Study site

Removal of juniper occurred in southwest Idaho, USA in the northern Great Basin (Fig 1) and began in August 2019. The study area is composed primarily of big sagebrush (Artemisia tridentata) and low sagebrush (Artemisia arbuscula) interspersed with one conifer species, western juniper (Juniperus occidentalis; hereafter, juniper). In areas classified as < 10% conifer cover at our study site, juniper has an average height of 2.7 m ± SD 2.1 and an average stem density of 19 trees/ha ± SD 25. In areas classified as >20% conifer cover at our study site, junipers have an average height of 3.6 m ± SD 2.3 and an average stem density of 198 trees/ha ± SD 193, but individual trees could reach heights of 12 m, and areal coverage of juniper can approach 60% [44]. Cheatgrass (Bromus tectorum) has invaded much of the study area, especially at lower elevations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study site.

Location of avian predator surveys in the Owyhee Mountains of southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.g001

The climate of the study area is typified by hot summers and cold, snowy winters with an average of ~ 35 cm of precipitation annually [40]. Elevation ranges from 1,250 − 1,920 m and topography is varied with low-lying riparian areas interspersed with open sagebrush tablelands and rocky ridgelines. Before European settlement, juniper was most likely limited to rocky outcrops in small portions of the study area, presumably by historical fire regimes [45]. Since European settlement, juniper has expanded into sagebrush communities forming a gradient of tree cover across the landscape, and areas of dense juniper are currently found along ridgetops and in drainages [31]. Lesser, more scattered juniper cover typifies open sagebrush flats. Cattle grazing is ubiquitous across the study site, occurring at lower elevations during April-May and moving to higher elevations as summer progresses.

The BOSH project will eventually remove juniper cover classified as < 20% (at a 2-ha scale) from habitat formerly dominated by sagebrush across a ~250,000-ha landscape to support sage-grouse and other sagebrush-obligate species. Our sampling encompassed ~30,000 ha, and in the fall of 2019 15,000 ha of juniper was hand-cut within our study site. Juniper was cut using chainsaws and scattered so that no debris was higher than one meter.

Raven and raptor counts

We conducted repeated-visit surveys for three years before and one year after juniper removal began. We surveyed 800-m transects (n = 37) to assess the effects of habitat structure on the probability of habitat use by avian predators within our study site. We surveyed each transect three times per year with at least two weeks between visits. We conducted surveys between May 1^st and July 15^th each year. We selected locations for survey transects using random points in a GIS stratified by category of juniper cover (1–10%, 10–20%, and >20%) and location with respect to treatment plans for juniper removal (Fig 1). Eighteen of 37 transects were within areas where juniper would be removed. Surveys consisted of walking transects with three stationary, 10-minute observation periods placed at the beginning, middle, and end of each transect [23, 46]. Observers recorded any avian predators seen or heard while walking between or while at stationary observation points. We recorded birds if they were perched, calling, or flying/circling within the transect sample area. Each survey lasted ~ 45 minutes in total. Along each survey transect, the amount of time spent at stationary survey locations was consistent, but walking surveys between stationary survey locations varied based on terrain. We limited our data to observations of birds within 500 m of the observer to allow for more precise estimates of the effect of habitat on the probability of use and because this allowed us to assume that birds were influenced by the habitat surrounding the survey transect. We used laser range finders to estimate distances for birds detected visually and by sound. A single observer conducted each survey and three observers conducted 98% of surveys over four years. We conducted surveys before 10 am local time and never during steady rain or when estimated wind speeds exceeded 10 kph.

Prey abundance

To test the effect of prey abundance on habitat use by avian predators, we estimated abundance, density, or presence for important prey groups. We considered the following measures for groups that are known to be common prey of both ravens and red-tailed hawks: densities of the most common species of small mammals including deer mouse (Peromyscus maniculatus) and Great Basin pocket mouse (Perognathus parvus); the relative abundance of songbirds; and ground squirrel presence/absence. Belding’s ground squirrels (Urocitellus beldingi), which occur in large semi-colonial populations, are the most common ground squirrel at our study site.

