Faunal assemblages and multi – scale habitat patterns in headwater tributaries of the South Fork Trinity River – an unregulated river embedded within a multiple – use landscape

Faunal assemblages and multi–scale habitat patterns in headwater tributaries of the South Fork Trinity River – an unregulated river embedded within a multiple–use landscape.— Headwaters can represent 80% of stream kilometers in a watershed, and they also have unique physical and biological properties that have only recently been recognized for their importance in sustaining healthy functioning stream networks and their ecological services. We sampled 60 headwater tributaries in the South Fork Trinity River, a 2,430 km2, mostly forested, multiple–use watershed in northwestern California. Our objectives were: (1) to differentiate unique headwater types using 69 abiotic and vegetation variables measured at three spatial scales, and then to reduce these to informative subsets; (2) determine if distinct biota occupied the different tributary types; (3) determine the environmental attributes as� sociated with the presence and abundance of these biotic assemblages; and (4) using niche modeling, determine key attribute thresholds to illustrate how these biota could be employed as metrics of system integrity and ecologi� cal services. Several taxa were sufficiently abundant and widespread to use as bio–indicators: the presence and abundance of steelhead trout (Oncorhynchus mykiss), herpetofauna (reptile and amphibian) species richness, and signal crayfish (Pacifastacus leniusculus) represented different trophic positions, value as commercial resources (steelhead), sensitivity to environmental stress (amphibians), and indicators of biodiversity (herpetofauna species richness). Herpetofauna species richness did not differ, but abundances of steelhead trout, signal crayfish, and amphibian richness all differed significantly among tributary types. Niche models indicated that distribution and abun� dance patterns in both riparian and aquatic environments were associated with physical and structural attributes at multiple spatial scales, both within and around reaches. The bio–indicators responded to unique sets of attributes, reflecting the high environmental heterogeneity in headwater tributaries across this large watershed. These niche attributes represented a wide range of headwater environments, indicating responses to a number of natural and anthropogenic conditions, and demonstrated the value of using a suite of bio–indicators to elucidate watershed conditions, and to examine numerous disturbances that may influence ecological integrity.


Introduction
Understanding ecosystem dynamics and cross-scale interactions (Peters et al., 2007) are challenging in dendritic riverine ecosystems where geomorphology, hydrology, and landscape processes influence biotic communities at multiple spatial and temporal scales (Ward, 1989(Ward, , 1998;;Wiens, 2002;Allan, 2004;Lowe et al., 2006).Whole watersheds are logical units for examining these ecological processes because they are contained within distinct natural boundaries (i.e., Montgomery, 1999;Benda et al., 2004;Lowe et al., 2006).There is also a growing body of knowledge about watershed-scale processes and how they shape and influence ecosystem services (e.g., Naiman & Bilby, 1998;Fagan, 2002;Wiens, 2002;Allan, 2004).This knowledge can guide comparative studies and allows key principles of ecosystem processes to be uncovered (e.g., Grant et al., 2007).It can also promote the development of management strategies designed to enhance and protect the functionality of these systems (Lowe et al., 2006).By learning how far ecosystems can be perturbed without harming their integrity, resource managers can make informed decisions regarding natural resources.Ecological pro� cesses like stable food webs that provide abundant fish for harvest, or high and stable biodiversity that can facilitate ecosystem resilience, are essential to maintain sustainable riverine ecosystems.Only eco� systems that are managed sustainably will provide perpetual services without losing process elements and system integrity (Westra et al., 2000;Hooper et al., 2005;Karr, 2006).
First, 2 nd and 3 rd -order channels (Strahler, 1957) (hereafter headwaters) can comprise over 80% of channel length in a watershed (Dunne & Leopold, 1978).These small tributaries and their integral riparian environments are hotspots for watershed faunal diversity (e.g., Naiman & Decamps, 1997;Ward, 1998;Ward & Tochner, 2001;Fernandes et al., 2004;Sabo et al., 2005).The loss of this diversity can have negative consequences for entire ecosystems and their ability to function and provide sustainable services (e.g., Naeem, 1994;Loreau et al., 2002;Duffy, 2003;Dobson et al., 2006).Here, we focus on the ecological attributes of headwater tributar� ies and their unique aquatic and riparian animal assemblages (Lowe & Likens, 2005;Richardson et al., 2005).Headwater streams provide key functional links with terrestrial (Nakano & Murakami, 2001) and downstream environments (Wipfli et al., 2007;Freeman et al., 2007); they improve water quality, sort, clean, and deliver coarse organic substrates needed by stream organisms for cover and repro� duction, and provide nutrients for fish.Knowledge of how watershed level processes can potentially affect these functions is paramount for managing and maintaining the ecological integrity of riverine ecosystems (Meyer et al., 2007).
Our first objective was to describe unique low-or� der tributary types within the South Fork Trinity River (SFTR) watershed based on attributes representing a wide range of conditions, ecological processes, and disturbance regimes.To do this we used a combination of upland, riparian, and aquatic attributes associated with 60 randomly selected tributaries from across the entire watershed.We initially considered 69 variables, representing three spatial scales and numerous eco� logical processes, in a cluster analysis to distinguish unique tributary types.We followed with a series of scale-specific discriminant analyses, which served to reduce the number of independent variables and to detect those most informative attributes within and across scales.Our second objective was to determine if the abundance, evenness and species richness of common species or species groups differed between reach types, seeking potential bio-indicators.Our third objective was to model the distribution patterns of the bio-indicators using refined sub-sets of those independent variables used to differentiate the reach types.Understanding the environmental gradients that influence these bio-indicators can reveal important thresholds and key relationships that enable their uses as indicators of ecosystem integrity.

