We used a nested design and selected three clustered sites within a 1.5 km radius circle and repeated this design across the region . The sites exist along a landscape gradient of semi-natural habitat density and agricultural land use density within a 1.5 km radius. Organic farms sites were all certified organic and used less intensive practices than paired nearby conventional farms that used more intensive practices . Many farms grew a single crop variety, most commonly Brussel’s sprouts, strawberry, broccoli, or lettuce. Farms with multiple crop varieties included a mix of vegetables , beans, squash, tomatoes, herbs, and strawberries. Within each farm pair, we chose monitoring sites in the center of fields planted in annual crops with similar proximity to natural habitat. We used one detector for each site and placed detectors within fields and at the edge of woodland patches .We sampled bats at all sites with passive acoustic bat detectors from mid-June to early September 2014. Bat detectors were mounted on t-posts and microphones were elevated on 3 m PVC poles attached to the t-posts. We monitored bat activity levels and species richness at all sites. We did not compare feeding buzzes due to high subjectivity in distinguishing between a bat inspecting research equipment, a novel structure in their environment, and pursuing insect prey . We sampled each site for 6–7 nights during one sampling period from sunset to sunrise to account for high variability of bat activity across nights; the three sites clustered within each landscape were sampled simultaneously to reduce variability and increase sampling efficiency . Recorded calls were processed using Kaleidoscope V2.3.0 to filter noise files and split files to a max duration of 5 s,fodder system which we defined as a bat pass and used in subsequent analyses.
Files were automatically classified using Sonobat V3.1 West and then manually vetted by a team of trained technicians. To account for potential bias among technicians, we worked as a team until our classification decisions were at least 90% in agreement and used a decision key to finalize call identifications. Remaining calls that could not be positively identified to species were classified by characteristic frequency into three phonic groups, with species that comprise each phonic group listed in parenthesis: 50 kHz , 40 kHz , and 30 kHz . To account for differences in how bats perceive their environment, we grouped a subset of the species recorded in the study region based on their echolocation call structure, characteristic frequency, and foraging ecology into two functional guilds: clutter-adapted and open space bats. Of the 12 species recorded in the study region, two were classified as clutter-adapted bats and four were classified as open-space bats . These groups include the most common species in the study region and represent 97% of recorded calls. We used guidelines described in Schnitzler and Kalko and Buchalski et al. and echolocation call parameters described by the Humboldt State Bat Lab to partition species into functional guilds. Species in the open-space guild have call characteristic frequencies < 30 kHz and > 6 ms in duration, which experience less environmental attenuation and are therefore suited for foraging in uncluttered, open areas. Species in the clutter-adapted guild have calls with characteristic frequencies of > 45 kHz and duration < 6 ms. Bats in the clutter-adapted guild are able to forage in highly cluttered forest habitat by using short duration, high frequency calls to distinguish insect echoes from echoes produced by background clutter .We designed bucket style light traps with 12 W black light bulbs and clear plastic vanes and a brewer’s funnel with a mesh collecting bag containing a 2cm2 piece of pesticide strip .
Light traps were deployed simultaneously with acoustic detectors and were programmed to turn on at civil sunset for three hours during the first two nights that the detectors were deployed. Light traps were placed 10 m from bat detectors to prevent potential high-frequency interference from light traps from being recorded by bat detectors, either within fields or along linear habitat edges. Insects were later sorted and identified to order. We measured the length of each insect to estimate biomass using relationships described for terrestrial California insects in Sabo et al. .We surveyed local vegetation within concentric circles centered on acoustic monitoring sites. Within a 25 m radius, we measured the number of weed morphospecies and visually estimated the total number of flowers within the crop field. Within a 50 m radius, we measured the number of crop varieties, average crop height, and the percent of different ground cover types. For the percent of ground cover types, we focused on cover classes exceeding 5% of ground cover, and included bare ground, crops, non-crop herbaceous vegetation, and woody vegetation. Within both a 50 m and 100 m radius, we measured the number of trees > 30 cm circumference and the number of tree species. For all farms, we estimated the percent of field margins in insectaries , the percent cover of insectaries in crop fields, and the percent of weedy field margins. We gathered information about pesticide application on each farm . Most growers applied broad spectrum insecticides and/or fungicides, although a few smaller, organic operations reported that they had not applied pesticides in years. Some of these pesticides were widely used by organic and conventional growers; only conventional growers used organophosphates. For each farm, we categorized frequency of pesticide applications according to these categories: 1=pesticides applied every 1–10 days; 2=pesticides applied every 11–30 days; 3=pesticides applied less than once a month; 4=pesticides applied less than once a year.
