Private land sites may use high-powered grow lights, drying fans, and visual barrier fencing, which could create potential wildlife disturbance . Such practices are less common on public land. It is possible that as cannabis production expands, particularly in the licensed industry, these forms of indirect impact may be more typical of cannabis production overall. Indeed, indirect effects of production practices on wildlife space use and behavior is a common concern for other agricultural crops , and may also interact with direct effects on mortality . Therefore, it is critically important to study both indirect and direct effects of cannabis on wildlife communities, particularly on private lands where research is lacking. Because outdoor cannabis farming is a land use frontier and therefore often characterized by different land use practices and patterns from traditional established farming in the US, it is uncertain whether other agricultural systems provide the best models to predict wildlife responses to cannabis development. Wildlife may use, avoid, or display differential responses to cannabis development, depending on whether production more resembles small scale countryside farming , industrial agriculture , or exurban/suburban development . In the case of differential responses, it’s also unclear whether cannabis production would have widespread enough effects to trigger mesopredator release , or generate novel food sources that could be exploited by behaviorally adaptable species like omnivores and small mammals . The small-scale, private-land cannabis farms for this study included one licensed recreational production site, one medically licensed production site, and six unlicensed sites. All farms were producing cannabis for sale, though in different markets depending on their access to licensed markets. We also had cameras placed in three hemp fields next to cannabis farms. We selected these eight cannabis growing equipment farms because they: were representative of the size and style of cultivation predominant in Josephine County in the years immediately following recreational legalization in 2015 , were all established after recreational legalization except for the medical farm, did not replace other plant-based agriculture, granted us permission to set up cameras on site, and were located next to a large section of unfarmed land that could grant researchers access in order to place cameras across a gradient of distance to cannabis farms.
Our sampled farms were small , had conducted some form of clearing for production space, and three had constructed some form of fence or barrier around their crop. Nonetheless, specific land use practices and production philosophies differed between farms . We cannot disclose farm locations, as per our research agreement for access. Monitored farms were clustered within each watershed: one farm in Slate Creek, five in Lower Deer Creek, and two in Lower East Fork Illinois River; however, most farms were also located near other nearby cannabis farms that were not directly monitored in this study. We placed unbaited motion sensitive cameras on cannabis farms as well as in random locations up to 1.5 km from the monitored farms. This is an expansion on previous camera research that only assessed on-site wildlife at these same farms . We placed cameras approximately 0.5 m off the ground to capture animals squirrel-sized and larger. We set cameras to take bursts of 2 photos, with a quiet period of 15 seconds. To guide the placement of cameras, we overlaid the area surrounding each cannabis farm cluster with a 50 x 50 m grid and then selected a random sample of at least one quarter of grid cells . We selected a 50 x 50 m grid size because we wanted to be able to detect fine scale space use responses of wildlife. The random sample was stratified by vegetation openness and distance to cannabis farm in all watersheds, and additionally by distance to clearcut in the Slate Creek watershed, such that cameras were placed in proportion to the landscape attributes and a distance gradient was achieved. When a selected site was inaccessible, we selected a new one that also met the same stratification criteria. We rotated 15-20 cameras through the sampled grid cells, ensuring each camera was deployed for at least one round of two week duration. Because of rotations and field constraints, all cannabis sites were not monitored at the same time or for the same length of time . Altogether, we monitored a total of 149 camera stations for a combined 4,664 trap nights. We then used a team of researchers trained to identify species found in the study area to sort photos by hand, grouping by species. We calculated spatial and descriptive covariates for each site to use in wildlife occupancy and detection models . First, we calculated spatial distance covariates. Our main covariate of interest was distance to cannabis farms.
