Applying ecological site concepts and state‐and‐transition models to a grazed riparian rangeland

Abstract Ecological sites and state‐and‐transition models are useful tools for generating and testing hypotheses about drivers of vegetation composition in rangeland systems. These models have been widely implemented in upland rangelands, but comparatively, little attention has been given to developing ecological site concepts for rangeland riparian areas, and additional environmental criteria may be necessary to classify riparian ecological sites. Between 2013 and 2016, fifteen study reaches on five creeks were studied at Tejon Ranch in southern California. Data were collected to describe the relationship between riparian vegetation composition, environmental variables, and livestock management; and to explore the utility of ecological sites and state‐and‐transition models for describing riparian vegetation communities and for creating hypotheses about drivers of vegetation change. Hierarchical cluster analysis was used to classify the environmental and vegetation data (15 stream reaches × 4 years) into two ecological sites and eight community phases that comprised three vegetation states. Classification and regression tree (CART) analysis was used to determine the influence of abiotic site variables, annual precipitation, and cattle activity on vegetation clusters. Channel slope explained the greatest amount of variation in vegetation clusters; however, soil texture, geology, watershed size, and elevation were also selected as important predictors of vegetation composition. The classification tree built with this limited set of abiotic predictor variables explained 90% of the observed vegetation clusters. Cattle grazing and annual precipitation were not linked to qualitative differences in vegetation. Abiotic variables explained almost all of the observed riparian vegetation dynamics—and the divisions in the CART analysis corresponded roughly to the ecological sites—suggesting that ecological sites are well‐suited for understanding and predicting change in this highly variable system. These findings support continued development of riparian ecological site concepts and state‐and‐transition models to aid decision making for conservation and management of rangeland riparian areas.


| INTRODUC TI ON
Riparian areas threading through upland rangelands boost landscape-level biodiversity (Sabo et al., 2005), filter water (Tate, Atwill, Bartolome, & Nader, 2006), and provide other valuable ecosystem services (George, Jackson, Boyd, & Tate, 2011). They also provide forage and water for livestock, which tends to congregate these areas, potentially degrading riparian resources (Belsky, Matzke, & Uselman, 1999;Kauffman & Krueger, 1984). Accordingly, efforts to improve the outcomes of riparian management are common, but the highly variable and site-specific responses of rangeland riparian zones can complicate managers' ability to make reliable predictions about the effects of management (George et al., 2011).
Ecological site descriptions and state-and-transition models are currently regarded as useful organizing frameworks for understanding and predicting the patterns and processes on rangelands (Spiegal et al., 2016;Sayre, 2017). These models have been extensively developed for upland rangelands in the United States, but only recently has attention been given to developing them for riparian systems (Stringham & Repp, 2010).
Major determinants of riparian rangeland vegetation composition include fluvial processes and their controls on channel geomorphology (McBride & Strahan, 1984;Stella, Battles, McBride, & Orr, 2010), depth to water table and soil moisture dynamics (Stringham, Krueger, & Thomas, 2001), inundation frequency (Sankey, Ralston, Grams, Schmidt, & Cagney, 2015), annual fluctuations in precipitation (Lunt, Jansen, Binns, & Kenny, 2007), and flood disturbance regimes (Campbell & Green, 1968). As a result of frequent disturbances and spatial heterogeneity, vegetation will likely never reach "climax" stages (Campbell & Green, 1968), and biotic drivers such as cattle grazing may have limited effects on vegetation composition (Lunt et al., 2007). Nevertheless, cattle tend to congregate in riparian areas and can have exaggerated effects on these systems (Kauffman & Krueger, 1984), and managers need models that consider the role of abiotic disturbances, livestock management, and site potential.
Rangelands in Mediterranean-type climates, which are predictably mesic in the winter and xeric in the summer, have distinct flora, fauna and unique management systems, conservation challenges, and threats Perevolotsky & Seligman, 1998).
Riparian systems in Mediterranean-type regions have high interannual and intra-annual weather variation coupled with a "flashy" hydrology produced during the relatively short wet season, creating periodic fluvial disturbances and drought which structure biological communities (Gasith & Resh, 1999). Models that consider these abiotic perturbations may be necessary to describe vegetation dynamics in Mediterranean-type riparian systems.
State-and-transition models are usually represented by box and arrow diagrams and descriptive text that catalogs all the known vegetation states (boxes) and transitions between states (arrows) for a given site. They were developed to model nonlinear vegetation dynamics in rangeland systems (Westoby, Walker, & Noy-Meir, 1989) and are a useful tool for communicating vegetation dynamics to managers. Ecological sites describe divisions of the landscape with similar environmental characteristics that support the same range of states and transitions (Spiegal, Larios, Bartolome, & Suding, 2014).
Given the variable nature of rangeland riparian sites, ecological site descriptions and state-and-transition models may be the optimal framework for cataloguing and making predictions about their ecology (Stringham & Repp, 2010)-but more information is needed about how to best classify ecological sites, states, and phases for rangeland riparian areas.
Upland sites are largely classified based on soils, climate, and landscape position, which are relatively stable over timescales relevant to management (Caudle, DiBenedetto, Karl, Sanchez, & Talbot, 2013). These factors are probably not sufficient to describe differences in rangeland riparian sites, because riparian sites are also influenced by differences in fluvial processes, channel geomorphology, and hydrologic cycles between sites (Stringham & Repp, 2010).
Processes governing temporal variation within riparian ecological sites differ somewhat from those in uplands as well. In addition to climatic and management drivers associated with interannual variation in uplands, fluvial processes and soil-water characteristics may drive temporal variation in vegetation composition (Stringham & Repp, 2010;Stringham et al., 2001). Linking characteristics of channel geomorphology, soils, and hydrologic properties to differences in riparian vegetation states is necessary to help pair riparian ecological site descriptions with state-and-transition models.
Given their value to conservation and management, it is important to understand riparian systems in rangelands so that their management can be improved. This study addresses the following research questions:  (Vaghti & Greco, 2007); due to its extent and management history, Tejon Ranch provides an ideal location to study a relatively intact network of Central Valley riparian forests.
Five creek segments were selected for study within the area of interest: Chanac Creek (CH), El Paso Creek (EP), Lower Tejon Creek (LT), Tunis Creek (TU), and Upper Tejon Creek (UT)-hereafter referred to as "creek segments." Within each of these creek segments, three locations were selected randomly within areas with woody vegetation for a total of 15 study reaches-hereafter referred to as "study reaches" (Figure 1). In the winter of 2014-2015, one study reach on each stream segment was randomly chosen to receive a cattle exclosure. The exclosures were in place for the remainder of the study. Reaches that received exclosures were CH2, EP3, LT1, TU2, and UT3.
Although somewhat drier than the "true" Mediterranean climate (Aschmann, 1984), the study area is in a Mediterranean-

