1. Birds play many roles within agroecosystems including as consumers of crops and pests, carriers of pathogens, and beloved icons. Birds are also rapidly declining across North America, in part due to agricultural intensification. Thus, it is imperative to identify how to manage agroecosystems to best support birds for multi-functional outcomes (e.g., crop production and conservation). Both the average amounts of services/disservices provided and their temporal stability are important for effective farm planning.

2. Here, we conducted point-count surveys for four years across 106 locations on 27 diversified farms in Washington and Oregon, USA. We classified birds as ecosystem service or disservice providers using indices spanning supporting, regulating, provisioning, and cultural services/disservices. We then examined service/disservice index pairwise correlations and assessed the relative importance of local, farm, and landscape complexity on the average and temporal stability of avian service/disservice provider indices.

3. Generally, service provider indices (production benefitting birds, grower appreciation, and conservation scores) were positively correlated with each other. Foodborne pathogen risk, grower disapproval, and identity/iconic value indices were also positively correlated with each other. However, the crop damaging bird index generally had low correlations with other indices.

4. Farms that implemented more conservation-friendly management practices generally had higher average service provider indices, but farm management did not impact disservice provider indices, except for grower disapproval. Average disservice provider indices were lower on farms in complex landscapes.

5. Local vertical vegetation complexity tended to increase the temporal stability of service provider indices but did not affect the disservice provider indices. Greater landscape complexity was generally associated with increased temporal stability of service and disservice provider indices. Increased landscape complexity may stabilize bird communities by increasing bird community evenness, which in turn, positively predicted temporal stability of all service/disservice provider indices.

6. Policy implications: Our results suggest that farmers can effectively manage their farms to harness ecosystem services from birds through farm diversification. Disservices provided by birds, however, appear to be most negatively impacted by landscape-level complexity. Thus, greater incentives for farmers to increase seminatural cover at the landscape-scale are likely necessary to achieve multifunctional outcomes for conservation and agriculture.

2.1 Study system

Across four years (2016–2019), we surveyed bird communities on a total of 27 farms in Oregon (n = 15) and Washington (n = 12; Fig. 2; Fig. S1) states, USA. We obtained permission to conduct surveys on all farms from the farm owners and/or managers. All farms fell into the Northern Pacific Rainforest Bird Conservation Region. Farms were highly diversified, largely used organic practices (20 were certified), and grew a range of crops (no monocultures; mean = 46.5 ± 19.8 (SD) crops grown per farm) including cereals, vegetables and melons, fruits and nuts, oilseed crops, roots, spice crops, beverage crops, medicinal crops, commercial flowers, and grasses and fodder crops, among others. Livestock were integrated into farming operations for at least one year of the study on 18 of these farms in a variety of forms. Farms spanned a range of landscape contexts, from intensified agriculture to primarily seminatural (e.g., Fig. 2; Fig. S1; range: 2.19% to 95.7% seminatural).

2.2 Bird point count surveys

Bird surveys were conducted twice per farm each year between 20 May and 8 August 2016–2019 to coincide with peak produce production in the region. We moved along a south to north transect among farms (Fig. S1a) for each of the two annual survey periods. Survey one each year roughly corresponded with the nesting season along the south to north transect, while survey two roughly corresponded with the fledging and flocking periods for gregarious species. One point count location (“point”) with a 100-m radius was surveyed for every 4 ha of farmed land to maintain a consistent point density. Points were systematically stratified to capture the range of land uses present on farms (e.g., Fig. 2c-d; Smith et al. 2020b). Point count centers were at least 200 m apart to avoid double counting individuals. In total, 106 points were surveyed across the 27 farms included in this analysis (mean per farm: 3.9 ± 3.3 (SD); range = 1 – 14).

At each point, the observer recorded the number of unique individuals per species seen or heard within a 100-m radius during a 10 min period. Surveys were conducted between sunrise and 10:45 AM, only in the absence of heavy rain, by the same skilled observer (OMS) to eliminate biases due to observer differences. Additionally, points within farms were surveyed in a different order each visit to reduce detection biases due to time-of-day effects (Smith et al. 2020b). If structures interfered with visual detectability of birds, the observer moved within points to see around structures (Šálek et al. 2017). Because our study traded geographic breadth for within-season temporal replication, we were unable to account for detection probability in our analyses. Thus, detection probability was assumed to be constant during each survey period across farms. For conciseness, we use the term “abundance” in the article to refer to the number of individuals detected, but it should be noted that we may have missed individuals. Our research was conducted with the approval of Washington State University’s Institutional Animal Care and Use Committee (ACC protocol ASAF #04760). This manuscript uses point count data (which only requires passive observation of birds) and weights the point count estimates by foodborne pathogen prevalence estimates derived from mist-netting birds reported previously (Smith et al. 2020a, in press).

