Divergent climate impacts on C3 versus C4 grasses imply widespread 21st century shifts in grassland functional composition

Grasslands cover a third of Earth's landmass and provide critical ecosystem services. Anticipating how perennial C3 (cool‐season) and C4 (warm‐season) grasses respond to climate change will be key to predicting future composition and functioning of grasslands. Here, we evaluate environmental drivers of C3 and C4 perennial distributions and assess how C3 and C4 grass distributions shift in response to future climate change.

abundance declined across 74% of areas, while C 4 abundance increased across 66% of areas. C 3 grasses expanded in mid-to higher-latitude areas with increasing temperature and decreasing seasonality of precipitation. In contrast, C 4 grasses increased in higher-latitude regions, but declined in lower-latitude, dryer regions. Results were surprisingly robust across climate scenarios, suggesting high confidence in the direction of these future changes.
Main Conclusions: Findings imply C 3 and C 4 perennial grasses will have highly divergent responses to climate change that may result in grassland functional compositional changes. Increasing temperatures and precipitation variability may favour some C 4 grasses, but C 4 habitat expansion may be constrained by soil conditions in western USA. Results provide actionable insights for anticipating the impacts of climate change on grass-dominated and co-dominated ecosystems and improving large-scale conservation and restoration efforts.

| INTRODUC TI ON
Predicting how ecosystems will respond to climate change remains a major challenge. Grasslands cover nearly 40% of Earth's landmass, provide critical ecosystem services including biodiversity, forage for livestock and wildlife, and carbon sequestration, and support the livelihoods of more than 1 billion people worldwide (Asner et al., 2004;Bengtsson et al., 2019). In recent decades, grasslands including those in arid and semiarid regions of the western United States have experienced significant warming and drying (Diffenbaugh et al., 2008;Pachauri et al., 2015;Seager et al., 2007).
Dry grasslands are expected to be particularly susceptible to climate change since they contain relatively shallow-rooted species whose productivity is highly sensitive to climatic variation (Knapp & Smith, 2001;Throop et al., 2012).
Rising temperatures and altered precipitation patterns associated with global climate change can have major consequences for perennial grass productivity and mortality and has been linked to regional grassland die-offs (McAuliffe & Hamerlynck, 2010;Munson et al., 2013). In response to climate disruption, grass species distributions may shift as regional environmental conditions become mismatched with species niche requirements. Within grassland plant communities, shifting species habitat suitability can result in species reordering or loss, and ultimately, altered long-term community composition (Jones et al., 2017). Further, habitat suitability declines across multiple species may also precipitate overall losses of plant cover with consequences for broader ecosystem functioning. For example, perennial grass cover losses can trigger ecological regime shifts, including grassland to shrubland (Báez et al., 2013;Briggs et al., 2005), perennial-to-annual dominance (Koteen, 2009) and/or native to invasive grass state transitions (Abatzoglou & Kolden, 2011;Munson et al., 2013). These state transitions are often accompanied by ecosystem service losses including decreased forage availability for livestock and wildlife (Izaurralde et al., 2011), declines in soil carbon and nutrient availability (Schlesinger et al., 1995) and increased bare ground cover, soil erosion rates and dust storm activity (Munson et al., 2011;Schwinning et al., 2008). While perennial grass declines have been reported worldwide in response to climate change, our ability to predict the relative sensitivities of grass species and functional groups remains limited.
Perennial grasslands are comprised of C 3 (cool-season) and/or C 4 (warm-season) grass species. Understanding how these plant functional types will respond to climate change is important for anticipating the future structure and function of grasslands. C 3 and C 4 species have different environmental niche requirements, which have created distinct geographic distributions for these functional types (Epstein et al., 1997) and will likely influence their responses to climate change (Sage & Kubien, 2007;Winslow et al., 2003). The C 4 photosynthetic pathway is a morphological and biochemical trait complex that confers a higher optimal temperature for photosynthesis and greater water-use efficiency relative to C 3 species (Sage & Kubien, 2007). These traits often yield a strong competitive advantage over the C 3 pathway in warm, drought-prone regions (Pau et al., 2013). However, since the C 4 pathway is more energetically expensive than the C 3 pathway, C 4 species have historically been competitively dominant only in hotter, high-light environments, while C 3 species dominate in cooler habitats and/or earlier periods in the annual growing season (Kubien & Sage, 2004;Pau et al., 2013).
Given these differences, in the context of climate change, we might expect C 3 species with cooler niche requirements to be more sensitive, while C 4 species might respond neutrally (Avendaño-González & Siqueiros-Delgado, 2021) or favourably (e.g. Palmquist et al., 2021) under future climate. Though, recent work highlights that grassland responses to climate change may not be predictable from key traits of C 3 and C 4 species alone (Knapp et al., 2020). Further, observations of declines in some C 4 grasses following aridification and/or drought (Knapp et al., 2020;Ludwig et al., 2017;Rondeau et al., 2018) have generated considerable uncertainty about how grasslands may respond to changing climate.
Anticipating responses of C 3 and C 4 grasses to climate change has a variety of practical implications for grassland management at local, regional and/or continental scales. For example, since C 3 species generally have higher forage quality relative to C 4 species, shifts in C 3 -C 4 dominance could result in changes in forage quality for wildlife and livestock (Barbehenn et al., 2004). Additionally, since cool-season (C 3 ) and warm-season (C 4 ) grasses occupy different phenological niches, changes in C 3 -C 4 abundance could result in to temporal mismatches between seasonal forage availability and wildlife and/or livestock grazing activity (Chamaillé-Jammes & Bond, 2010). Predicting changes in habitat suitability for C 3 and C 4 grasses in space can help land managers with conservation planning including identification of habitat refugia and/or assisted migration of priority grass species (Vitt et al., 2010). Similarly, this knowledge could aid restoration practitioners in selecting more resilient species in revegetation efforts in grass-dominated and co-dominate ecosystems in the context of global change Doherty et al., 2017).
To assess long-term impacts of climate change on the distributions of C 3 and C 4 perennial grass species in the western USA, we fit integrated species distribution models (SDMs) to identify historical environmental drivers of relative densities and predict how climate change will alter the distributions of for 11 high-priority native, perennial grass species common across many grasslands in the western USA. Specifically, we fit integrated SDMs for each species to (1) evaluate climate and soil drivers of C 3 and C 4 perennial grass biogeography, (2) assess how C 3 and C 4 grass relative abundances shift in response to future climate change and (3) characterize K E Y W O R D S C 3 grass, C 4 grass, climate change, grassland, habitat suitability, perennial grass, species distribution modelling (SDM) climate-driven uncertainty in projections by evaluating when and where shifts in relative abundance are robust across climate models.