| Sampling abiotic site factors
Fourteen variables were used to classify the 15 study reaches into ecological sites (Table 1). These include remotely sensed values measured in a geographical information system, such as elevation, slope, sinuosity, watershed size, and geology ( Table 1). Measurements of stream geomorphology were made in the field using a total station. Soil samples were collected in the field in 2013 and analyzed for soil texture using the hygrometer method at the UC Davis Analytical Laboratory. the first plant hit within one meter above the ground was recorded. A line-intercept transect was used to record the linear distance of shrubs and trees overhanging the greenline transect. Any plant (regardless of species) overhanging the tape between one and 3 m of height was recorded in the "shrub" category, and any plant overhanging the tape above 3 m in height was recorded in the "tree" category.

| Statistical methods
Our analytical approach proceeded as follows: (1)

| Ecological site cluster analysis
Hierarchical cluster analysis can be used to classify ecological sites based on groupings of key abiotic environmental variables in rangeland uplands (Spiegal et al., 2014). Similarly, it has been used to classify stream reaches from geomorphic and hydrologic measurements of stream channels and is especially useful if applied within a distinct physiographic unit where it can yield objective classifications (Kondolf, Montgomery, Piegay, & Schmitt, 2003). A suite of indicators used in stream classification and those used in upland ecological site classification were combined in a cluster analysis to create the riparian ecological site classification (Table 1).
The ecological site cluster analysis was performed using Gower's distance, which calculates similarity for each variable in the matrix separately (using a method according to the variable type) and is therefore able to analyze both continuous and categorical variables together. The final distance metric is an average of the partial similarities (Borcard, Gillet, & Legendre, 2011). Analysis was performed in R using the packages "vegan" and "cluster" (Maechler,

