Drought and invasive species are threats to primary productivity in grassland ecosystems worldwide. A central challenge in ecological restoration is establishing plant communities that can withstand these abiotic and biotic stressors. We tested the efficacy of a trait-based framework for enhancing drought tolerance and invasion resistance in experimental restorations of contrasting grassland systems, a perennial mixed-grass prairie in Wyoming, and an annual interior grassland in California. Each experiment included four trait-based seeding treatments: drought-tolerant, invasion-resistant, functionally diverse, and a random control. All seeded communities were subjected to extreme precipitation reduction (-50 % annual) and invasion by non-native annual grasses. We asked two questions: (1) Can we establish communities from seed that meet and maintain these desired trait-based targets? (2) Did the trait-based seeding treatments tolerate extreme precipitation reduction or resist invasion by non-native annual grasses? Based on analyses of compositional dissimilarities, we found that trait-based restoration targets were more difficult to maintain in an annual-dominant grassland rather than perennial-dominant, and targets composed of multiple community-weighted mean traits were more difficult to meet than targets that maximized functional diversity. In both grasslands, plant communities with drought tolerant traits maintained growth rates under reduced precipitation, but these effects diminished after multiple years of drought and during drought release. This highlights the importance of including a diversity of strategies when restoring plant communities. Our proposed invasion resistant traits did not consistently reduce non-native annual grass establishment, but communities with traits that conferred drought tolerance were the most effective at resisting invasion, suggesting that similar traits enhance drought tolerance and resist invasion in these grasslands. Our findings indicate that restored plant communities with high functional diversity may be able to respond to variable conditions better than communities with traits designed to meet specific restoration objectives. The results from this restoration experiment suggest that our understanding of the traits underlying drought tolerance is better than our understanding of the traits underlying invasion resistance. However, if drought tolerance enhances competitive ability in arid and semiarid grassland ecosystems, then physiological theories of resource use can be applied to enhance invasion resistance.

Site Descriptions

In Wyoming, the restoration experiment occurred on a topographically flat crested wheatgrass (Agropyron cristatum) pasture at the USDA Agricultural Research Service High Plains Grassland Research Station near Cheyenne, WY (latitude 41.1776 °N, longitude -104.8983 °W; 1900 m elevation, 370 mm mean annual precipitation, 6 °C mean annual temperature). We prepared the site (0.6 acres total area) for restoration by spraying glyphosate in September 2018 and again in May 2019 to kill the existing vegetation. Sixty-four experimental blocks (each 1.5 m x 3.7 m) were established in a grid.

In California, we established this restoration experiment at the University of California South Coast Research and Extension Center in Irvine, CA (latitude 33.683 °N, longitude -117.717 °W; 125m elevation, 365 mm mean annual precipitation, 17.5 °C mean annual temperature). The site (one acre total area) consisted of fifteen experimental rainfall manipulation structures (3 m (w) x 7.3 m (l) x 2.6 m (h)) (Figure 1, Step 5). In November 2019, existing vegetation and root systems from a prior experiment were manually pulled and soil was tilled to a maximum depth of 15 cm. A 1.5-meter perimeter fence was constructed to prevent entry by small mammals, but evidence of minor gopher disturbance was present in some plots at this site.

Seed Mix Design

We generated assemblages of species that met specific trait profiles using the ‘selectSpecies’ function in the R package ‘Select’ (Laughlin and Chalmandrier 2018) parameterized with species-level trait data measured in each grassland (Appendix S2).

For the Wyoming site, drought tolerant communities were generated to exhibit a drought tolerant trait target with high LDMC (T_k = 0.75 quantile), low leaf TLP (T_k = 0.25 quantile) as estimated by leaf osmotic potential (Bartlett et al. 2012), and high diversity of fine root diameter (maximum Rao’s Q of root diameter). Invasion resistant communities were generated to exhibit an invasion resistant trait target with low leaf nitrogen (T_k = 0.25 quantile), high SRL (T_k = 0.75 quantile), and a high diversity of clonal growth strategies (maximum Rao’s Q of clonality).