To estimate the density of small mammals, we deployed 740 traps at five of the avian-predator transects in 2017 and 896 traps at the same five avian-predator transects in 2018 and 2019. We selected trapping locations for small mammals that represented a gradient of juniper cover, cheatgrass cover, and shrub structure (see [44]). In 2017, we used one trap array that consisted of 148 traps at the center of the avian predator transect. In 2018 and 2019, we used three trap arrays of 64 traps each at the center and ends of avian predator transects. We trapped small mammals over nine days broken up into four- and five-day sessions one month apart.

To estimate songbird abundance, we conducted 1,269 point-count surveys over four years along our avian-predator transects. Each survey transect for avian predators had three point counts stations placed at either end and in the middle of the survey transect. We surveyed songbirds concurrent with avian-predator surveys. We conducted 10-minute point counts three times per year at each avian-predator transect. We limited songbird observations to within 100 m of the point count station. We conducted songbird surveys from sunrise to 10 am and never during steady rain or winds stronger than 10 kph.

We noted the presence or absence of ground squirrels within 100 m of a survey transect because ground squirrels are an important food resource for red-tailed hawks. Ground squirrel presence is therefore treated as a categorical predictor variable.

Prey measures

We used measures of individual prey groups as predictor variables to assess the importance of prey for habitat use by avian predators. We estimated the density of small mammals in response to habitat characteristics at transect locations using a spatially explicit capture-recapture design (see [44]). We predicted small mammal density using the top-ranked models for both deer mice and pocket mice for sites where we did not sample mammals. We used vegetation measurements taken at each raptor-raven survey transect to predict small mammal density. The top model used to predict deer mouse density included a quadratic term for juniper cover within 100 m of the survey transect. The top model used to predict pocket mouse density included a quadratic term for mean sagebrush cover. We recorded the presence of ground squirrel colonies on survey transects using a categorical presence/absence approach. Ground squirrel presence on a survey transect was consistent across years, so each survey transect received a single classification for ground squirrel presence.

To calculate an index of songbird abundance, we pooled observations of all individuals and species and estimated the mean relative abundance of songbirds at the three points per transect. We did not adjust songbird observations for detection probability because differences in detectability among species would bias our abundance estimates, and previous studies examining the influence of songbird abundance on habitat use by avian predators have also used unadjusted relative mean counts [11].

Habitat characteristics

We focused our sampling and analysis on natural and topographic features that may influence habitat use by avian predators. We also included anthropogenic landscape features in our data collection (e.g., cabins, roads) because the effects of human subsidies such as artificial vertical structures, roads, and food are well-established as characteristics that influence the distribution of raptors and ravens. However, our study site was relatively free of human infrastructure compared to similar studies of habitat use by avian predators. (e.g., [15, 23]).

We classified juniper cover across our study site with imagery collected at a 1-m scale from the National Agriculture Imagery Program (NAIP). To identify juniper in the image, we conducted a supervised image classification in ArcGis [47]. Next, we manually corrected any misclassifications for each of our transects using visual inspection and ground truthing. We then calculated the area of juniper cover within 100, 250, 500, 750, 1000, 1500, 2000, 2500, and 3000 m of our survey transects. These distances span a range of reported movements during the breeding season for ravens (mean movement 570 m from nest, 6.6 km² nest territory) and 40 km² core use area for non-breeding individuals [15, 23, 48] and breeding red-tailed hawks ( = 3.88 km² nest territory [14]). Three of our survey transects featured juniper mixed with mountain mahogany. We included mountain mahogany in our juniper-cover layer because individual tree species are not discernible in our classification, and we are interested in the effect of vertical structure on habitat use by avian predators. Thus, we assumed that avian predators respond to both tree species similarly, and that any association with trees in sagebrush-dominated habitats is a function of structure and not tree species composition. We calculated distance to cliff from each transect by defining cliffs as areas with greater than a 60° slope. We used a 10-m resolution digital elevation model (DEM; U.S. Geological Survey), and our definition of cliff areas as > 60° slope captures a 90° cliff that is 17 m high, a 100° cliff that is 25 m high, and a 110° cliff that is 54 m high. We also included a measure of distance to water using a GIS stream layer in combination with ponds created for livestock. We calculated distance to improved road and distance to nearest human dwelling for each transect.