| Vegetation cluster analysis
A cluster analysis was performed on the greenline vegetation cover data to investigate patterns of riparian plant community structure within the 15 study reaches over 4 years. The 60 unique Reach × Year combinations were clustered based on absolute cover of all live plants along transects. Proportional cover data from the shrub and tree layers were generally much higher than data from the herbaceous layer; therefore, herbaceous layer data were square-root transformed so that the shrub and tree layers did not overly influence the cluster assignments (McCune, Grace, & Urban, 2002). Shrub and Tree cover was not transformed. Similarly, all species occurring on <2 Reach × Years were removed from the analysis so that very rare species did not disproportionately influence the analysis. All species × canopy class combinations were treated as unique species. The cluster analysis was performed using Bray-Curtis distance, which calculates similarity based on species found to be present on study reaches rather than mutual absences (Zuur, Ieno, & Smith, 2007

| Indicator species analysis
In addition to showing the optimal location to prune the cluster dendrogram, the "significant" indicator species show which species best characterize each cluster. Indicator species are those that are common within study reaches of one cluster, and relatively scarce in study reaches of other clusters (Dufrene & Legendre, 1997). Based on these criteria, species are given an indicator value (0-1), and a randomization test is performed to determine the statistical significance of the indicator value. "Significant" indicator species are those with <5% probability of having no difference between groups.

| Identifying states, phases, and transitions
The USDA Natural Resources Conservation Service (NRCS) implementation of state-and-transition models and ecological site descriptions is largely a "top-down" process where elements of state-and-transition models are drawn and populated by expert opinion and afterward validated with data (Jackson, Bartolome, & Allen-Diaz, 2002). In contrast, in this study, plant species data are aggregated to build vegetation states from the "ground-up" with fewer preconceptions about what constitutes a "state." The NRCS also differentiates between minor, easily reversible changes in vegetation labeled "phase-shifts" and the more resilient "states" they occur in (Bestelmeyer et al., 2003;Stringham, Krueger, & Shaver, 2003).
These distinctions formalize some general aspects of the original state-and-transition approach and provide useful categories that can be the basis of testable hypotheses.
In this study, we performed cluster analysis to define meaningful vegetation assemblages. Many of these clusters had similar vegetation structure and functional group composition-thus having similar implications for management-and transitions between some of these clusters would likely occur without threshold dynamics. As a result, the optimal number of clusters from the vegetation cluster analysis was considered vegetation "phases," not "states." The more general "states" were defined by considering: potential drivers of spatial and temporal variation (e.g., irreversible geomorphological changes), differences in Bray-Curtis distance between the clusters, and ecological characteristics of the dominant and indicator species of each cluster (e.g., wetland vs. upland plants). The resulting states are still based on the original vegetation cluster dendrogram, but represent a deeper "cut" of the dendrogram with fewer terminal nodes.
The vegetation cluster analysis was performed on data from all 15 study reaches, and the resulting states and phases were subsequently divided into the two ecological sites. This procedure was chosen because (1) it allowed evaluation of how well the ecological sites corresponded to observed differences in vegetation dynamics, and (2) although study reaches are represented by discreet ecological sites, they represent a gradient of site characteristics and are therefore expected to share some vegetation states. Combining data from all study plots showed which states are unique to each ecological site, and which are shared between them.
In our scheme, a "temporal transition" occurs when the state at a study reach moves in species cluster space between years (sensu Spiegal et al., 2014). "Spatial transitions" are evident in cases in which different vegetation clusters occur in different areas within the same ecological site and are differentiated by spatial-instead of inherently temporal-processes (also see Bestelmeyer, Goolsby, & Archer, 2011).

| Classification Tree (CART)
A classification tree was built to determine which environmental factors best predict the observed vegetation states and to inform our ecological site classification approach. The response variable (the data to be partitioned) was the clusters from the vegetation cluster analysis, and the independent variables were the environmental variables used in the ecological site cluster analysis, annual precipitation, and grazing treatments (exclosures). A classification tree uses top-down recursive binary splitting to partition the response data into a tree that optimizes the classification of response variables at each node with respect to each of the predictor variables (James, Daniela, Trevor, & Robert, 2013).
The classification tree was built using the "tree" package in R (Ripley, 2016). The tree was pruned using the function "cv.tree," which determines the optimal number of terminal nodes by minimizing the deviance in a K-fold cross-validation (Ripley, 2016). Pruning the CART tree to seven terminal nodes resulted in the lowest deviance in the CART analysis. This resulted in only six of the fourteen abiotic factors being included in the construction of the classification tree (Table 2).