We also gathered information about distance to habitat features that are important for bats, and that we could not standardize across the trio of clustered sites. We used Google Earth to calculate the shortest distance to water and distance to the nearest linear habitat element, such as a treeline, hedgerow, or forest edge .We created land cover maps from the National Agricultural Imagery Project 2014 using manual likelihood classification in ArcMap V10.3.1 . We categorized agriculture and semi-natural habitat, and smoothed and resampled maps to a resolution of 7 m2 to reflect ground-truthed land cover classes. We then calculated the land use density of agriculture and semi-natural habitat within a 1.5 km radius, an ecologically relevant scale for bats and insects .We used total bat activity, clutter-adapted bat activity, open-space bat activity, diversity, and species richness as response variables. Bat activity is a measure of relative abundance calculated for each site over a minimum of six nights of sampling, and thus not true count data. Therefore, we opted to log transform the average number of bat passes per night to meet assumptions of normality with a Gaussian error distribution,fodder system for sale instead of using the Poisson error distribution recommended for count data . We used the Chao1 species richness estimator to assess the completeness of species inventories and to compare species richness across site types . We calculated estimated species richness and diversity using all calls identified to species and the 40k phonic group with the package “vegan” in R software, version 3.3.1 To compare response variables between organic farms, conventional farms, and natural sites, we used a randomized block ANOVA, with sites blocked by landscape cluster, and conducted multiple comparison tests using Tukey contrasts as implemented in the “multcomp’’ package in R . We checked that assumptions of normality and homogeneity of variance were met using the Shapiro-Wilk test and the Brown-Forsyth test in the “HH” package in R . To compare community composition between organic farms, conventional farms, and natural sites, we used a one-way Analysis of Similarity test. We used the Bray-Curtis index to characterize species dissimilarity between habitat types using the “vegan” package in R . We used linear mixed models to compare differences in responses to local, on-farm management practices. We selected predictor variables that represent agroecological farming practices and/or practices characteristic of intensive agriculture that growers could potentially manipulate. We included the percent of weedy field margins, number of crop varieties , average height of herbaceous vegetation , distance from a linear element, and percent woody vegetation . We did not include the percent of field margins planted in insectaries or the percent of field planted in insectaries as covariates because many farms did not utilize either of these practices. Selected predictor variables were not strongly correlated , although they were somewhat correlated with a number of other farm characteristics . In addition, we included distance to water as a fixed effect and landscape cluster as a random effect to account for variation in the surrounding landscape, and to avoid spatial autocorrelation problems.
We used standardized model residuals obtained by the top model for total bat activity to create a variogram and verify spatial independence . We performed model selection using the dredge function in the package “MuMIn” after testing for normality . We log-transformed total bat activity, clutter-adapted bat activity, and crop diversity, and used a square root transformation for open-space bat activity to meet assumptions of normality. We tested all possible combinations of explanatory variables and selected the top model if the corrected Akaike’s Information Criterion score was at least two points less than the next best model . When the top models had similar AICc scores , we selected the top models within two points of the lowest AICc score and adopted a model averaging approach to obtain estimated coefficients and the relative importance of predictor variables from the top model set. We visually inspected the relationships between predictors in top models and responses, and noted an outlier site for crop diversity and weed morphospecies, with most sites ranging from 1 to 10 crop varieties, and one site with 39 crop varieties. We decided to remove the outlier after exploring its leverage on LMMs and Pearson correlation coefficients. The effect of the outlier was minimal in most cases, but it changed the significance of the effect of crop diversity on open-space bat activity, and the strength and significance of the relationship between crop diversity and weed morphospecies within 25 m. We asked if the same local management practices and farm characteristics that predicted bat activity would also predict the biomass of insect orders that are known to be consumed by bats and were commonly collected in light trap samples . We tested LMMs that contained the same set of predictors as bat models to assess if variables that predicted bat activity operated on bats via changes in abundance of insect prey. All models contained a random effect of landscape cluster. To meet assumptions of normality, we used a square root transformation for Diptera biomass, and a log transformation for Lepidoptera and Coleoptera biomass. We assessed Pearson correlations between bat activity and insect biomass for these common prey orders. We used LMMs to test for an interaction between local practices that significantly predicted bat responses and semi-natural habitat density using LMMs. We built a set of candidate models using the top model from model averaging and added an interaction with semi-natural habitat density , and compared models using AICc scores.Across 648 detector nights, we recorded 36,294 bat calls. Of these, we were able to confirm identification to species for 14,969 calls, and remaining calls were identified to phonic group. A total of 12 different species were recorded , and of those, Tadarida brasiliensis and Myotis yumanensis were the most common across all sites. Three recorded species are listed as species of special concern in California: Lasiurus blossevellii , Corynorhinus townsendii , and Antrozous pallidus . The species inventory appears to be complete for natural habitat, but not for farms, according to rarefaction analysis . We recorded fewer, high quality calls on farms compared to natural habitat. Bat calls recorded on farms had less noise interference than recordings from natural habitat, allowing us identify a greater number of species from fewer calls, resulting in steeper species accumulation curves that did not reach an asymptote.Categorical comparisons of lower- and higher-intensity agricultural systems have documented mixed effects on bat activity, diversity and species richness, and provide limited insight on the effects of specific, agricultural management practices.