Study area and sampling
The SFTR is a 2,430 km 2 catchment in the Klam� ath-Siskiyou bioregion of northwestern California, USA.(fig.1).This bioregion is a globally significant area of biodiversity due to its age, range of geo� morphologies, soil types and moisture gradients, conditions that have created high endemism and many relict species (Whittaker, 1960(Whittaker, , 1961;;Welsh, 1994;DellaSalla et al., 1999;Sawyer, 2006).The SFTR is dominated on the west side by Douglas-fir (Pseudotsuga menziesii) mixed conifer/hardwood forest, with lesser amounts of ponderosa pine (Pinus ponderosa), montane hardwood-conifer, montane hardwood, montane riparian, and blue oak (Quercus douglasii)-gray pine (Arceuthobium occidentale) for� ests (Mayer & Laudenslayer, 1988); drier forest types are more dominant on the east side.Ownership is a mixture of federal (US Forest Service) and private lands.Periodic fires constitute an important natural disturbance with the SFTR experiencing median fire intervals of 11.5-16.5 years (Taylor & Skinner, 2003).The conservation strategy of the US Northwest Forest Plan (Thomas et al., 2006) lists the SFTR as a key watershed for the preservation of salmonid fishes.
A stratified random approach was used to distrib� ute 60 sample locations across the entire watershed in order to capture the full range of headwater aquatic and riparian conditions (fig.1).Headwa� ters were located using a Geographic Information System (GIS; ESRI, Redlands, CA) grid (cell size 1 km 2 ) overlaid on the watershed.Fifteen equally sized polygons were created from north to south, with four 1 km 2 cells randomly selected in each one.The centers of these grid cells were used to locate the closest headwater tributary.GIS-derived locations were visited and searched for potential sample reaches within 2 km of each starting point.
The search criteria consisted of locating a ≥ 300 m stream reach of accessible perennial surface flow shallow enough to sample without diving.Reaches near abrupt changes in vegetation (i.e., edge) and those surrounded by highly heterogeneous forest types were avoided.Although the SFTR contains many small roads used for timber harvesting and private access, this network was not used to find locations; channel access at or near roads required locating reaches ≥ 50 m upstream.

Data collection
We sampled sub-basins 14 to 1,900 ha in size and measured variables at three nested spatial scales that are not mutually exclusive.

The sub-basin scale
Attributes at the sub-basin scale were coarse in resolution, including topographic features and vege� tation mosaic elements representative of the entire sub-basin.Features included relative amounts of the primary forest types, annual minimum and maximum air temperatures and solar illumination, and geogra� phic features including aspect and mean elevation.
Values for sub-basin attributes were determined in GIS to characterize the larger spatial context in which reaches were embedded (appendix 1; 22 sub-basin scale variables).

The reach scale
Reach scale variables characterized the proximate tributary environment by measuring the structure and plant species composition immediately surrounding each reach (e.g., tree species composition by size class, ground cover vegetation).Variables were collec� ted in three circular plots centered on the reach and spaced equally at 50, 150, and 250 m.Each reach included 1/10th and 1/5th ha concentric circles, and one soil station per side, 25 m above the channel (appendix 1: 28 reach scale variables).