Occupancy models

To assess the effect of habitat characteristics and juniper removal on habitat use by avian predators, we used Bayesian multi-season occupancy models [49]. Multi-season occupancy models [50, 51] are an extension of single-season occupancy models that allow for the estimation of changes in occurrence probability between seasons (in this case, years) through the estimation of parameters of extinction and colonization probability. Multi-season occupancy models assume that occupancy of a survey location is closed during each year but may change between years. If there is the possibility that the closure assumption is violated, as often happens for highly mobile species, the occupancy estimator may instead be considered the probability of habitat use [50]. We used the auto-logistic formulation of the multi-season occupancy model [49] to allow for inference on the effects of habitat covariates on overall occupancy probability as opposed to a decomposition of occupancy into colonization and extinction parameters. Use of the auto-logistic formulation is suitable for a limited sample size of unique sites and puts the inference focus on occupancy probability for each site for each year, as opposed to colonization and extinction probability [51].

To account for the imperfect detection of avian predators, we tested covariates that can affect the probability of detection using leave-one-out cross-validation [52]. We allowed detection to vary by year for all models and tested the effect of a terrain roughness index (TRI) and time of year on detection. We did not include an observer effect because a single observer (ACY) conducted 344 of 409 avian predator surveys (85%). Because we were interested in testing the effect of juniper cover on habitat use by avian predators, we did not include juniper cover as a covariate on detection in our occupancy models. We assumed that juniper cover would not lower our ability to detect large, loud species such as ravens. However, to validate this assumption we tested an exploratory model that included juniper cover within 100 m of the transect as a covariate on detection and with no covariates on habitat use. This exploratory model was not included in our set of candidate models. We standardized all predictor values and used normally distributed, non-informative priors for all covariate parameters (mean = 0, precision = 0.01).

Variable and model selection

Uncertainty about the spatial scale at which habitat features may influence ecosystem processes is common for ecological studies. As a result, researchers are often interested in testing the effect of a habitat covariate at several spatial scales. However, the inclusion of all habitat covariates at all potentially relevant scales can lead to models that are difficult to interpret [53]. We reduced the number of spatial scales for variables included in our models following the screening procedure recommended by Stevens and Conway [53]. We fit univariate multi-season occupancy models for each scale (100 m– 3,000 m) of juniper cover that we quantified. We also tested other juniper metrics at these scales, including clustering and proportion of landcover comprised of three cover categories (<10% cover, 10–20% cover, and >20% cover). We estimated clustering using a K-means nearest neighbor analysis in ArcGIS. We estimated proportion of juniper cover categories by categorizing 30 m² pixels based on our juniper cover classification. We then compared the predictive power of each model in the set with leave-one-out cross-validation. Preliminary results from our model set testing the effect of different spatial scales of juniper measurement on habitat use suggested that ravens respond most strongly to the proportion of >20% juniper cover category habitat within 100 m of a transect. (β = 2.83, 90% Crl 0.41, 6.16; 99% posterior direction). However, because we were also interested in the relationship between all categories of juniper expansion and habitat use by ravens, we used continuous percent cover within 100 m of a survey transect, which was also a competitive model, as a covariate in our occupancy model set (β = 1.32, 90% Crl 0.13, 3.02; 99% posterior direction, Table A1 in S1 File).

Once we determined the spatial scale and measurement of juniper that best predicted habitat-use probability for ravens and red-tailed hawks, we constructed a model set that included combinations of habitat features in addition to juniper that may influence habitat use for avian predators (Table 1). We then ran each model for at least 300,000 iterations using JAGS called from R [54]. We assessed model convergence using values generated by JAGS [55] and model convergence using visual inspections of traceplots [56]. We compared the predictive power of each model in the set using leave-one-out cross-validation [57, 58] and assessed model fit using the Freeman-Tukey test statistic to generate a Bayesian p-value [59]. Models with p-values close to 0.5 adequately fit the data, while extreme values close to 0 or 1 indicate a lack of model fit due to over- or under-dispersion. Finally, for each covariate in a model, we calculated the posterior direction, which is the percentage of the posterior distribution that has either a negative or positive value, depending on the sign of the covariate point estimate [60]. We calculated the average probability of habitat use and average detection probabilities for each year. In the final stage of modeling, we carried forward our top habitat structure model and tested for additive or interactive effects of prey abundance and habitat structure on the probability of habitat use for avian predators (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Models used to test effects of habitat features on detection and probability of habitat use for common ravens (Corvus corax) and red-tailed hawks (Buteo jamaicensis) in southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.t001

