Enhancing sustainable pollination on urban farms using native plant conservation strips and outreach

The overall purpose of this project is to 1) better understand how native plant conservation strips influence the composition of bee communities on urban farms and their associated ecosystem services and 2) determine the best approach to increase awareness and implementation of pollinator management on urban farms and, 3) create outreach materials and conduct field days and workshop series with our partnering organizations to increase overall awareness of native bees and their pollination services on urban farms.

Objective 1: Improve in-farm habitat for pollinators by utilizing native plant conservation strips on urban farms.

We established 46.5 m² native plant provisioning habitats on nine urban farms and gardens in Detroit and Lansing, Michigan, U.S.A. in May and June of 2024. Habitats were composed of 240 individual plant plugs belonging to 15 perennial species that were sourced from a local plant nursery (Wildtype Native Plants, Mason, MI). Plant species were selected for their pollinator resource value and ability to thrive well in urban environments based on growth pattern, heat tolerance, and overall resiliency. The shape and dimensions of the habitat and species arrangement were left to farm or garden manager’s discretion, resulting in random habitat design and species distributions which were not accounted for in this study. In 2024 and 2025, we maintained species composition by removing plants from habitats that were not part of the original 15 species selected.

Prior to observing floral visitation, we recorded floral abundance for each species. For plant species with larger, spatially separated The floral unit of all other species was the capitulum, spike, or secondary umbel (Baldock et al. 2015; Lowenstein et al. 2019). We then observed all floral visitation interactions between bees and native plants in the provisioning habitats between 9:00 and 14:00. We recorded temperature, percent cloud cover, and windspeed during each observation period and did not survey when there was precipitation, high windspeeds, or when temperatures were below 10℃ (Bloom et al. 2021; Roedel and Szendrei 2025). Visitation surveys occurred once monthly at each site throughout the squash growing months of June, July, and August in 2025. To sample pollinators at different times of day within our sampling window, we randomized site visit order each month. During the 30-minute observation periods, the provisioning habitat was split in half with one observer recording all floral interactions in each half concurrently. Therefore, the entire habitat was observed for a combined one hour each month, resulting in a total of 27 sampling hours. As our study focused on foraging, not pollination, contact to floral reproductive parts was not required for an observed visit, although this incidentally often occurred.
           

Non-invasive sampling methods were used for our bee identification to avoid over collecting on small urban farms. Larger and easily recognized species were identified on the wing. When species identification was not possible, specimens were collected in individual vials that were labeled with a unique identification and stored in a cooler. Specimen were removed after 20-30 minutes so that photographs could be taken with an OM System TG-7 high-resolution digital camera. We later used photos and supporting resources, including Discover Life bee species guide and world checklist (Hymenoptera: Apoidea: Anthophila; Ascher and Pickering, 2020) and “Bees an Identification and Native Plant Forage Guide” (Holm, 2017), to identify specimens to the lowest taxonomic level possible. Identifications using images were extremely conservative and should largely be considered verified “on the wing” as species characteristics are rarely visible in photos. To further guide identification, we used a reference collection of locally sampled bees available in the Vegetable Entomology Lab at Michigan State University (East Lansing, MI). The collection included bees identified by an expert (Jason Gibbs, University of Manitoba, CA) for a similar study (Roedel and Szendrei 2025). When we were unable to identify a specimen on the wing or collect them, they were recorded as ‘wild spp.’ in surveys.

After bee identification, we classified specimen as squash pollinators or non-squash pollinators. Squash pollinator status was determined by previous literature observing bee-squash flower visitation. As many studies report only genus level identifications of pollinating visitors and many of our identifications are not down to species, classification of bees as squash pollinators is often by genera. In the case of Bombus, this resulted in the classification of the cleptoparisitic species B. citrinus as a squash pollinator. Although this species efficiency as a pollinator is likely low, the potential that they are among unidentified Bombus from previous studies can’t be ruled out. Other taxonomic groups classified as squash pollinators despite unresolved efficiencies include Ceratina and Triepeolus remigatus. We did not use reports of very few individuals (three or less) foraging on squash flowers as support of their role as pollinators if this was not reiterated in multiple studies.