| Focal C3 and C4 perennial grasses
We focused on 11 native C 3 and C 4 perennial grass species (Table 1).
These species were chosen based on their ecological and cultural importance and because they are among the most abundant grass species in grasslands in the western USA. Five of the species have a C 3 photosynthetic pathway, while six have a C 4 pathway. For each species, we gathered presence-only (PO) and presence-absence (PA) records and analysed these data relative to environmental drivers in an integrated species distribution model.

| Species occurrence and absence records
Species occurrence records, or presence only (PO) data, for the 11 grass species were extracted for the western USA (latitude = 25° N to 50° N, longitude = 130° W to 100° W) from the SEINet data portal (https://swbio diver sity.org/seine t/). Only records collected since 1980 that were reliably georeferenced (i.e. those records with a GPS georeferencing source, coordinate uncertainty <2000 m or verbatim coordinates) were retained. Species presence-absence (PA) data were acquired from the 2020 United States Bureau of Land Management Assessment, Inventory, and Monitoring (BLM AIM) Terrestrial AIM Database (TerrADat) species inventory dataset; these data were collected using the line-point intercept method (Herrick et al., 2005) at 15,568 sites situated on BLM-managed lands throughout the western USA. For Achnatherum hymenoides, Poa secunda, Pascopyrum smithii and Bouteloua gracilis, a subset of BLM AIM sites (n = 14,556) was used. For these species, PA records within the state of Wyoming were removed because the dataset reported zero occurrences of these species even though they are common species in this region (USDA-NRCS, 2022). Complete maps of species' PO and PA records can be found in the (Figures S1-S11).

