However, this is a simplification of likely future changes, as soil and air temperature, frost, CO 2 and vegetation change are also all likely to impact flower and insect emergence and abundance in snow-covered landscapes (Høye et al., 2021, Høye et al., 2013, Kudo and Cooper, 2019. Comparisons of dates of flowering and insect visitation between the two regimes also revealed interesting patterns, with common insects apparently spending short periods foraging from the most common plant species, suggesting that they tended to move between flowering species at their peak. The additional availability of flowers later in the season in Deep is also likely to have had an effect of prolonging the period of peak availability of flowers and density of the insect species. For example, ichnuemonid wasps and chironomids both appear to visit D. octopetala flowers in Deep after peak flowering in Ambient, even though some flowers of this species were still available in Ambient. This pattern of visitation can be explained by theories of optimal foraging (Macarthur and Pianka, 1966), whereby the relative abundance of high quality food resources determines the usefulness of lower quality food sources. As early flowering plants decrease in their density and quality they become less apparent and attractive, particularly given the emergence and peak density (and presumable quality) of other food sources Wookey, 2003, Kudo and. These changes in abundance can be important factors in interaction assembly and disassembly in networks (Caradonna et al., 2017), and plant switching in response to changing abundances has been found in other food webs (Carnicer et al., 2009).
The visitation patterns of insect families may also be indicative of the neutrality hypothesis of mutualistic networks, whereby relative species abundances predict interaction frequencies (Vázquez and Aizen, 2004), rather than species traits or phenology. If this is the case, changing flower phenology will perhaps have lower impact on network properties in such a general, species-poor network, as most of the insects are flexible in their food sources. The alternative, the biological constraints hypothesis, suggests that forbidden links determine network structure Jordano 2007, Olesen et al. 2010), where links between a plant and insect-visitor are not possible due to morphological or phenological incompatibility.
In this case, changing plant phenology would lead to an increase in forbidden links locally, as seen in our Deep regime above. Our experiment did not allow the detection of any complete mismatches, rather a reduced probability of interaction at low levels of overlap. In any case, Vazquez et al. (2009) suggest that both relative abundances and spatiotemporal distribution of plants and insects are together important contributors of network structure and resilience.

22
In summary, our data suggest that in snow covered landscapes such as High Arctic tundra, changes in the temporal and spatial snow melt conditions may benefit some flowering plants and insects by providing more opportunities to interact, and such flexibility in resource use could enhance the network's resilience to perturbations such as phenological changes (Burkle et al., 2013, Caradonna et al., 2017. However, this is likely to be dependent on the timing of insect emergence in changing conditions. For plants, a reduced probability of visitation in areas of deep snow and therefore late melt will likely impact patch occupancy, partly through limiting reproduction if snow melt change is variable spatially and temporally. If snow depth increases or snow melt is delayed more widely throughout the landscape, both trophic levels are likely to be negatively impacted by the shortening season (Høye et al., 2013). In general, the impact of climate change factors on the timing of peak densities of individual flower species, and the overlap of these peak densities, may be an important determinant of species interactions at the community level (Hegland et al., 2009, Rafferty et al., 2015. We suggest that further experimental and long-term observational research would help to confirm these findings and to assess the likelihood of these patterns predicting future effects for snow covered landscapes at lower latitudes.         The coloured boxplots for each taxon indicate the period from the first to the last interaction