We observed squash flowers after completing visual surveys in the provisioning habitat whenever squash was blooming, which varied by farm. Squash plantings were a minimum of 46.5 m² and were planted within 305 m of the native provisioning habitat. On some farms, squash plantings were unsuccessful or planted too late in the season, resulting in flowering outside of the June-August sampling window. This limited squash observations to 5 farms, which were surveyed 1-2 times. During squash surveys, we visually observed up to 30 random open flowers (15 female, 15 male) for 1 minute each. For each 1-min trial, we counted the number of visits that resulted in a bee contacting the squash anthers or stigma. We identified all bee visitors following the same methodology previously described. Ultimately, we completed 127 1-min observation trials, or 127 sampling minutes.

All analyses were conducted using R Statistical Software (v4.; R Core Team 2024). Species accumulation estimates of bee species and morphotypes (unresolved species) were conducted using the ‘vegan’ package (Oksanen et al. 2020). We used the ‘DHARMa’ package (Hartig 2024) to investigate residual homoscedasticity and uniformity of all models, and the ‘performance’ package (Lüdecke et al. 2021) to check for model collinearity and singularity. For all models, we ran an analysis of variance modeling with the ‘Anova’ function from the ‘car’ package (Fox and Weisberg 2019). When significant differences were detected (p < 0.05), we investigated pairwise differences with a means comparison from the ‘emmeans’ package (Lenth 2025). Observations of floral visitation events were used to construct bipartite networks weighted by interaction frequency with the ‘bipartite’ package (Dormann and Fründ 2008). Network data were pooled together across the sites to create monthly networks for June, July, and August and all months together were used to create an overall network. We analyzed the monthly networks at the community (network) level using the networklevel function. Network-level analyses were performed using species and morphotypes, but for visualizations, we binned morphotypes by genus except in cases where they represented only two difficult to distinguish species. Although analyzing at different taxonomic levels is not ideal, it is accepted for field identification (Kremen et al. 2002; Moretti et al. 2009; Lowenstein et al. 2019). 

 Because network size heavily impacts all network values except for nestedness, we used a randomized null model and z-score approach to make comparisons across networks (Dormann and Fründ 2026; Dalsgaard et al. 2013; Gilarranz et al. 2015; James et al. 2013; Nielsen and Bascompte 2007; Sebastián-González et al. 2015; Takemoto and Kajihara 2016; Takemoto et al. 2014; Trøjelsgaard and Olesen 2013; Welti and Joern 2015). Additionally, we used null models to ensure that observed network indices were statistically different from random (Dormann et al. 2009; Dormann and Fründ 2026). To do this, we created 5000 weighted `r2dtable` null models (Patefield, 1981) for each network. Random null models were also used to compare site-level mean differences for each metric in the overall network with and without the observations of two extremely prolific Bombus species, B. bimaculatus and B. impatiens (henceforth referred to as BbBi), which accounted for ~47% of all interactions (Dormann and Fründ 2026).

Our overall network was also analyzed at the node (species) level using the specieslevel function. Node-level metrics were calculated using only fully resolved species (excluding morphotypes). For bees, we evaluated overall pollination service index (PSI), proportional generality, and two different measures of weighted centrality, betweenness and closeness. Proportional generality is the quantitative measure of normalized degree, revealing the number of interaction partners per species (Dormann and Fründ 2026) and PSI is the relative importance of a pollinator to community functionality based on species proportion and interaction (Dormann et al. 2008). For this study, PSI is theoretical as we were investigating foraging, not pollination, but it helps measure activity in the conservation strips and has strong translation to actual pollination services. Both centrality measures are commonly used to identify important species within a network (Russo et al. 2013; Jordán et al. 2006); betweenness ranks species as community connectors, closeness measures species proximity to others in the network (Freeman 1979; Martín González et al. 2010). Betweenness values greater than zero indicate a species is important for network cohesivity and connection (Martín González et al. 2010; Newman 2004), while a higher closeness value suggests a species readily impacts others within the community (Martín González et al. 2010). We then compared the mean node level metric values of squash pollinators and non-squash pollinators using generalized linear models with gamma distributions for all metrics except for betweenness, which was modelled using negative binomial distribution with the ‘glmmTMB’ package (Brooks et al. 2017).
 