The habitat unit scale
Habitat unit scale variables characterized conditions within each reach, including canopy cover and stream channel morphology.We deployed water temperature data-loggers from June to October at the bottom of each reach to measure summer water temperatures.These data were used to calculate a mean weekly maximum water temperature (MWMT; Dunham et al., 2005).MWMT is derived by averaging the daily maximum water temperatures for the hottest week of the summer.At this latitude, greater daily extremes occur in summer and are more limiting than winter minimums for cool temperate-adapted fauna such as salmonids and many amphibians (Magnuson et al., 1979;Huey, 1991).We estimated fine sediments by calculating mean sediment depths from 10 pools in each reach (e.g., Welsh & Ollivier, 1998) (appendix 1: 19 variables).

Animal sampling
Two teams collected fish, amphibian and reptile data during daylight hours in June through Sep� tember from 2000-2003.The herpetofauna team used a four-tiered aquatic/riparian/upland ap� proach consisting of: (1) a channel-focused visual encounter survey (VES; Crump & Scott, 1994) of each 300 m reach, (2) 10 area-constrained (ACS) cross-channel aquatic sampling belts, (3) one ½hour seep-focused VES search, and (4) one upland (terrestrial) 4-hour VES search conducted on clear days between 10 am and 4 pm (Welsh & Hodgson, 1997).The channel-focused VES consisted of a single observer walking slowly upstream recording all observations.The observer walked three to four paces, stopped and scanned the wetted channel and bank-full width for animals.The 10 ACS consisted of one to two observers systematically searching a defined stream area, using acrylic view boxes to search underwater before, during, and after remov� ing all detachable channel substrates.A small dip net, held immediately downstream of the view box, captured dislodged animals.Locations of the ACS belts were determined by dividing the reach into five 60 m sub-sections and locating each belt us� ing random numbers.Within each sub-section, one "fast water" and one "slow water" habitat unit was selected based upon relative water velocity (e.g., Hawkins et al., 1993).Within each habitat unit, 1 m long cross-stream belts were situated in accessible habitat; areas that prohibited thorough searching (i.e., large downed logs) were avoided.The seep VES consisted of searches in seeps or springs detected during the channel VES.The upland VES occurred at least 10 m above riparian vegetation and was conducted for two-person hours on each side of a reach; data were collected at 40 reaches in 2003.The herpetological richness analysis included incidentals from 20 reaches re-sampled in 2001 and 2002.Fish data were collected in 2001, but due to re� duced late-summer flows and equipment failure only 55 of 60 reaches were sampled.The two-person fish team used a Smith-Root backpack electrofisher to sample fish.While electrofishing, we minimized cur� rent and set voltage to reduce animal trauma while facilitating capture (Reynolds, 1996).Each 300 m reach was divided into three sections.Within each 100 m subsection, we used six random numbers to locate sample units (18 units sampled per reach).When arriving at the first random distance from the bottom of each section, we electrofished the closest intact habitat unit (fast or slow) using a multiple-pass method.Block nets were not used because tributaries were small and water velocities low.Stunned animals were captured with nets and held in stream water.Animals seen but not captured were counted and identified to species and size category.To ensure conservative estimates, the number of un-captured animals enumerated during the successive n + 1 passes could not exceed the number encountered during pass n, unless obvious differences in size were observed.After the first habitat unit, the next Welsh et al.
random distance upstream was used to find the nearest opposite type (fast or slow water).The fish team sampled ≥ four days before or after the herpetofauna team.

Data analysis
Cluster analysis of environmental variables Our overall objective was to examine the full range of aquatic and riparian conditions that characteri� zed SFTR headwaters, and relate those conditions to particular faunal distributions.Therefore, we did not constrain our sampling to a specific set of attributes other than stream order.This enabled us to incorporate the considerable heterogeneity along multiple environmental gradients and across multiple spatial scales to characterize both reaches and the surrounding sub-basins in which they were embedded.However, we did assume that sufficient commonalities would exist among the 60 tributaries that would allow us to detect a reasonable number of unique sets based on their shared positions in the dendritic network and along environmental gradients.This would enable us to both compare reach types and to discern possible reasons for differences in animal distributions and abundances.To this end, we used non-hierarchical K-Means cluster analysis (Hintze, 2000), which minimizes the within-cluster sums of squares.We eliminated one variable each from six highly correlated (r ≥ 0.70) pairs, resulting in 69 variables from three spatial scales (appendix 1), variables were used with a K-Means algorithm (with 100 random starts and 1000 iterations) to determine unique reach types and assign group membership for each of the 60 reaches.