| Ecological sites
The Mantel correlation test showed that the optimal number of clusters was 2 (r = .644), representing two ecological sites: Lower Tejon Creek and all other study reaches ( Figure 2). The r value for the next highest correlation (for five clusters) was substantially lower at r = .572.
These two ecological sites differ in several regards. Ecological Site 1 is more widespread in the study area and as a result is more variable. Reaches in Ecological Site 1 (all study reaches except those on Lower Tejon Creek) have higher elevations, higher channel slopes, smaller watershed sizes, lower entrenchment ratios, more silt and less sand in the soil, and more diverse geologies and upstream geologies than reaches in Ecological Site 2 (those on Lower Tejon Creek). The variables that do not substantially differ between the two ecological sites are sinuosity, width:depth ratio, greenline height above thalweg, and percent clay in soil (Table 2).

| Vegetation states
The vegetation cluster analysis showed that Reach × Years generally clustered most closely with the same reach in other years. The Mantel correlation test pruned the resulting dendrogram to 10 clusters (r = .663). However, eight clusters had the most significant indicator species p-values and were therefore selected by indicator species analysis. As the Mantel correlation coefficient was very close between eight and ten clusters, (r = .645 and r = .663 respectively), eight clusters were selected to represent the vegetation groups ( Figure 3).
Each of the eight clusters has statistically significant indicator species. All clusters include perennial woody species as indicators, and all clusters except Clusters 1 and 2 include herbaceous species as significant indicator species (Figure 4). Indicator species are always most abundant in the cluster they are assigned to; however, in Ecological site clusters --Gower distace this analysis, they generally occur in other clusters as well, so their mere presence is not diagnostic of cluster membership. Just five of the 41 indicator species occurred in only one cluster, and five occurred in all eight clusters. Thirty-five of the 41 indicator species were in the top five species (by cover) for their canopy layer in the cluster they belonged to.
Per methods described previously, three vegetation "states" were defined among the eight vegetation clusters (Figure 3). States 1 and 2 occur exclusively in Ecological Site 1, while State 3 occurs almost exclusively in Ecological Site 2, but has a limited distribution on Ecological Site 1 (Figure 3). These three states represent a deeper "cut" of the dendrogram and also have a high Mantel correlation value (r = .61). The three states are:

| Vegetation state 1
This state comprises four of the eight vegetation clusters

| Vegetation state 2
Although most of the study reaches classified as Ecological Site 1 are clustered in a relatively cohesive area of the vegetation cluster

| Vegetation state 3
In Ecological Site 2, there are three closely related vegetation clusters (clusters 2, 4, and 7). All three share a common branch of the cluster dendrogram ( Figure 3)  cluster dendrogram, these three vegetation clusters are all considered phases in Vegetation State 3 (Figure 4).

| Transitions and phase shifts
Spatial variation within each of the ecological sites was more pronounced than temporal change over the study period. In total, seven community phases (i.e., the vegetation clusters) comprising three vegetation states were observed across the reaches in Ecological Site 1, and three community phases were observed across the reaches in Ecological Site 2 (Figure 4). Of all the potential "spatial transitions," compelling evidence only exists for the cause of one transition in Ecological Site 1 between Vegetation State 2 and Vegetation State 1 (T2, Figure 4). In Ecological Site 1, one minor "temporal" phase shift and one more significant "temporal" transition were also observed over the 4 years of the study; and only one phase shift was observed on reaches in Ecological Site 2. A summary of these transitions and phase shifts is below: F I G U R E 4 State-and-transition diagram for the riparian study reaches. The top diagram shows the states and phases occurring on study reaches in Ecological Site 1, and the bottom diagram shows state and phases occurring on reaches in Ecological Site 2. Species listed in each phase are the significant indicator species for that phase, listed by descending order of indicator value. Solid arrows indicate "temporal transitions" and phase shifts, the dotted arrow shows the only "spatial transition" with a plausible driver. Wetland codes are provided in parentheses after each species name (Lichvar, Banks, Kirchner, & Melvin, 2016). An * indicates that the species is not included in "The National Wetland Plant List" (Lichvar et al., 2016). The wetland status of species with one * is inferred from congeners on the list. Species with two ** do not have congeners on the list, and their wetland status is hypothesized from authors' field observations. More information on plant species is included in Table S1. Descriptions of the states and transitions are in the text of the Results section

| Transition from vegetation state 3 to vegetation state 2 (temporally-observed) (T1-Ecological Site 1)
In Ecological Site 1, the study reach CH2 changed from Vegetation  Figure S1). Reversing this transition may require several years of wet conditions to reestablish these tree species.