For the California site, drought tolerant communities were generated to exhibit a drought tolerant trait target with high LMA (T_k = 0.67 quantile), high seed mass (T_k = 0.67 quantile), and high diversity of fine root diameter (maximum Rao’s Q of root diameter). Invasion resistant communities were generated to exhibit an invasion resistant trait target with low leaf nitrogen (T_k = 0.33 quantile), high SRL (T_k = 0.67 quantile), and low RMF (T_k = 0.33 quantile). Trait quantiles in California were less extreme than in Wyoming in order to obtain feasible solution sets because it improved the chance that the set of 10 randomly drawn species from the species pool could achieve the target T_k.

Functionally diverse assemblages were generated by maximizing the functional diversity (Rao's Q) across all six traits used in each site.

Random (null) communities were generated independent of traits to serve as controls to test whether drought tolerant, invasion resistant, or functionally diverse communities performed significantly better than randomly assembled communities. The ‘rsad’ function from the R package ‘sads’ (Prado et al. 2018) was used to randomly generate log normal species abundance distributions (Harpole and Tilman 2006, Ulrich et al. 2010).

Seed mixes were designed to generate communities as close as possible to the model-derived relative abundances by incorporating estimates of pure live seed (PLS) and germination into seeding rates (Figure 1, Step 4; see details in Appendix S1: Section 1: Methods).For the Wyoming site, drought tolerant communities were generated to exhibit a drought tolerant trait target with high LDMC (T_k = 0.75 quantile), low leaf TLP (T_k = 0.25 quantile) as estimated by leaf osmotic potential (Bartlett et al. 2012), and high diversity of fine root diameter (maximum Rao’s Q of root diameter). Invasion resistant communities were generated to exhibit an invasion resistant trait target with low leaf nitrogen (T_k = 0.25 quantile), high SRL (T_k = 0.75 quantile), and a high diversity of clonal growth strategies (maximum Rao’s Q of clonality).

For the California site, drought tolerant communities were generated to exhibit a drought tolerant trait target with high LMA (T_k = 0.67 quantile), high seed mass (T_k = 0.67 quantile), and high diversity of fine root diameter (maximum Rao’s Q of root diameter). Invasion resistant communities were generated to exhibit an invasion resistant trait target with low leaf nitrogen (T_k = 0.33 quantile), high SRL (T_k = 0.67 quantile), and low RMF (T_k = 0.33 quantile). Trait quantiles in California were less extreme than in Wyoming in order to obtain feasible solution sets because it improved the chance that the set of 10 randomly drawn species from the species pool could achieve the target T_k.

Functionally diverse assemblages were generated by maximizing the functional diversity (Rao's Q) across all six traits used in each site.

Random (null) communities were generated independent of traits to serve as controls to test whether drought tolerant, invasion resistant, or functionally diverse communities performed significantly better than randomly assembled communities. The ‘rsad’ function from the R package ‘sads’ (Prado et al. 2018) was used to randomly generate log normal species abundance distributions (Harpole and Tilman 2006, Ulrich et al. 2010).

Seed mixes were designed to generate communities as close as possible to the model-derived relative abundances by incorporating estimates of pure live seed (PLS) and germination into seeding rates (Figure 1, Step 4; see details in Appendix S1: Section 1: Methods).

Wyoming perennial grassland – Seed mixes were broadcast by hand on to each plot and then raked into the soil at the start of the 2019 growing season. We measured the species composition of each plot in August 2019, and re-seeded species with poor establishment in November 2019 to achieve relative species distributions that were as close as possible to the targets. In the second growing season (2020), we trenched around each block to 15 cm deep and installed thick black plastic sheeting to prevent lateral movement of subsurface water. All plots were irrigated and weeded throughout the first (2019) and second (2020) growing season (Appendix S1: Section 1: Methods). We considered communities to be established by the 2021 growing season. During the peak growing season (July), absolute foliar cover was estimated for every species (precision to the nearest 1 % for cover up to 20 %, to the nearest 5 % up for cover 50 %, and to the nearest 10 % for cover > 50 %) in each of the 256 plots in Wyoming annually between 2021 and 2023.

California annual grassland – Plots were initially seeded in December 2019. The top 2.5 cm of soil was removed from each plot, seeds were broadcast by hand, and new soil from the site was laid over the seeds. In mid-February of the first growing season (2019), species composition for each plot was observed, and additional seeds were broadcast, or individual seedlings were hand removed to achieve relative species distributions as close as possible to our target composition. Plots were hand weeded biweekly throughout the first growing season, and irrigated throughout the first (2019) and second (2020) growing season (Appendix S1: Section 1: Methods). We considered communities to be established by the 2021 growing season During peak growing season (March), we visually estimated absolute foliar cover for every species (precision to the nearest 1 % for cover up to 10 % and to the nearest 5 % for cover > 10 %) in each of 180 plots in California annually between 2021 and 2023.