To identify priority plants, we calculated mean proportional generality and both measures of centrality. These values also came from the overall network, as monthly networks were too small to calculate node level metrics after separating by site. To make comparisons across species, we used generalized linear mixed models and included site as a random effect. Proportional generality and closeness were modelled using gamma distributions with the ‘lme4’ package (Bates et al. 2015) but betweenness was modelled using negative binomial distribution. We used simple chi-square test analyses comparing total visitation frequency and visitor species richness across pollinator status to determine if squash or non-squash pollinators were supported more by each plant species. This was done excluding the visits of BbBi for a more balanced comparison. Highly attractive plants, those visited more frequently than expected given their floral abundance, were identified as those outside of the 95% confidence interval for the correlation between floral abundance and total visitation (Russo et al. 2013; Lowenstein et al. 2019). This correlation was modelled using zero truncated negative binomial distributions under four scenarios: all bees, all bees except for BbBi, all squash pollinating bees except for BbBi, and all non-squash pollinating bees. For the first two models, floral abundance, month, and plant species were fixed effects with farm as a random effect, but due to singularity in the remaining models, the random effect was removed (all squash pollinators -BbBi) or replaced with flower (non-squash pollinators).

We investigated the impacts of month and plant species as well as all other variables (bee origin and pollinator status, sampling time period, temperature, wind speed, cloud coverage) on total floral visitation rates. Here, we used the ‘dredge’ function from the package “MuMIn” (Barton, 2025) to select the top model. We repeated this process without BbBi observations to assess if these two Bombus species were shaping floral visitation patterns (Plant et al, 2025). Finally, we compared visitation in provisioning habitats to observed squash flower visits. Due to variation in sampling intensity between these two flower types (native and squash) and limited squash observations, differences are only visualized as proportion of total visits by observed squash pollinators to squash flowers and to all native flowers.

Objective 2: Understand urban farmer perceptions and decision-making tools used for making pollinator management decisions.

The primary objective of our survey was to identify and evaluate factors that informed pollinator management decisions initiated by urban growers. This included attempting to elucidate the relationship between the use of pollinator management methods and grower dependence on insect pollinators based on the crops they grow. Additionally, we aimed to investigate how decisions are impacted by grower demographics, perception of pollinators and conservation, motivation for urban farming and pollinator management, resource use and accessibility, and their source of educational information. A secondary objective was to evaluate pollinator educational resources used by growers to increase our understanding of their learning preferences and improve pedagogy.
           

To support our objectives, we developed a survey with 45 questions that were a mix of three-point Likert-style, multiple choice, true or false, and open response. The survey was made available in English and Spanish and was designed to be completed in approximately 10 minutes. No questions were required; therefore, participants could refrain from answering at their discretion. Questions were divided into four parts: farm/garden description, farm management, pollinator knowledge and resource accessibility, and participant demographics. Responses throughout these sections were used to determine crops grown and pollinator management strategies implemented by growers. They also informed our predictor variables, which fell within five major categories: demographics, perception, motivation, resource use and accessibility, and source of information.

Demographics: Studies have shown that grower demographics can impact the implementation of pollinator management strategies on farms (Hevia et al., 2021). Since grower demographics can influence the types of crops grown (Philpott et al., 2020), we inquired about the following grower demographics: age, gender identity, highest level of education, years of agricultural experience, and the percentage of overall income from farming. Age, gender identity, highest level of education, and percentage of overall income from farm were formatted as multiple-choices, while years of agricultural experience was open-response. We also included a multiple-choice for respondents to indicate whether their farm location was urban, suburban, or rural.

Perception: To understand how grower perception, or the knowledge and beliefs of pollinators and their management, influences the use of pollinator management strategies, we inquired about their limitations for implementation using a three-point Likert-scale (Table S1). We also included true or false questions to inquire about each respondents’ pollinator knowledge. These true or false questions tested grower perception of common beliefs regarding pollinators and their management.

Motivation: Based on previously found motivations for engagement in urban agriculture (McDougall et al. 2019), we asked respondents to indicate motivations for urban farming and gardening on a three-point Likert-scale consisting of the following statements: access to fresh and healthy foods, access to culturally appropriate foods, supplementing the food supply for self and household, environmental sustainability, the ability to provide food to the community, enhance social bonding within the community, and additional income.

Resource use and accessibility: Access to grower information and resources, or resource use and accessibility, can either limit or increase farmers’ capacity for pollinator management. Additionally, numerous urban farmers have experienced limitations when attempting to implement pollinator management strategies. To further understand this, we inquired about the following limitations using a three-point Likert-scale: financial investment, manual labor, city regulations, and access to plant resources.

Source of educational information: In this part of the survey, we asked respondents to select any of their preferred educational, pollinator sources from a list of online, hard copy, and in-person materials.