To calculate distance to cannabis, we combined the location data for participating farms in our study with mapped data on Josephine County cannabis farms from 2016 aerial imagery . Then we calculated the minimum distance from each camera to its nearest farm using the package sf in R. We transformed distance to cannabis using a square root to help fit potential thresholds in wildlife responses. Next, we again used the sf package, this time to calculate the distance from each camera to the nearest major paved roadway, which was primarily highway 99 for most sites. For our two raster-based covariates, we used the raster , and exactextractr packages in R. We calculated the proportion of forested land cover within a 50 m buffer around each camera, and extracted the elevation in meters at each camera site. We also included some non-spatial covariates. We included a covariate for Julian date of each interval, as well as Julian date squared, to capture seasonal peaks. We then included an estimated distance at which a camera could still detect an animal , which was measured at camera setup. We also generated activity indices for dogs and humans by calculating the number of observations of humans or dogs, respectively, at each camera within the last three days, divided by the number of days the camera was active. This produced an activity rate where the beginnings or ends of placement rounds were on the same relative scale as all other days. All continuous variables were scaled so that they centered on 0 with a standard deviation of 1 and checked for correlations in R. Finally, we used additional categorical covariates to account for potential effects of geographic region and camera type. We assigned each camera a binary region variable based on which USGS Unit 12 watershed it was located in, such that Region1 represents Lower Deer Creek, Region2 for Lower East Fork Illinois River, and Region3 for Slate Creek. We created a binary variable for camera type. We gave a 0 to camera models that generally performed well in our study system and a 1 to camera models that generally seemed to perform worse or were older models .To assess the local space use response of wildlife to cannabis production, we used single-season, hierarchical single and multi-species occupancy models. Our approach is a departure from the typical use of these models to estimate occupancy in that we knowingly violated multiple assumptions of occupancy models: first, because cameras were spaced relatively close together compared to the home range of species included in the study, we have likely violated the assumption of independent cameras; second, as a result of the aforementioned spacing as well as sampling across two years , we likely violated the model’s assumption of geographic and demographic closure . We have done our best to account for these violations in our use of regional fixed effects, as well as our narrow interval of replication . However, given our interest was in space use associations and not estimates of occupancy, we believe the violations are a minimal issue. This use of occupancy models is not particularly unusual, as the use of occupancy modeling to assess space use is becoming more common in wildlife response studies, and even traditional uses of occupancy modeling are influenced by wildlife space use .
With the closure assumption violated, the occupancy probability estimate represents the likelihood that the animal occupied the site at any point during the study period, while the detection probability represents a combination of the probability that the species is detected and the intensity of use of the site within its larger range . This interpretation is common in camera trapping studies , but we proceed while being careful to acknowledge where appropriate that any covariate’s influence on detection probability is a combination of its effect on detection and the intensity with which an animal uses a given space. In addition, we have taken care to include variables in the detection process to account for what we anticipate to be the largest sources of variation in detectability, so that the other variables should primarily reflect space use intensity. We therefore interpret occupancy for the models as space use rather than true occupancy . We operationalize detection as a combination of intensity of use, and camera detectability or error . To examine animals’ space use in relation to distance from cannabis grow table farms, we first conducted single species occupancy analyses on nine wild and one domestic species . We summarized species observations on and surrounding cannabis farms and created detection histories using the package CamtrapR in program R using Rstudio . We used a 24-hr time interval because our focus was on estimating space use associations instead of occupancy , and a short interval reduced the likelihood of the same individual animal being detected on neighboring cameras . We modeled the space use probabilities of the most commonly detected species or those of particular ecological interest, including: black-tailed deer , black bear , bobcat , coyote , gray fox , black-tailed jackrabbit , striped skunk , California ground squirrel , tree squirrels , and domestic dog using the NIMBLE and nimble Ecology packages in Program R . We selected these species because they had sufficient detections to model , and because they covered a range of functional groups, including predators and mesopredators , omnivores , large and small prey , and a domestic predator . We included dogs as an added check on our modeling approach, as their general distributions and associations are already well known in the study system, unlike wildlife species. We modeled the observed data as a binary variable where 1 was an observation for a given species at camera station s, and 0 was a non-detection. We modeled the observed data for each species as a product of both true occurrence of a given species at a site and our probability of actually detecting it , which is also influenced by intensity of use at a given site. The model assumes that true occupancy is an outcome of a Bernoulli-distributed random variable, denoted zs~ Bern, where is is the probability that a given species used site s on any day during the survey period. We assumed that occurrence and detection probabilities varied by species, and that cannabis might influence both in different ways. For occupancy, we expected that increasing distance from cannabis farms would increase animal space use for all species except domestic dogs, and ground squirrels. We also expected that elevation and forested land cover would influence space use based on their importance in other wildlife studies . We expected distance to highways to negatively affect space use, and to function as a proxy for other non-cannabis forms of human land use in our study system. While we initially wished to include distance from clear cuts as the other major source of human disturbance in the study system, it was highly correlated with distance to highways, so we did not include it in our models. Finally, we accounted for potential regional differences in the three watersheds by including a fixed effect of region. We parameterized regional fixed effects using region-specific intercepts as described in the following equations. For the single species occupancy models, occupancy and detection varied by species . Recall that for our models, we are interpreting occupancy as space use, and detection as a combination of detectability and space use intensity .