The study site was in the Kaibab National Forest in northern Arizona (Fig. 1a; 36.64–36.57°N, 112.38–112.51°W; 1800–2100 m elevation). Annual precipitation (30-year average) is approximately 406 mm and occurs primarily during the winter/spring and the summer monsoon season. Precipitation was within 10% of the 30-year average for the 2015 and 2016 water years (October–September), with a wetter-than-average (~ 160%) 2015 summer and drier than average (~ 60%) 2016 winter and spring (PRISM Climate Group). Soils are shallow, very stony fine sandy loams derived from Kaibab limestone and sandstone (Soil Survey Staff 2015) and lacked a robust soil biotic crust component. Vegetation was a patchy and diverse mixture of pinyon-juniper woodlands, big sagebrush shrublands, Gambel oak and cliffrose brushlands, and forb-rich grasslands, including both warm-season (C4) and cool-season (C3) grasses. We encountered 135 unique plant taxa, including 17 species of shrubs, 19 grasses, 4 cacti, and over 95 forbs (Appendix A). The majority of species were native (Appendix A); however, non-native, invasive species had increased in abundance following the Bridger Knoll (1996) and Slide (2007) fires, prompting concern that observed increases in B. tectorum abundance were facilitating cheatgrass-fire cycles akin to those of the Great Basin (Sisk et al. 2010).
Native grass seeding and targeted spring grazing experiment (Questions A, C and D)
The experimental design paralleled that of a sister targeted grazing and seeding experiment conducted at a highly-invaded Great Basin site (Porensky et al. 2018). At both sites, the experiment was hierarchical with three levels (split-split plot). At the broadest scale, targeted spring grazing treatments were implemented using a randomized complete block design. Eighteen 20.2 ha paddocks were arranged into nine blocks (Fig. 1a). Blocks were randomly located within broader areas that were non-randomly selected to represent plant communities with low tree cover and high B. tectorum cover (average B. tectorum cover in these study blocks was 11% in 2014, and 6 blocks included previously burned areas). For six of the blocks, one paddock per block was randomly assigned a targeted spring grazing treatment. Other blocks remained ungrazed due to unavailability of grazing animals for operational reasons; for these blocks, both paddocks were analyzed as ungrazed (N = 6 grazed and 12 ungrazed paddocks). Targeted spring grazing treatments were implemented 17 months after seeding treatments (described below) to minimize the negative impacts of targeted spring grazing on seeded species establishment. Twenty dry cows grazed each of the six spring-grazed paddocks for two weeks in April, 2016. We estimate that the grazing treatment achieved 30% utilization across grass species, based on (1) the reduction of grass size and weight as determined by ocular estimation and (2) clipping, drying, and weighing aboveground herbaceous biomass in 0.5 × 1 m frames at 9 random locations in each grazed and paired, ungrazed paddock soon after the removal of cows.
Within each paddock (middle level of the hierarchical design), we established eight 20 × 60 m experimental plots separated by ≥ 50 m. These plots were randomly assigned to four native grass seeding treatments, with two replicates of each treatment per paddock. Treatments included all combinations of two spatial arrangement treatments and two seed rate treatments (Fig. 1b). In the monoculture spatial arrangement plots, five native perennial grass species (described below) were seeded separately into adjacent 4-m wide strips within each 20 × 60 m plot. In mixture plots, the same amount of seed per species was used as in monoculture plots, but all seeds were mixed together before planting and seed was distributed across the entirety of each 20 × 60 m plot. Thus, monoculture and mixture plots differed only in the spatial distribution of seed. To avoid disturbance to soil and resident plants and because drill seeding was often infeasible due to very rocky soils, seeds were added using a broadcast seeding method. Plots were broadcast seeded using handheld broadcast seeders (for monoculture plots) or UTV-mounted broadcast seeders (for mixture plots) in November 2014. No rakes or chains were used following broadcast seeding.
At the finest experimental design scale (split-split plot), each experimental grass plot included three 20 × 20 m subplots, one of which was randomly selected for a coated seed treatment. Seeds were coated with a non-ionic alkyl terminated block co-polymer surfactant coating based on C1–C4 alkyl ethers of methyl oxirane–oxirane copolymers (Aquatrols Corporation of America, Paulsboro, NJ). This surfactant has been used to increase the wettability of water‐repellent soils (Fernelius et al. 2017; Kostka 2000) and can also improve plant drought tolerance in wettable soils (Ahmed et al. 2018).