2.3 Ecosystem service and disservice provider classification

We calculated several metrics spanning supporting, regulating, provisioning, and cultural ecosystem services and disservices provided by birds (Fig. 1A; Table S1). We acknowledge that all service and disservice proxies have limitations, which we note for each proxy used in Table S1. We used abundance (total number of individuals detected during each point count survey) as a metric of supporting services.

To estimate the risk of foodborne pathogen delivery to crops (regulating disservice), we generated a foodborne pathogen risk index. To do so, each bird observed was weighted by its species’ estimated Campylobacter spp. prevalence and crop contacts/survey point from Smith et al. (in press). Briefly, Smith et al. (in press) estimated Campylobacter spp. prevalence and number of crop contacts/survey point for 139 bird species by examining which of 11 species traits were most predictive of each. They then used the best-supported models to predict Campylobacter spp. prevalence and number of crop contacts/survey point for understudied bird species. Our analyses used the estimated prevalence of Campylobacter spp. because it is the most common foodborne pathogen found in birds (Smith et al. 2020c, a). The crop contact score represents the estimated number of individuals of that species in crop fields per survey point. We accounted for both the estimated Campylobacter spp. prevalence and estimated crop contact rate because the probability that an individual will deposit pathogens on crops is the joint probability that it will carry the pathogen, enter crop fields, and defecate on produce (Smith et al. 2020c, a).

We calculated a per point estimate of food safety risk as [Equation 1].

[Eqn 1]: Per-point foodborne pathogen risk index = Σspecies’(estimated Campylobacter spp. prevalence * crop contacts * number of individuals detected)

To generate a proxy for regulating services (pest consumption and pollination) and provisioning disservices (full or partial consumption of crops), all bird species detected were assigned to a diet guild following the protocol outlined in Smith et al. (2020b, 2021) (Data S1). Wilman et al. (2014) assigned species to diet guilds when the diet was ≥ 50% that item, and we followed this definition to assign species to guilds based on the majority items in the diet (if ≤ 50% in any category, the species was considered omnivorous). We then assigned insectivorous, carnivorous, and nectivorous species as “production benefiters” (species potentially provide pest control or pollination services); and frugivorous, granivorous, and herbivorous species as “crop damagers” (species potentially inflict crop damage/loss through foraging on fruits, grains, seeds, or vegetation of crop plants) (Peisley et al. 2015; Smith et al. 2021). To calculate per survey point estimates of production benefit services [Equation 2] and crop damage disservices [Equation 3], we calculated the abundance of birds falling into each guild and weighted those abundances by the summed proportion of the diet in those categories from Elton Traits 1.0 (Wilman et al. 2014).

[Eqn 2]: Per-point production benefitting bird index = Σ[(number of individuals from insectivorous, carnivorous, or nectivorous species)*(Total percent of the species’ diet composed of invertebrates, endothermic vertebrates, carrion, plus nectar)]

[Eqn 3]: Per-point crop damaging bird index = Σ[(number of individuals from granivorous, herbivorous, and frugivorous species)*(Total percent of the species’ diet composed of fruits, seeds, plus plants)]

We then considered several metrics of cultural ecosystem services (Table S1). We first estimated identity and iconic value to the US population as a whole using the popularity score of “celebrity” birds (Schuetz & Johnston 2019). We created subsets of all species to include those that were ranked as “celebrity” (n = 37), which are those that have above average national interest when considering national-level encounter rates (“popularity”) and have low geographic alignment in interest, or interest outside of where they are found (“low congruence”). This is because prior work has demonstrated that people may only perceive a subset of birds around them (Belaire et al. 2015). Therefore, low popularity scores likely indicate lack of awareness rather than a disservice per se. Thus, we used weighted abundances of “celebrity” species by weighting observed abundances by the species’ continuous popularity scores [Equation 4].

[Eqn 4]: Per-point identity and iconic value index = Σcelebrity species’(number of individuals detected * continuous popularity score)

We then generated a metric of cultural ecosystem service provisioning to the growers whose farms we surveyed using data from Smith et al. (2021). Smith et al. (2021) distributed a grower questionnaire survey to 54 farmers, including the 27 who managed farms included in this study alongside more farmers who managed similar farms in California, USA. These questionnaire surveys were conducted under the Washington State University Office of Research Assurances Institutional Review Board (IRB) that deemed it exempt from the need for IRB review (certification number 16610-001). Farmers provided open-ended responses to questions asking which species were considered beneficial or harmful to the farm and why, as well as which species farmers were attempting to attract/repel and why. Based on these open-ended data, we generated a metric of cultural ecosystem service provisioning to the farmers by first calculating a salience/interest metric similar to Schuetz & Johnston (2019) and then conducting a sentiment analysis (Lennox et al. 2019). We began by quantifying the number of times different species, families, suborders, and orders of birds were mentioned to the finest level possible (Data S2 in this Dryad Dataset, n = 263 data points). Of the 263 rows we coded, we were able to identify order in 90.4%, family in 79.8%, genus in 49.0%, and species in 46.8%. Considering this frequency distribution, we used family as the taxonomic unit in our analyses and excluded the comments from which family could not be inferred (e.g., “songbirds” [Oscine suborder of Passeriformes] and “birds of prey” [representing birds from Accipitriformes, Falconiformes, and Strigiformes in our region]).