Data Analyses

All analyses were conducted in R version 4.3.3 (R Core Team 2024). After normalizing all six population-level traits within each site (LDMC, TLP, leaf nitrogen, SRL, root diameter, and clonality for Wyoming; leaf nitrogen, LMA, SRL, root diameter, and seed mass for California), we calculated the CWM and diversity (Rao’s quadratic entropy) for each trait, in each community, in each year (2021-2023). We used the ‘functcomp()’ and ‘dbFD()’ in the FD package (Laliberté and Legendre 2010) to calculate CWM’s and functional diversity, respectively. In order to calculate valid community-level trait values for only the native portion of the communities, we excluded plots for which 20 % or more of the community was composed of invasive species (across all years, 19 % of observations were removed in WY and 15 % were removed in CA).

To answer our first question (Q1: Can we establish communities from seed that meet and maintain a desired trait-based target?), we computed the Euclidean distance between the traits of the realized communities and the desired restoration targets, where smaller distances were more successful. We did this for each community in each year (2021-2023). Rather than using the trait quantiles (i.e., 0.25 or 0.75) to define the restoration targets, we used the ranges (i.e., minimum and maximum, respectively) of the traits because our goal would be met if we achieved community-level trait values that were above our upper quantiles or below our lower quantiles. We then compared the distances between the CWM in each plot to the target CWM’s for each seeding treatment to the random controls using ANOVA to determine whether trait-based treatments were more similar to their targets than randomly generated communities across the three years of the experiment. These models included fixed effects for precipitation treatments, an interactive effect of seeding treatment and precipitation treatment, and a main effect of year. We quantified the relationship between compositional dissimilarity and functional trait dissimilarity from trait targets for each seeding treatment across years and precipitation treatments using a general linear model to assess if changes in composition were related to changes in the distance to trait targets as we expected.

To answer our second question (Q2: Did experimental communities tolerate precipitation reduction or invasion by non-native annual grasses?), we used generalized linear mixed effects models to determine the effect of treatments and dissimilarity from trait targets on annual growth rates in each experimental community. We assessed the tolerance of all communities to drought by calculating an annual growth rate for each community as the log ratio of total cover in the current year (t) to total cover in the previous year (t-1) in the reduction and control precipitation treatments:

Annual growth rate = log(cover_t/ cover_t-1) (6)

To test if any of our seeding treatments increased community drought tolerance, we first modelled the annual growth rate in each community as a function of the seeding treatments. To assess the effects of the different realized trait distributions on drought tolerance, we then modelled the annual growth rate in each community as a function of the Euclidean distance to each of the seeding treatment trait targets. All these models included fixed effects for the precipitation treatment and year, and random intercepts for experimental block (Wyoming) or rain manipulation structure (California) to account for spatial variation within each site. We tested including invasion as a binary factor in these models, but there were no significant differences in growth rates as a result of invader addition, nor the interaction of invader addition and drought, and therefore we exclude the invasion treatment in the analysis of growth rates under drought for simplicity. We again excluded the plots for which 20 % or more of the community was composed of invasive species.

To test if any of our seeding treatments increased community invasion resistance, we modelled the relative cover of the non-native annual grasses (B. tectorum in Wyoming and F. perennis plus two naturally recruiting invasive grasses, B. madritensis and B. diandrus, in California) in each invaded community in the final year of the experiment (2023) as a function of the seeding treatments. To assess the effect of different functional trait targets on invasion resistance, we then modelled the relative cover of invasive grass in 2023 as a function of the Euclidean distance to each of the seeding treatment trait targets. In Wyoming, we assessed any plots with B. tectorum cover in 2023 for growth of invasive grasses (N = 190). In California, we assessed any plots with invasive annual grass cover in 2023 for growth of the invasive grasses (N = 90). These models included fixed effects for precipitation treatment and random intercepts for experimental block (Wyoming) or structure (California) to account for site-level spatial variation. We did not exclude any plots from this analysis based on the proportion of non-native species since we also want to assess the distance to our trait targets in communities that succumbed to invasion.