Permits and ethics statement

Field work for this study was carried out under Idaho Fish and Game research permit 161213 (T.N.J.). Sampling of small mammals was approved by the Institutional Animal Care and Use Committee of the University of Idaho (protocol #2017–14).

Effect of structure on habitat use

We completed 409 avian predator surveys on the 29 transects sampled from 2017 to 2020 and eight additional transects sampled from 2018 to 2020. In 2017, we completed three surveys at 21 transects, two surveys at 10 transects, and one survey at two transects (n = 76 surveys). In 2018–2020, we completed three surveys at each of 37 transects (n = 111 surveys/year). Mean juniper cover within 100 m of all survey transects (n = 37) was 7.5% (8.1% SD) before juniper removal and 5.2% (8.4% SD) after removal, representing an average reduction in juniper cover of 30.6% per 10 ha. At transects where juniper was removed in fall 2019, average juniper cover within 100 m of treated transects decreased from 4% (4% SD) to 0.009% (0.02% SD), representing an average reduction in juniper cover of 99% per 10 ha (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Sampling effort for surveys of common raven (Corvus corax) and red-tailed hawk (Buteo jamaicensis) in southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.t002

We observed interannual variability in the number and spatial distribution of detections for both avian predator species (Table 2). Raven habitat use was positively influenced by percent cover of juniper and proximity to water (Fig 2A, Table A2 in S1 File). As juniper cover increased within 100 m of a transect, the probability of habitat use for ravens increased (β = 1.74, 90% Crl 0.32, 3.76; 99.5% posterior direction), and the highest probability of habitat use for ravens was areas with > 20% juniper cover. Credible intervals for the effect of distance to water on habitat use by ravens overlapped zero (β = -0.65, 90% Crl -1.40, 0.03), but most of the posterior distribution for distance to water indicated that the probability habitat use by ravens declined as distance to water increased (95% posterior direction). The average probability of habitat use for ravens declined by 55% in the year following juniper removal (2019 = 0.77, SE 0.05, 2020 = 0.34, SE 0.06, Table 3). Posterior predictive checks and visual inspections did not indicate a lack of fit or lack of convergence for the most predictive model (Freeman-Tukey p value = 0.35, values = 1). We included TRI as a covariate for detection in our raven MSO models because TRI had the most predictive power for raven detection (β = 0.22, 90% Crl 0.008, 0.44; 90% posterior direction, Table A3 in S1 File). Juniper cover had a positive effect on the probability of detection of ravens (β = 0.43, 90% Crl 0.23, 0.59, 99% posterior direction, 0.12 Bayesian p-value). The probability of detection of ravens ranged from 0.43–0.58 (90% Crl range 0.27, 0.69, Table 3) over four years.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Most predictive models.

Probability of habitat use for A.) common ravens (Corvus corax) and B.) red-tailed hawks (Buteo jamaicensis) in southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Derived parameters for the top-ranked common raven (Corvus corax) occupancy model in southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.t003