Multi-scale discriminant analysis
The clustering procedure established four unique reach groups, but it provided no information on the relative importance of the 69 variables.We employed discriminant analysis (DA) to determine which vari� ables differed among the four types and to rank their relative importance (Green & Vascotto, 1978;McCune & Grace, 2002).Ecological subsets were arranged by spatial scale and analyzed in a hierarchical series of DAs (SAS Institute Inc., 2003) to identify those variables that best determined group membership in each subset and at each scale.Variables at each scale were divided into subsets representing structural, compositional, or climatic attributes of the landscape, forest stand or stream environment (appendix 1) (e.g., Welsh & Lind, 1995).Four-group DAs were performed on each subset at each scale (appendix 1).The null hypothesis tested was that there were no differences between reach types for the variables within each subset.For model-building, variables were entered if the P value for the partial F statistic was ≤ 0.10 and removed it if it was > 0.05.A linear or quadratic discriminant function was calculated based on the variables selected.Bartlett's modification of the likelihood ratio analysis was used to test the homogeneity of variance-covariance matrices (SAS Institute Inc., 2003).
We then combined the significant variables from the DAs of the ecological subsets and performed composite DAs at each spatial scale.Our objective was to derive a reduced set of variables that best distinguished the reach types at each scale.We then ran a final multi-scale DA with the reduced number of variables from each scale-specific DA.With this iterative approach we were able to find those variables that were best able to discriminate between the reach types at each scale, and across scales, and thus reduce the initial number of environmental variables from the cluster analysis to just those that provided both the greatest discriminatory power and the most  information on how the tributary types differed.We tested the ability of our DA models to accurately predict whether or not the data from a given reach fit a particu� lar reach type (i.e., classification success) using both a jackknife procedure and a re-substitution test (SAS Institute Inc., 2003).Cohen's Kappa statistic (Titus et al., 1984) was computed for each model to indicate the classification success compared with chance.For this test, we equalized the prior probabilities of group membership because the true proportion of sites in each of the reach groups was unknown prior to the analysis (SAS Institute Inc., 2003).

Analysis of animal distributions
We ran ANOVAs (SAS Institute Inc., 2003) to examine the abundances, richness and evenness of faunal assemblages and individual species among the reach types, testing the null hypothesis of no difference in abundance for each assemblage or species across the four types.Our approach was based on the assumption that differences in animals among the reach types could be directly or indirectly linked to the different ecological attributes of these types.An example of an indirect link is the occurrence of tailed frogs (Ascaphus truei), a cold-water-adapted species whose presence can represent the capacity of streams to support similarly adapted fauna such as coho salmon (Oncorhynchus kisutch; Welsh & Hodgson, 2008).For parametric ANOVA we used log or square root transformations to reduce skewness.The distribution of crayfish could not be normalized so we used a non-parametric Krus� kal-Wallis ANOVA.When ANOVAs were significant, we used the Student-Newman-Keuls (SNK) a posteriori multiple comparisons to test group differences.We set α ≤ 0.10, as this level reduces chances of type II errors and is more appropriate for detecting ecological trends (Shrader-Frechette & McCoy, 1993).