| Climate data
Gridded weather data were used to represent historical climate conditions , downloaded from the WorldClim2 version 2.0 database (Fick & Hijmans, 2017) at a resolution of 30 arc-seconds, or ~900 m at the equator. We considered a broad array of climatic covariates for our models (see Table S1) and reduced covariates prior to model fitting to avoid collinearity (avoided pairs with |r| > 0.70) and retain variables with recognized importance for plant species niches in the western USA that represent relatively independent dimensions of climate variability (Butterfield & Munson, 2016). Six climatic variables (Table 2, Figure S12) were ultimately selected: mean annual temperature (MAT), mean annual precipitation (MAP), precipitation seasonal timing (PTcorr; the Pearson correlation between monthly mean temperature and precipitation), precipitation seasonality (PSeas; the coefficient of variation in monthly mean precipitation), annual temperature range (TRange; temperature difference between the warmest and coldest month) and summer aridity index (monthly mean precipitation divided by monthly mean potential evapotranspiration (Trabucco & Zomer, 2018) averaged across summer (June-August) months).

| Soil and species sampling effort data
Gridded soil texture (i.e. percentage sand and clay) and depth to bedrock data were acquired from the SoilGrids 1.0 application (ISRIC, 2013) TA B L E 1 Characteristics and presence absence data for the 11 perennial grass species included in our study. Species presence-absence data column includes the final number of records included in SDMs from the BLM AIM TerraDat database after screening and resampled for the western USA at spatial resolution of 250 m using a bilinear resampling function. Soil composition data were averaged over the top 0-100 cm of the soil profile after removing pixels from layers that fell below the total predicted soil depth in that average computation. To account for differences in sampling effort in the PO data, a sampling intensity covariate was estimated from all Poaceae records in SEINET (https://swbio diver sity.org/seine t/). Specifically, the number of records downloaded from the SEINet data portal per grid cell was used as a proxy for sampling intensity. All covariates were scaled prior to inclusion in statistical models.

| Statistical model
We fitted models for each species that estimated relative abundance, λ, values across the study area based on PO and PA records as well as environmental covariates. Estimated abundance is relative because line transect data (our PA data) have a species-specific relationship with overall abundance in a (~900 m) grid cell. Data were combined in an integrated species distribution model derived from Koshkina et al. (2017) by assuming detection probability equal to one during PA surveys. This is a reasonable assumption given that every grass species on the transects that form are PA dataset is identified and our 11 species are well known and easily identified. In our model, the expected abundance, λ i , in a transect in grid cell i, is related to a matrix of covariates, X i , specific to the grid cell according to the following equation: where is an intercept and represents a vector of slopes associated with each covariate in X i . The probability of a count, C, >0 on a transect in grid cell i (i.e. the probability of a species is present, i ) is then given by: which is equivalent to a Bernoulli generalized linear model with a complementary log-log link (McCullagh & Nelder, 1989).
Presence only (PO) data are generated when a species is present in an area, the area is sampled, and an individual of the species is detected in the sampled area (Yackulic et al., 2013). This leads to thinning (loss of records) and frequently sampling bias (i.e. uneven thinning with respect to environmental covariates). In the framework introduced by Koshkina et al. (2017), the degree of thinning, b j , in the j th grid cell, is modelled using a logit-linear link, based on covariates W j : where γ is an intercept and δ represents a vector of slopes associated with each covariate in W j . The expected density of PO data in the j th grid cell is then given by the product of b j and i .