Probability of habitat use by red-tailed hawks was most strongly influenced by distance to a cliff, but differences in estimates of expected log predictive density (ELPD) suggest that this top model was only marginally more predictive than the null model. Probability of habitat use for red-tailed hawks declined as distance to cliff increased (β = -4.09, 90% Crl -7.57, -1.08; 97.3% posterior direction; Fig 2B, Table A4 in S1 File). Bayesian p-values suggest overdispersion in the models. We did not detect a statistically significant change in habitat use estimates for red-tailed hawks following juniper removal (Table 4). Red-tailed hawks responded most strongly to the clustering index of trees at a scale of 500 m, responding positively to more dispersed juniper. However, confidence intervals overlapped zero (β = 1.72, 90% Crl—4.82, 8.02; 77% posterior direction, Table A5 in S1 File). We did not identify important predictors for detection probability for red-tailed hawks (Table A6 in S1 File). Percent juniper cover did not have a strong effect on detection probability for red-tailed hawks (β = -0.15, 90% Crl -0.39, 0.09), which ranged from 0.18 − 0.24 over four years (90% Crl 0.10, 0.32).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Derived parameters for the top-ranked red-tailed hawk (Buteo jamaicensis) occupancy model in southwest Idaho, 2017–2020.

https://doi.org/10.1371/journal.pone.0286478.t004

Effect of prey on habitat use

Estimates of prey abundance alone did not influence habitat use for either ravens or red-tailed hawks. For ravens, a model that included the relative abundance of songbirds along with juniper cover ranked higher than the top-ranked model that included habitat structure only. However, as songbird abundance increased the probability of habitat use by ravens decreased (β = -0.80, 90% Crl -1.73, 0.12; Table A7 in S1 File) suggesting that ravens are less likely to use habitat with a higher abundance of songbirds. For red-tailed hawks, models including songbirds and small mammals as predictors of habitat use were ranked higher than distance to cliff alone (Table A8 in S1 File). However, increased songbird abundance was associated with a lower probability of habitat use for red-tailed hawks (β = -4.88, 90% Crl -9.17, -0.54), and credible intervals for the effect of small mammal density on habitat use by hawks widely crossed zero.

Effect of structure on habitat use

The establishment of trees and other woody vegetation in rangeland habitats is occurring globally as a consequence of altered fire regimes, climate change, enhanced atmospheric CO₂, and livestock grazing [65, 66]. Conifer trees have expanded into shrub-steppe habitats in southern Canada [67], acacia (Acacia saligna) and pines (Pinus spp.) have become established in the fynbos shrublands of South Africa [68, 69], and pines have invaded the high Andean paramos ecosystem [70]. The establishment of trees in these systems has had negative impacts on some native wildlife associated with vegetation that existed before expansion. As the expansion of woody plants continues in other rangeland ecosystems across the globe, understanding how expansion affects distributions of both predator and prey species and subsequent predator-prey dynamics will be an important part of conserving grassland and shrubland-dependent species. Where habitat structure can be modified through the removal of expanding woody species, it will be important to assess whether there are limitations to the effectiveness of removal as a restoration technique, or if additional efforts are required. For instance, tree removal can increase cover of understory vegetation such as bunchgrasses and sagebrush shrubs [71, 72]. However, variation in soil condition, hydrologic factors, and the pre-removal dominance of conifer can limit the response of understory plants to conifer removal [73]. Further, variation in plant responses to tree removal techniques (i.e., mastication, hand cutting, etc.) may lead to the establishment of invasive annual grasses and contribute to differences in the recovery of faunal communities [74].

Effects of generalist predators are a widespread conservation concern for sensitive prey populations, and efforts to control predator populations can be difficult and controversial [75]. The conversion of sagebrush habitat to conifer woodlands has reduced and fragmented habitat for sagebrush-obligate wildlife, and increased abundances of generalist predators can compound the negative effects of habitat loss on prey species [76]. For example, conifer expansion has been linked to population declines for greater sage-grouse, and increased predator populations are suspected to play a role in these trends [32, 33]. However, evidence for an association between juniper expansion and numerical or functional responses of avian predators has been limited or speculative (e.g., [15, 23, 77]). Our results show that ravens are more likely to use sagebrush habitat experiencing conifer expansion, a relationship that has implications for the conservation of sagebrush-associated wildlife.