We estimated “interest/salience” of those families by extracting the residuals from a linear regression of the total number of individuals from each family observed across farms that returned the grower questionnaire survey (predictor variable) against the number of mentions of that family by farmers (response variable) (Fig. S2). We did so to account for how much more, or less, interest each family generated than expected for a given encounter rate to mirror our identity/iconic value derived from Schuetz & Johnston (2019). We re-centered interest/salience values by adding the most negative residual value to make all values positive. In this way, our index of interest/salience captures the degree to which a given farmer is interested in a particular family of birds, accounting for differences in how likely they were to encounter a bird of a given family (as measured by the total number of birds observed in a family).

Next, we conducted a sentiment analysis on the same grower questionnaire survey data to understand farmers’ attitudes towards the birds they mentioned. Two of the authors (OMS and AE) independently scored each mention to range from -3 (“very negative” sentiment) to 3 (“very positive” sentiment) from a scale that included “very negative,” “moderately negative,” “slightly negative,” “slightly positive,” “moderately positive,” and “very positive.” (See Data S2, sentiment code book). As is common in attitudinal studies (e.g., Thelwall & Buckley (2013); Hutto & Gilbert (2014)), we averaged the two sentiment scores (84% of the 263 rows were assigned the same score by the two coders) and averaged the sentiment across mentions of that family.

We then generated an index of grower appreciation using species with positive averaged sentiment values and an index of grower disapproval using species with negative averaged sentiment values. For each, we first summed the abundance of individuals from each bird taxonomic family per survey point. We then multiplied each family’s abundance by its interest/salience score and by its sentiment score. For the service [Equation 5] and disservice [Equation 6] indices, we summed the weighted abundances across species with positive and negative sentiments, respectively.

[Eqn 5]: Per-point grower appreciation index = Σfamilies with positive averaged sentiments(number of individuals detected * positive sentiment * interest/salience)

[Eqn 6]: Per-point grower disapproval index = Σfamilies with negative averaged sentiments (number of individuals detected * negative sentiment * interest/salience)

Finally, we estimated conservation value using the maximum Combined Conservation Score from the North American Bird Conservation Initiative State of North America’s Birds (2016). We considered ‘conservation need’ as a cultural ecosystem service in itself because of the greater value people assign to species in need of conservation (Schuetz & Johnston 2019). We used the Combined Conservation Score instead of species’ binary listing because only 3% of total observations were of species listed as at least sensitive at the state level (Data S1), precluding analyses. To calculate the per point conservation value index, we weighted abundances of each species that had Moderate (9-13) or High (14-20) Combined Conservation Scores by its Maximum Combined Conservation Score (CCSmax). We then summed across species’ weighted abundances for each point for each survey for the conservation value index [Equation 7].

[Eqn 7]: Per-point conservation value index (combined conservation score index) = Σspecies with moderate to high CCSmax (CCSmax * number of individuals detected)

We repeated analyses using all species weighted by their CCSmax, which yielded similar results. Therefore, we refer the reader to Tables S2-S5 and Fig. S3 for results from analyses using all species.

2.4 Local, farm, and landscape complexity

2.4.1 Local complexity

To capture the structural complexity of each survey point, we estimated the percent cover of ground herbaceous vegetation (0–0.5-m height class), low shrubs/crops (0.5–2 m), tall shrubs/crops (2–6 m), and trees (>6 m) within a 10-m radius of each point count location’s center (Fig. S4a). We divided the 10-m radius circles into four equal quadrants divided along the four cardinal directions (Kennedy et al. 2010). During each survey, we estimated the percent vegetative cover in each height class for each of the four quadrants. We then averaged estimates across the four cardinal directions for each height group to estimate percent cover by vertical strata. Vegetation surveys were conducted at each bird point-count location at each bird survey occasion. Finally, we averaged the ground, shrub, tall shrub, and tree cover estimates across the 8 surveys (4 years x 2 repetitions per year) for each point, giving us 106 averaged values for each of the four vertical strata to estimate the local complexity.