For all seed addition treatments, five perennial grass species were selected from a pool of species native to the region, emphasizing a diversity of life-history strategies. Species included four locally abundant C3 grasses: Elymus elymoides (Raf.) Swezey [squirreltail], Pascopyrum smithii (Rydb.) Á. Löve [western wheatgrass], Poa fendleriana (Steud.) Vasey [muttongrass], Achnatherum hymenoides (Roem. & Schult.) Barkworth [Indian ricegrass], as well as a locally common C4 grass (Sporobolus cryptandrus (Torr.) A. Gray [sand dropseed]). Following NRCS recommendations, seed rates for the low rate treatment were: E. elymoides 6.11 PLS kg/ha; P. smithii 10.65 PLS kg/ha; P. fendleriana 1.3 PLS kg/ha; A. hymenoides 8.31 PLS kg/ha; and S. cryptandrus 0.22 PLS kg/ha; these rates were doubled for the high rate treatment. For a given seed rate, all plots received the same total number of PLS.
Pretreatment vegetation data were collected from eight 20 m transects per paddock in June 2014. These included 4 randomly-located transects within plots designated for seeding treatments, and 4 randomly-located transects in unseeded areas. For 10 evenly-spaced 20 × 50 cm quadrats per transect, we visually estimated aerial plant canopy cover (i.e. plant material that would intercept a raindrop) by species. For portions of the quadrat without plant cover, we estimated cover of litter (detached plant tissue less than 2 cm diameter), wood (woody plant matter larger than 2 cm diameter), rock (larger than 1 cm), moss/biocrust, and bare ground. We also recorded plant density and number of inflorescences by species for all vascular plants in two quadrats per transect.
Post-seeding data were collected in June 2015, and post-seeding and post-targeted spring grazing data were collected in June–July 2016. We recorded species-specific aerial cover and density at multiple sampling stations per seeded plot (Fig. 1b) as well as once at each unseeded plot. We estimated plant cover in 1 × 1 m quadrats using the visual estimation methods described above and recorded seedling densities of all seeded species in the northeast 50 × 50 cm corner of each quadrat frame. Because of small seedling sizes, it was difficult to confidently distinguish among some planted grass species, so seeded species seedling densities and aerial cover of seeded species seedlings were recorded in aggregate. We also recorded density and inflorescence counts by species for all vascular plants in the northeast 25 × 25 cm corner of the frame at two sampling stations per plot (Fig. 1b) and within each unseeded control.
In June 2015 and August 2016, we quantified standing biomass in each paddock by clipping all aboveground non-shrub biomass rooted inside 1 × 1 m and 0.5 × 0.5 m frames, respectively. We sampled at 5 locations per paddock in June 2015 (one randomly selected seeded plot and four unseeded locations), and at 22 locations per paddock (once per seeding treatment plot and at all 12 unseeded locations) in August 2016. Clipping locations were shifted 3 to 10 m each sampling effort to avoid resampling the same areas. Biomass was separated into four categories (B. tectorum, forbs, seedlings of seeded species, and all other native grasses), then oven-dried and weighed.
Species-specific B. tectorum exclusion study (Question B)
We sampled a subset of invaded paddocks to investigate the ability of different perennial grasses to potentially resist B. tectorum invasion. Sampling was conducted in October 2015 within the six paddocks (arranged into three blocks) that had the highest cover and density of B. tectorum. Within each paddock, we sampled along four random 20 m transects in unseeded areas. On both sides of the transect at each of ten evenly spaced points, we selected the nearest adult individual belonging to one of the five native perennial grass species seeded in the manipulative experiment. For each sampled plant, a 20 × 50 cm frame was centered against the edge of the plant base, extending perpendicular to and away from the transect. The frame was divided into three rectangular sampling regions (0–10 cm, 10–20 cm, 20–50 cm), with the 0 cm mark against the sample plant.
For each sample plant, the basal diameter and tallest leaf height were measured. Within each of the three sampling regions, B. tectorum density was recorded, and the cover class (1 = 0–10%, 2 = 10–50%, 3 = 50–100%) of each of five functional groups was visually estimated. Functional groups were: heterospecific grass (all but the species of the focal plant), conspecific grass (including only the species of the focal plant), forb, subshrub, and shrub.