Subsidized populations of generalist predators can significantly impact prey populations because predation may continue even after the prey density becomes very low [78]. There are likely fewer anthropogenic subsidies at our study site than in many areas of the western U.S. given the low human density of Owyhee County, but raven populations more broadly have benefited from human development, leading to increased abundance in the Great Basin [79]. Housing density is low at our site (0.02 houses/km²), and two lightly used dirt roads run through the study site (road density is 0.08 km/km²). There are no agricultural fields within 10 km (distance from a transect to agriculture ranged from 10–43 km) and no transmission lines within 21 km (distance from a transect to a transmission line ranged from 21–55 km) of the study site. As a result, the effects of natural structures on the probability of habitat use are less likely to be confounded by the positive effects of human subsidies and structures. However, further research is needed to elucidate the relative effects of conifer woodlands and anthropogenic resources on the probability of habitat use for ravens. The addition of human structures to a landscape that also contains conifer woodlands may have an additive effect on the probability of habitat use for ravens. Ravens are ‘incidental’ hunters, consuming a wide variety of prey including small mammals, songbirds, and lizards [80]. As the amount of habitat used by ravens increases, the likelihood that ravens will incidentally prey on sage-grouse nests may also increase [36, 81].

For highly mobile species such as avian predators, violations of the assumption that animals remain in a survey location throughout a season may require that researchers consider the effects of habitat on an animal’s presence and availability to be observed (i.e., “availability”; [50, 82]). Availability is defined as the probability that an animal is present at a survey location and able to be observed. Ravens are large, conspicuous birds that are easily seen and heard over long distances. Previous studies have sometimes assumed that observer error in detecting ravens is close to zero and estimated availability of ravens in relation to habitat, often referred to as use (e.g., [13]). For example, a study of ravens in the eastern United States estimated a 0.99 probability of detecting ravens over three visits to cliff locations if those locations were in fact occupied [83]. When the objective is to estimate the effects of habitat on an observer’s ability to detect an avian predator, researchers most often test for effects of habitat features that can obscure avian predators, including trees or rough terrain. For example, O’Neil et al. [23] found support for using a ‘viewshed index’ that factored tree cover and TRI into an observer’s ability to detect ravens. We found that juniper cover and TRI positively influenced detection of ravens, indicating that ravens are spending more time in habitat featuring juniper, regardless of whether or not ravens were being obscured from the observer by juniper or landscape features. At 0.43–0.58, our estimate of detection probability for ravens is higher than other studies (e.g., 35%, [23]). Given that we truncated our detections to 500 m, a smaller area than other studies (e.g., [23, 84]), we assume that our ability to see and hear ravens was high and that our estimate of detection probability represents availability to a large degree. Therefore, the fact that both the probability of habitat use and availability are positively influenced by juniper cover provides further support for a relationship between juniper cover and habitat use by ravens. For red-tailed hawks, a relatively low detection rate in our data is likely a result of low density and home range sizes larger than the area of our survey transects. Low detection rates can bias occupancy estimates if there is unmodeled, non-random movement in and out of sampling units often associated with the breeding season or migration [50]. However, we tested a model with time of year as a covariate for detection and this model was not supported. When the closure assumption is violated but individual movements in and out of sampling units is random, the occupancy estimator is likely unbiased. In this case, occupancy estimates can be interpreted as probability of use [50].

Effect of prey on habitat use

Similar to others, our study suggests that habitat structure is relatively more important than any specific prey resource for generalist avian predators [11, 27]. In fact, ravens were less likely to use habitat with higher abundances of songbirds and densities of small mammals. Densities of small mammals at our site, largely driven by deer mice, are highest at 10% juniper cover and decline as juniper cover increases [41]. Songbird abundance and diversity also increased in areas featuring early juniper expansion because the habitat can support sagebrush, ecotone, and conifer-associated songbirds [44, 67]. Longer-term data on prey-population distributions may reveal more complex predator-prey relationships that influence habitat use by generalist avian predators. It may also be that for generalist predators, no one prey population strongly influences habitat use. Data collected on prey populations at finer scales that account for within-season population variability and estimation errors associated with predictions for small mammal density and songbird abundance may improve inference. However, given the strong relationship we observed between habitat use for ravens and juniper woodlands, a disconnect between habitat use and prey abundance may be explained by the strong influence of vertical structure on habitat use (also reported in [23]).