| Model selection
In all models for all species, W included a single covariate, the sampling intensity for Poaceae, to account for biased sampling of PO records. In contrast, we considered a number of covariates and covariate transformations as drivers of variation in relative abundance (i.e. in the matrix X). A priori we assumed that species' relative densities might respond to both linear and quadratic transformations of the three soil and six climate covariates, as well as pairwise interactions among the covariates. These covariates and transformations have 2 90 potential models and necessitated a model selection strategy (Morin et al., 2020). Our model selection strategy depended on an out of sample metric derived for each model structure and consisted of two stages.
To identify a model structure for each species out of the model set that transferred well among regions and thus might be expected to produce more accurate long-term projections, we applied a fourfold cross-validation process to score each potential model structure. Specifically, in each fold, three of four groups were used to fit The best model for each species was then evaluated based on out-of-sample area under the curve (AUC) values using a similar process to fourfold cross-validation process to calculate log-likelihood in which data were fit to three-quarters of the data and predictions were made for the other one-quarter of the data. AUC values were derived independently for the PO and PA data held out in each fold.

| Drivers of present distributions of C 3 and C 4 perennial grass species
Environmental drivers of perennial grass species distributions differed between C 3 versus C 4 groups and among species (Figure 1; Tables S1, S2). Temperature-related covariates were included more often in C 3 versus C 4 species SDMs (Figure 1). Both C 3 and C 4 species had unimodal responses to variation in mean annual temperature (MAT) with C 3 species tending to occupy cooler climatic niches relative to C 4 species ( Precipitation-related covariates were included in a similar proportion of C 3 and C 4 species SDMs overall (Figure 1) Figure 2). The correlation between monthly temperature and (4) Δ = ( future − historic) ∕ (max( historic)) precipitation (PTcorr) was similarly included in most SDMs for C 3 (80%) and C 4 (83%) species. C 3 species had variable responses to PTcorr, while C 4 species habitat suitability tended to increase as PTcorr increased (Figure 2). Summer aridity index was included in more C 4 versus C 3 SDMs (67% and 40% of SDMs, respectively). In C 3 grasses, habitat suitability of H. comata and P. secunda decreased with increasing summer aridity index. C 4 species responses to variation in summer aridity index were more variable. Some species (P. Soil characteristics also played important roles in structuring species niches (Figures 1, 2). Overall, C 4 species were more sensitive to variation in soil texture relative to C 3 species. C 4 several species had relatively narrow soil niche optima and tended to occupy sandier soils (Figure 2). Percentage sand optima for C 3 species was lower (40.3%) relative to C 4 grasses (58.6%). C 4 species habitat suitability increased with increasing soil sand content (B. gracilis and S. cryptandrus; Figure 2) but conversely decreased with increasing soil sand content in C 3 species P. smithii. Soil clay content was an important driver of habitat suitability in all C 4 grasses, although relationships were species specific: Optimal suitability occurred at clay values ranging from 21.1% (A. purpurea) to 42.7% (S. cryptandrus) (Figure 2), while others (S. airoides and B. gracilis) increased exponentially with increasing clay. Only C 3 grass P. secunda was sensitive to clay, decreasing as clay increased. Soil depth was included in 80% of C 3 and 50% of C 4 species SDMs. Both groups had niche optima in deeper relative to shallower soils ( Figure 2). All species distribution models had similar in-sample and out-ofsample AUC, indicating that they were not overfit and should have good transferability (AUC for in-sample, training data mean = 0.87, range = 0.82-0.96; AUC for out-of-sample, test data mean = 0.86, range = 0.73-0.98; Table S2).

| Relative abundance of C 3 grasses declines, while C 4 species expands in response to climate change
The western USA is forecasted projected to experience increased annual temperature, variable changes in annual precipitation and greater summer aridity under RCP 8.5 by the end of the 21st century (2070; Figures S13-S18; Table S3). In response, our models project widespread declines in suitable habitat for C 3 species (4/5 species), but modest habitat increases in C 4 species (4/6 species; Table 3, Figure 3, Table S4). On average, C 3 species decreased across 74% of all land area within the western USA (Δλ = −0.013; Table 3)