| Phase shift from vegetation cluster 1 to 6 (temporally-observed) (PS1-Ecological Site 1)
In Ecological Site 1, the study reach EP3 changed from Vegetation Cluster 1 to Vegetation Cluster 6 between 2015 and 2016 sampling.
This represented a phase shift from a community dominated by Vitis californica and Salix laevigata in the herbaceous layer to one characterized by a suite of herbaceous hydrophilic plants ( Figure 4). This phase shift followed an unusual summer flood in 2015 that cleared out some of the woody plant understory.

| Phase shift from vegetation cluster 7 to 2 (temporally-observed) (PS1-Ecological Site 2)
The only phase shift observed in Ecological Site 2 was when the study reach LT2 changed from Vegetation Cluster 7 to Vegetation

| Results of CART analysis
The root split in the classification tree was channel slope (500 m), indicating that it explained the most variation in vegetation phases.
After that, a combination of soil texture, geology, watershed size, and elevation were the factors chosen to further partition the cluster assignments. The reach-scale stream geomorphological measurements, cattle exclosures, and annual precipitation were not included in the pruned classification tree, indicating that they did not consistently predict the different vegetation clusters (Figure 6). Overall, the pruned CART model correctly classified 90% of the Reach × Years, with only six Reach × Years misclassified.

| D ISCUSS I ON
The distribution of the vegetation states and phases was largely explained by the two ecological sites (Figures 3 and 4). Similarly, phases and states from each of the ecological sites occurred largely This implies that none of these geomorphological variables consistently predicted differences in the vegetation phases. Stream cross-sectional profiles differed at large spatial scales, as seen in the differences between ecological sites (Table 2), but they also varied at relatively small spatial scales throughout the study area (e.g., between study reaches on a creek segment). Channel geomorphology was only measured once at each reach during the study; however, it was considered relatively stable over the study period because of the below-average rainfall.
Cattle exclosure and precipitation were also not significant variables in the CART analysis. This makes sense given that (1) cattle exclosures were only in place for two years; (2) rainfall was not highly variable over the study period; and (3) cluster indicator species contained many perennial woody species. The apparent lack of influence from cattle grazing raises important questions for management of this system, including: Creek affect the rate at which the head cut in that stream F I G U R E 6 Results from a classification tree with vegetation clusters as the categorical response variable, and the abiotic Ecological Site variables, total annual precipitation, and cattle exclosures as the factors used to split the data. The splits farther up in the tree explain more of the overall variation in vegetation Could management strategies-such as seasonal grazing regimes, bank stabilization using restoration planting, or moderating peak stream discharge using an existing dam-slow or stop the movement of the head cut and prevent vegetation from transitioning from State 2 to State 1?
These questions can be formulated as formal hypotheses and tested through longer-term monitoring of exclosures or riparian pastures with prescribed stocking rates. The ecological sites and vegetation states/phases identified in this study provide ecological context that can guide managers' selection of study locations, treatments, and monitoring methods to efficiently answer these questions.

| CON CLUS IONS
By including riparian-specific criteria, ecological site classifications can be built for riparian systems. On Tejon Ranch, riparian ecological site descriptions and state-and-transition models provided a unified framework linking abiotic and management factors to vegetation dynamics. These models were able to incorporate and organize highly variable riparian site factors and vegetation assemblages. By cataloging known phases, states, and transitions on each ecological site, these models created an organized approach to understanding the complex and site-specific responses of rangeland riparian areas. They provided a framework for predicting vegetation states and transitions, and for generating and testing hypotheses linking weather, management, and site characteristics to vegetation changes over time and space.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R CO NTR I B UTI O N
FR, JB, SS, and MW developed the ideas and methodology; FR, JB, SS, and MW collected the data; FR analyzed the data, with contributions from JB, LM, SS, and MW; FR led the writing of the manuscript with substantial contributions from JB, LM, SS, and MW. All authors gave final approval for publication.

DATA ACCE SS I B I LIT Y
Data will be made available in the Dryad Digital Repository.