To obtain a single estimate of local vertical vegetation complexity for each survey point location, we conducted a principal components analysis using the ‘prcomp’ function in the ‘stats’ package in R version 3.6.3 (R Core Team 2020). First, we standardized values by calculating a z-score for each. The first two principal components (PC) combined accounted for 67.5% of the variation (Fig. S4b). Increasing values of PC1 (“local vertical vegetation complexity”; 41.0% of the variation) were associated with increased fullness of the shrub, tall shrub, and tree layer. Increasing values of PC2 (“ground cover”; 26.5% of the variation) were primarily associated with increased cover in the ground layer. We used PC1 in subsequent models because we were interested in local vertical vegetation complexity.

2.4.2 Farm-wide High Nature Value index

We measured farm intensification/extensification, or conservation-friendly management practices, by modifying the High Nature Value index (Pointereau et al. 2010; Smith et al. 2021). In brief, the High Nature Value index is a continuous metric from 1 (lowest conservation value/most intensive) to 30 (highest conservation value/most extensive). Farms are classified using 3 sub-component indices (“diversity of crops,” “extensive/intensive practices,” and “landscape elements”), which each get equal weight (10 points max). Farms that score highest on the “diversity of crops” indicator are typically small with high crop diversity and/or integrate livestock. Farms that score high on “extensive/intensive practices” typically use few inputs, are certified organic, and maintain low stocking densities of livestock. Farms that score high on “landscape elements” incorporate seminatural elements within their farms, such as hedges or wet grassland. See Smith et al. (2021) for full details on our modification. Each farm had one High Nature Value score to represent management across years.

2.4.3 Landscape complexity

To characterize landscape context, we calculated the percent seminatural land cover based on the 2016 National Land Cover Database (Dewitz 2019) using a 2.1 km radius buffer from each point count location (Fig. 2e-f) using R and FRAGSTATS 4.1 (McGarigal & Marks 1994; R Core Team 2020). Seminatural land cover included forest (deciduous, evergreen, and mixed), scrubland (dwarf scrub and shrub/scrub), herbaceous (grassland/herbaceous, sedge/herbaceous, lichens, and moss), and wetland categories (woody and emergent herbaceous wetlands). Categories not included in seminatural land cover were water, ice/snow, developed, barren, pasture/hay, and cultivated crop classes. We used a 2.1 km radius as the biologically relevant landscape scale (Jackson & Fahrig 2015) because it was the estimated weighted average home range size for birds detected on our farms (Smith et al. 2020b).

2.5 Final service and disservice provider index derivation 

We first estimated ecosystem-service-and-disservice-weighted abundance indices for each point per survey per year (which we label “average ecosystem-service-and-disservice-weighted abundance indices”). Then, we calculated the coefficient of variation for each of the indices per survey point across the eight temporal replicates as an estimate of temporal variability. The coefficient of variation is calculated by dividing the standard deviation by the mean (CV = σ/µ). Stability is the inverse of the coefficient of variation (1/CV) (Blüthgen et al. 2016), which is the metric we used in temporal stability analyses. 

Our results suggested that landscape context was important in promoting temporal stability. We hypothesized that this was due to a shift away from dominant, highly nomadic species (i.e., an identity or selection effect; Fig. 1c). Therefore, we conducted analyses examining the relative importance of local, farm, and landscape complexity on evenness of the overall bird community at each survey point. We averaged the evenness values across the 8 surveys and repeated our analyses described above used to examine temporal stability. We then examined the influence of evenness as a predictor of temporal stability on each of the ecosystem-service-and-disservice-weighted abundance indices examined.

Data S1: Species service and disservice classifications. All bird species used in analyses and their ecosystem service and disservice provider classifications. Tab 2 has a meta-data key.

Data S2: Sentiment analysis. Code book of grower open-ended responses to questions asking which species were considered beneficial or harmful to the farm and why, as well as which species farmers were attempting to attract/repel and why. Tab 1: examples of how statements were coded for sentiment. Tab 2: all responses from growers to our open-ended questions, "Please indicate which bird species you consider are the most beneficial to your farm and why (Please specify)," "Please indicate which bird species you consider are the most harmful to your farm and why (Please specify)," "Which bird species are you trying to attract to your farm and why? (Please specify)," and "If you use any repellent techniques, please indicate which species or bird groups (such as sparrows, starlings, finches, corvids, birds of prey, etc.) you target and why." Tab 3: scorer 1 (Alejandra Echeverri) coding. Tab 4: scorer 2 (Olivia Smith) coding. Tab 5: assigned sentiment scores and assignment of bird taxa (Order, Family, Genus, Species) mentioned. Tab 6: meta-data key.

Data S3: Data used in analyses. Tab 1: data used for pairwise correlations between average ecosystem-service-and-disservice-weighted abundance indices and stability of ecosystem-service-and-disservice-weighted abundance indices (data needed to recreate main text Figure 3). Tab 2: data used for average ecosystem-service-and-disservice-weighted abundance index analyses. Tab 3: data used for stability of ecosystem-service-and-disservice-weighted abundance indices. Tab 4: meta-data key.