| Geographic patterns in future changes C 3 and C 4 species relative abundance
Forecasted changes in grass species abundance varied geographically. C 3 grasses experienced robust declines in abundance across two thirds of the western USA, while most C 4 species experienced moderate abundance increases in higher latitudes, but declines in more arid, low-latitude areas (Figure 4). Declines in C 3 grass relative abundance corresponded to areas experiencing increasing mean annual temperature (MAT) and decreasing precipitation seasonal timing (PTcorr). Most C 3 species experienced habitat losses in southwestern warm deserts (i.e. Mojave, Sonoran, Chihuahuan), the Great Basin and the Great Plains though the magnitude of climate sensitivity varied by species (Figure 3). For example, declines in C 3 grasses H. comata and P. smithii were greatest (Δλ = −0.15 to −0.50) in higher-latitude regions projected to shift from the species' historical niche requirements. These regions include the Northern Great Plains, particularly in parts of Wyoming and Montana (Figure 3) experiencing substantial increases in mean annual temperature (MAT; Figure S14) and annual temperature range (TRange; Figure S17 Figure S15). In contrast to other C 4 species, S. airoides experienced few areas of abundance declines under future climate (Figure 4).

TA B L E 3
Relative abundance (λ) values for the 11 grass species for the western USA for: Historical (1970Historical ( -2000; future median (50th quantile of the ensemble future of 17 GCMs for RCP 8.5, 2070); pixel-wise percentage area where relative abundance is robustly increasing/ decreasing; the mean values for regions of robust increase/decrease; and overall mean robust change in abundance

| C 3 and C 4 grass distributions are driven by temperature, precipitation timing and soil characteristics
Understanding differences in the sensitivities of C 3 and C 4 grasses to climate change will be key to predicting the future structure and function of grassland ecosystems. The structures of the niche models developed here provide useful perspectives on the differences between C 3 and C 4 grass species niches and offer robust predictions of how perennial grasses will respond to changing climate. Results show C 3 species have historically occupied lower temperature habitats and have more variable responses to changes in precipitation (i.e. MAP, PTcorr and PSeas) relative to C 4 grasses. C 4 species occurred in areas with higher temperatures (MAT) and temperature variability (TRange), and in areas where precipitation is more seasonably variable (higher PSeas) and occurs during warm months (higher PTcorr). C 4 species were also more sensitive to variation in soil characteristics than C 3 species (Figure 2). These results support previous findings (Epstein et al., 1997;Paruelo & Lauenroth, 1996;Pau et al., 2013;Still et al., 2003) and expand their inferences by providing species-level variability at high spatial resolution over multiple ecoregions.
Both C 3 and C 4 grasses were more consistently sensitive to variation in temperature (i.e. MAT, TRange) than precipitation (i.e. MAP; Figure 1), suggesting C 3 and C 4 grasses alike may respond strongly to future warming irrespective of future changes in total annual precipitation. This result is consistent with previous studies that have shown temperature to be a strong driver of perennial grass species distributions (Butterfield & Munson, 2016;Gremer et al., 2018;Munson et al., 2012;Teeri & Stowe, 1976). C 3 grasses occupied cooler habitats relative to C 4 species, which favoured drier habitats at lower latitudes (Figure 4), as is consistent with broadly understood trait and biogeographical differences between C 3 and C 4 species distributions (Pau et al., 2013;Winslow et al., 2003), including higher optimal temperature and water-use efficiency in C 4 species on average (Ehleringer, 1978).
Accordingly, our models showed grass distributions were highly sensitive to precipitation seasonality (e.g. PSeas, PTcorr) with these effects differing between C 3 and C 4 groups. These differences may be partially explained by differences in seasonal phenology between these functional groups. Cool-season (C 3 ) grasses have earlier seasonal phenology and may be more likely benefit more from cool-season and stored soil moisture than warm season (C 4 ) species, which are often more dependent on summer rainfall (Skinner et al., 2002). Accordingly, C 3 species H. comata and P. secunda occupied regions with high precipitation seasonality (PSeas), possibly suggesting dependence on short-term temporal concentration of cool season precipitation events. In contrast, that many C 4 species (e.g. A. purpurea, B. eriopoda and S. airoides) occupied habitats with intermediate PSeas values suggests these species may rely on precipitation events spanning multiple seasons. C 4 and most C 3 species habitat was also positively associated with precipitation seasonal timing (PTcorr; Figure 2). This suggests perennial grasses, especially C 4 species, rely to some degree on availability of warm-season precipitation. Spatial trends in projections suggest that this may be particularly true in regions such as in the southwestern USA, where on average 40% of rainfall derives from the summer North American Monsoon (Adams & Comrie, 1997). This agrees with previous work showing perennial grasses in dry regions of the western USA rely heavily on summer monsoonal precipitation for recruitment, growth and survival .
Soil characteristics were also important drivers of C 3 and C 4 species distributions (Figures 1, 2). C 3 and C 4 species alike favoured habitats with relatively deep soils, but overall C 4 species were more sensitive to variation in soil texture ( Figure 1). C 4 species tended to occupy sandier soils relative to C 3 species (Figure 2). This result follows with the inverse texture hypothesis (Noy-Meir, 1973), which posits that plants growing in coarser textured soil experience less water stress than in finer textured soils in arid ecosystems (Sala et al., 1989) and suggests warm-season species may occupy well-drained soils to escape high evaporative zones in drylands. Accordingly, our results also indicate frequent interactions among climatic and edaphic covariates in determining grass species habitat suitability (see Table S2 for full model results). This is unsurprising given that soil texture consistently plays a strong role in determining soil moisture and nutrient cycling dynamics that impact grassland productivity (Anacker et al., 2021;Hook & Burke, 2000). Our results showing how interactions between climate and soil variables drive grass habitat suitability provide a more complete understanding than many analyses that use climate alone and suggest that both need to be factored into understanding grass responses to future climate change.