To answer question A, “Do plant functional groups of northern Arizona differ in their association with B. tectorum abundance?”, we used 2014 pretreatment data averaged at the transect scale to produce 8 data points per paddock. We used generalized linear-mixed-models with a model selection and model averaging approach to determine which plant functional groups were most closely associated with B. tectorum cover, density and inflorescence counts at each transect. Predictor variables in a global model included the cover of six different plant functional groups (C3 perennial grasses, C4 perennial grasses, shrubs, cacti, forbs, and non-B. tectorum annual grasses; Appendix A). Block and paddock nested within block were included as random factors. We compared the global model to all possible subsets, including a null (intercept + random effects) model. Using all models within 4 AICc units of the best model, we performed full model averaging to further isolate the most important predictors of B. tectorum abundance. To test whether B. tectorum invasion was associated with litter cover, we regressed pretreatment B. tectorum cover against pretreatment litter cover for each transect, including random factors as described above.
To answer question B, “Do adult native perennial grass species differ in their association with B. tectorum abundance?”, we used data from the species-specific B. tectorum density counts conducted in October 2015. We used generalized linear-mixed-models with a model selection and model averaging approach as described above to determine which predictors were most closely associated with B. tectorum density at different distances from adult native perennial grass individuals. Predictors included species identity of the focal perennial grass, scaled size of the perennial grass (bunchgrass volume/maximum bunchgrass volume measured in the study), and cover class of forbs, heterospecific grasses, conspecific grasses, subshrubs, and shrubs in the same zone as B. tectorum. We calculated bunchgrass volume as π * (bunchgrass diameter/2)2) * (bunchgrass height). Random factors included block, paddock nested within block, and transect nested within both paddock and block.
For question C, “What are the best methods for establishing native perennial grasses in these systems?”, data on seeded species establishment were analyzed using generalized linear-mixed-models with a negative binomial distribution and a log link function due to zero-inflated count data. This distribution fit the data significantly better than other potential distributions (e.g., Gaussian, Poisson, Gamma). We ran two models using seedlings of seeded species per m2 as the response variable. Two models were necessary because it was not possible to evaluate all treatments across all plots. For example, we could not include data from areas we didn’t seed in a model investigating the effect of using coated versus uncoated seeds. For both models, random effects included block, paddock nested within block, and plot nested within paddock and block. The first model (seeding X grazing model) determined if plots planted with native grasses seeded at high or low rates had greater seeded species seedling densities than unseeded controls, and if targeted spring grazing affected seeded species seedling densities. For year one data, collected prior to initiation of targeted spring grazing treatments, the only fixed effect in the model was seeding rate (high rate seeding, low rate seeding, or unseeded). For second-year data, fixed effects included seeding rate, targeted spring grazing treatment, and their interaction.
The second model (planting strategies model) assessed the effects of different planting methods on seeded species densities, using only data from plots seeded with native grasses. Fixed effects included seed rate (high or low), spatial arrangement, seed coating, targeted spring grazing treatment (for second-year data), and all interactions among these factors. Seed coating subplot was added as an additional random effect nested within plot, paddock and block.
To answer question D, “Can native seeding or targeted spring grazing treatments affect B. tectorum cover or biomass?”, we analyzed B. tectorum cover and density data in 2016 using linear mixed-models. Random effects included block, grazing paddock nested within block, and plot nested within paddock and block. Fixed effects included 2015 B. tectorum cover (covariate), targeted spring grazing treatment, seeding rate (unseeded, low rate seeding, high rate seeding), and the interaction of targeted spring grazing treatment and seeding rate. To further explore the impact of seeded species on B. tectorum, we also regressed 2016 B. tectorum cover and density on seeded species seedling densities within each seed rate category (high, low, or unseeded), and B. tectorum cover levels from 2015 were included as a covariate in these regressions.
Finally, we used linear mixed-models to analyze B. tectorum biomass in 2015 and 2016. Each year was analyzed separately due to differences in data collection intensity between years. Fixed effects included targeted spring grazing, seeding rate (unseeded, low rate seeding, high rate seeding), and their interaction. Random factors included block and grazing paddock nested within block. For 2016 data, a negative binomial model structure was used due to non-normality of biomass data.
Analyses were conducted in R 3.4.0 using the lme4 (Bates et al. 2015) and MuMIn (Bartoń 2015) packages, and with JMP (JMP®, Version 12. SAS Institute Inc., Cary, NC, 1989–2007). Data were transformed or variance-weighted when necessary to meet model assumptions. All results were considered significant at p < 0.05 and marginally significant at 0.05 < p < 0.10, and are reported as means ± standard error (SE). Means separations were done using Tukey HSD.