Pollinator management strategies: We evaluated the implementation (presence/absence) of eight common farm practices that positively influence pollinator abundance and diversity. These practices were selected because they provide resources for pollinators while requiring minimal financial or material investment. To avoid bias, pollinator management strategies were referred to as farm management methods in the survey. For each pollinator management strategy, participants were asked about their primary reason for implementation, with the following options: soil health, pest control, pollinator management, weed control, environmental impact, and "I don't use this method".

Crop richness and pollinator dependency: An estimate of farm crop richness was gathered by having respondents select the crops they grow out of a list of 21 common fruits and vegetables. These crops ranged in their pollinator dependency, which was used to categorize them into two groups. The highly pollinator dependent group included crops that pollinators are essential to (melons, summer squashes, winter squashes) or highly important for (cucumbers, apples, blueberries, cherries, peaches, plums, pears, raspberries, and blackberries) and the minimally pollinator dependent group contained crops with modest to no dependence (strawberries, currants, gooseberries, eggplants, broad-beans, sunflowers, tomatoes, peppers, and snap peas; Klein et al., 2007).
 

The proportion of high-pollinator-dependent crops grown was used to estimate the level of importance pollinators had for growers. We used this to classify growers into two groups: high and low pollinator importance. In the high importance group, 50% or more of the crops grown were highly dependent on pollinators (n=26). In the low importance group, less than 50% of crops grown were highly dependent on pollinators (n=46).

Data collection
Our survey was disseminated both online through Qualtrics (Qualtrics, Provo, UT) and in-person at various events from August to October, 2024. The survey was open to anyone in the United States (USA), but mostly circulated throughout Michigan, USA. Online distribution included social media posts, direct emails, various grower newsletters, university extension communications, Master Gardener Program leaders, and a post-webinar announcement. Hard copy versions and QR codes to the digital survey were made available at a ‘Pollinators in urban agriculture’ field day, the National Urban Agriculture Conference, a local farmers market, and at numerous events with the Greater Lansing Food Bank’s Garden Project. Growers were incentivized to complete the survey by providing entry into a post-survey raffle to receive a $100 gift card. A QR code and link to the raffle appeared after growers submitted their responses to the digital survey and it was also located on the last page of the paper survey. The winner was drawn one day after the survey was closed.

Data analysis
We received 126 initial survey responses which were analyzed using R studio version 2024.12.1+563 (Posit Team, 2025). Surveys with response time less than 4 minutes were removed from analysis (n=38). Surveys from non-urban growers (n=10), surveys without any responses, or that were duplicate entries were also removed from the analysis (n=4). After excluding these surveys, 74 were used in our analyses.
           

We used generalized linear models to investigate the correlation between the implementation of pollinator management strategies and grower pollinator dependency level. We also created generalized linear models to determine how implementation of pollinator management strategies and crop richness were influenced by the following predictor variable categories: grower demographics, perceptions, motivation, resource use and accessibility, and source of educational information. This was done by using a model selection process. We created separate global models for each of the five predictor variable categories for both pollinator management strategy implementation and crop richness as response variables. For grower motivation, ‘reasons for using various farm management methods’ was included only for crop richness as a response variable to avoid collinearity in the pollinator management implementation model. With crop richness models, we used three different response variables: total overall crop richness, total number of highly pollinator-dependent crops, and total number of low pollinator-dependent crops. We then used the ‘dredge’ function in the “MuMIn” package (Barton, 2025) to select the top model for each predictor variable category. Based on these results, we included the most important variables in our statistical models.

Using our model-selected variables, we created unique generalized linear models for each category for both response variables with the ‘glm’ function. Shapiro-Wilk tests were used to check whether the data were normally distributed. Gaussian error distributions were used for continuous response variables (crop richness) and binomial error distributions for dichotomous response variables (use of pollinator management strategies). We ran an analysis of variance modeling using the ‘Anova’ function with the “car” package (Fox and Weisberg, 2019). If significant differences were found (p < 0.05), we followed with a means comparison (package = ‘emmeans’; Lenth, 2025) to determine pairwise differences between predictor variables.

We also ran an analysis of variance to compare the overall effects of our five predictor variable categories on the implementation of pollinator management strategies. We created model averages from the five global models: demographic (N=16), motivation (N=64), resource use and accessibility (N=128), perception (N=1,024), and source of educational information (N=16,384). We then extracted the model averaged sum of weights for all variables for each management strategy. Next, we added the model averaged sum of weights for all predictor variables within a category together for each management method. We then frequency corrected these coefficients by dividing them by the number of variables in the predictor category they were from. We used the ‘Anova’ function to determine differences in the frequency-corrected coefficients across categories.