| C 3 abundance declines in response to climate change
Predicted changes in perennial grass habitat suitability were surprisingly robust to variation in climate projections, suggesting high confidence in the direction of future changes (Table 3, Figure 3).
Projections indicate robust declines in C 3 perennial grass abundance across much of the western USA by the late 21st century. These results align with spatial patterns of future climate and are consistent with observational (Gremer et al., 2015;Munson et al., 2013;Wertin et al., 2015;Winkler et al., 2019) and modelling studies (e.g. Palmquist et al., 2021;Winslow et al., 2003)  temperature with C 3 grass mortality (Wertin et al., 2015;Winkler et al., 2019). Because most C 3 grasses can continue to grow under high temperatures, declines associated with higher temperatures may be indirect via increased evaporative demand that drives down soil moisture (Golluscio et al., 2009;Sala et al., 2012). Among the C 3 grasses, only A. hymenoides deviated from this pattern of overall future decline. This may be explained in part by the fact that A. hymenoides was overall less sensitive to changes in temperature and suited to habitats with more intermediate values of PTcorr relative to other C 3 species ( Figure 2). Furthermore, the current distribution of A. hymenoides is in warmer, more southern latitudes, compared to other C 3 species. Among C 4 grasses, projections of abundance decline in B. gracilis were most spatially widespread (Figure 2). While B. gracilis has a relatively wide climatic niche breath in terms of annual temperature and precipitation, like other C 4 grasses experiencing regional abundance declines, the species also has a relatively narrow soil niche ( Figure 2). As such, sensitivity of B. gracilis to climate change and potential increases may be partially mediated by soil characteristics.

| C 4 abundance expands at higher latitudes but constricts at southern latitudes under future climate
Projected declines in B. gracilis abundance here contrast with a recent niche modelling effort by Avendaño-González and Siqueiros-Delgado (2021) that did not include soil characteristics, suggesting that B. gracilis' distribution will be relatively unaffected by projected climate change. However, declines in abundance for B. gracilis align with recent observations of reductions in cover and frequency of

B. gracilis in arid and semiarid ecosystems in southwestern North
America (Ludwig et al., 2017) and the semiarid short-grass steppe and central Mexico (Rondeau et al., 2018). Our results are also supported by experimental studies that have found large declines in B.

| Benefits of integrated models, study limitations and future research needs
Integrated species distributions are increasingly popular because they provide a means to combine the strengths of different datasets.
The presence-only data in our study was collected from a broader range of environmental conditions, but only provided information about relative densities. In contrast, the presence-absence data are from a more restricted set of environmental conditions but include absence data that allow for a better understanding of absolute presence or absence at the scale of the original surveys. While our analyses using integrated models combined the strengths of these data to reveal robust patterns of decreasing C 3 species and increasing C 4 species habitat suitability under future climate, these analyses have several key limitations. First, our models were developed only for 11 common species across the western USA. Future work could explore the responses of a broader range of species, including species with functional traits that may be suitable for future hotter, drier conditions across the region. Secondly, an important limitation of all niche models is that SDMs can only assess species realized niches but have no way of evaluating their potential niches as influenced by potential interactions between biotic interactions and species disequilibrium (Yackulic, 2017). As such, SDMs may be more prone to predicting extreme climate impacts relative to other modelling approaches (Beaumont et al., 2016). In addition, the SDM model-selection process employed here only included a priori environmental covariates.

| Practical implications for land management in a changing world
Although climate impact assessments have numerous sources of uncertainty, our approach of identifying robust changes in C 3 and C 4 species relative abundance can help resource managers identify regions of future change with relative confidence, providing a rigorous foundation for making long-term strategic resource management decisions. These results can provide actionable insights for land managers seeking to anticipate the impacts of climate change on grass-dominated and co-dominated ecosystems including: 1. Grassland conservation and management -Anticipating future shifts in C 3 and C 4 perennial grass distributions might be useful for conservation planning to preserve grassland connectivity with high-quality wildlife habitat. While C 4 habitat expansion was modest relative to C 3 suitability declines overall, some C 4 grass species (e.g. A. purpurea and S. cryptandrus) experienced relatively strong increases in abundance in some high-latitude regions that are the same locations where many C 3 species experienced their greatest declines (e.g. the Northern Great Plains). This suggests some C 4 species could become more abundant in regions where C 3 species experience declines in response to climate change. Such shifts in C 3 -C 4 abundance could stabilize grassland net primary production (Klemm et al., 2020) but could also have a variety of ecological implications, including altered inputs to soil carbon and nutrients.
2. Effects on seasonal forage availability -Shifting C 3 -C 4 abundance, for example, could alter the seasonality of forage availability, which in turn could alter regional and local wildlife and livestock grazing regimes. Relative to C 3 species, C 4 grasses tend to be of lesser forage quality due to an abundance of bundle sheath cells and lower protein levels (Barbehenn et al., 2004).
Moreover, an overall shift from C 3 to C 4 habitat suitability could coincide with temporal mismatches between forage availability and grazing and habitat needs of large herbivores since C 4 grasses have later phenology relative to C 3 species (Chamaillé-Jammes & Bond, 2010).
3. Opportunities for climate-adaptive management and restoration -There may also be geographic regions where C 3 grasses are projected to decline where C 4 grasses may not currently be abundant enough to compensate for the ecosystem service losses associated with C 3 species losses. These may be areas where assisted migration of suitable C 4 grasses could be helpful to maintain the ecosystem services delivered by perennial grasses (Vitt et al., 2010). At local and regional scales, these results could also be used to strategically identify suitable perennial grasses to use for restoration purposes Doherty et al., 2017). Future work should test and evaluate outcomes of such strategies via monitoring and adaptive management experiments.

ACK N OWLED G EM ENTS
We thank the members of the National Park Service Southeast Utah Group Climate Adaptation Working Group for participation in workshops and discussions that defined this study. We also thank Travis Nauman and Caitlin Andrews for assistance with data acquisition and organization, and Sarah Burnett, data manager for the BLM AIM project team, for providing access to the latest BLM AIM data F I G U R E 4 Projections of relative abundance (λ) for C 4 perennial grass species included in this study (Aristida purpurea, Bouteloua eriopoda, Bouteloua gracilis, Pleuraphis jamesii, Sporobolus airoides, Sporobolus cryptandrus) species included in this study showing: (a) historical  λ; (b) future median projected λ based on an 50th quantile of the ensemble future of 17 GCMs for RCP 8.5, 2070; and (c) areas of robust change in Δλ based on differences between historical and future median abundance. Robust change indicates areas where 90% of GCMs agreed in the direction of change in relative abundance for RCP 8.5, 2070. All areas in panel "c" are robust except for white shaded areas