Predictive models
To examine the relationships between the envi� ronmental attributes (appendix 1), and amphibian richness, and those individual species that varied by stream group based on ANOVA, we evaluated competing predictive models comprised of subsets of these attributes.Lizard diversity and western fence lizard (Sceloporus occidentalis) abundance varied significantly among stream groups based on ANOVA, but they are not riparian or aquatic obligates and their predictive models were weak so they were omitted in this final analysis.Using Spearman correlation analyses, we reduced the environmental attributes to those significantly correlated (α ≤ 0.1) with each of five bio-indicators.Relationships between these attributes and faunal metrics were assessed with non-parametric multiplicative regression analysis (NPMR) (McCune, 2006) using the software Hy� perNiche version 1.0 (McCune & Mefford, 2004).NPMR, designed for multivariate niche modeling, seeks to optimize a fit of detection data along multiple environmental gradients (i.e., in multi-dimensional attribute space) rather than adhere to a specific model form like linear or Poisson regression.NPMR considers interactions among all predictor variables in a given model (McCune, 2006).NPMR estimates a response at a given point in the predictor space by heavily weighting points that are near a target point, and giving less weight to distant points (using a minimum of three points); data points employed in the model comprise the ecological neighborhood.In model generation, we set the minimum neighborhood size to five percent of each sample.The term "tole� rance" is used to describe how broadly information is borrowed from nearby areas in predictor space while attempting to estimate the value of a particular attribute around a target point (McCune, 2006); it is thus akin to the niche breadth for that attribute.Tolerance is then the bandwidth used in the multi� plicative kernel smoother, given in the units of the environmental attribute (McCune, 2006).A species that is broadly tolerant to a particular attribute uses information from a large neighborhood of data points (McCune, 2006).We used a local mean estimator and Gaussian weighting function in all-possible-sub� sets regression for each set of models.Models were assessed using a leave-one-out cross-validated R 2 (xR 2 ), which is equal to one minus the ratio of the residual sum of squares over the total sum of squares (Antoine & McCune, 2004).We used the HyperNiche exhaustive search mode to determine best models, with up to six predictor variables, based on xR 2 (e.g., Giordani, 2007).Relationships between bio-indicators and variables are reported as positive (+), negative (-), or humped/U-shaped (^).

Cluster analysis
The K-Means cluster analysis considered options from two to six groups, with the four group solution the most informative; variation in the data explained dropped from 76.1 to 71.7% beyond four groups, declining more steeply thereafter.The numbers of tributaries in these groups were 11, 13, 16, and 20.We examined the Euclidean distance matrix values for the four group solution, and present scatter plots illustrating group separation (fig.2).

Discriminant analyses
Distinguishing tributary groups at the sub-basin scale We performed a DA of 22 sub-basin variables to detect differences in the landscape settings of the four tribu� tary groups (table 1).We found five geographic, three climatic, three disturbance, and one geologic attribute differed among the reach groups at the sub-basin scale (table 1).The best model was that of geographic relationships; the climate model was second, and the disturbance regime model third (table 1).When these 12 variables were combined in a sub-basin scale DA, seven contributed to the composite model; four geographic, two climatic, and one geologic (table 1).

Welsh et al.
Distinguishing tributary groups at the reach scale Twenty-eight variables were used to examine differen� ces in terrestrial environments adjacent to the tributary groups (table 2).The best model consisted of tree and log attributes, with five forest structure and tree composition variables differing (table 2).Other models indicated differences in understory and ground-level vegetation, ground cover, and the amount of upland forest canopy (table 2).The composite model deri� ved from the DA of these 13 variables contained six attributes, three showed differences in numbers of small and large conifers, and medium hardwoods, and three indicated differences in riparian forest width, and amounts of ferns and leaf litter (table 2).
Distinguishing tributary groups at the habitat unit scale Nineteen channel attributes were measured within the reaches (table 3).The aquatic conditions model was the best, indicating differences among groups in the amount of overhead channel canopy, per� cent of fine sediments (S-Star), and mean weekly maximum water temperature (MWMT) (table 3).Differences in aquatic substrates were indicated for percent boulders, pebbles, gravels, and visually estimated fine sediments (table 3).The composite model at this scale consisted of overhead canopy, fine sediments (S-Star), MWMT, percent boulders, and visually estimated fines (table 3).4).Using canonical scores from the greatly reduced set of variables in this multi-scale model (10 vs. 69 in the cluster analysis) we plotted the relationships of the 60 tributaries in three dimensions (fig.3).
The final multi-scale model improved group se� paration and provided useful information on the environmental gradients that separated the reach groups compared to the cluster analysis.

Stream groups and animal distributions
We found no differences in reptile or amphibian evenness, or in reptile richness among tributary groups.However, amphibian richness differed, with Group 2 having significantly greater richness than the other three groups (table 5).Several species including the southern torrent salamander (Rhyacotriton variegatus), the black salamander (Aneides flavipunctatus), the rough-skinned newt, and the Pacific chorus frog (Pseudacris regilla), were de� tected in numbers too low to test individually with ANOVA, but none-the-less contributed to differences in amphibian richness.Two anurans were sufficiently widespread and abundant for ANOVA.The foothill yellow-legged frog (Rana boylii) was more abundant in tributaries of Group 2 compared to the other types and the tailed frog was more abundant in tributaries of Group 1 compared with the other groups (table 5).Lizards (all species combined) were more abundant along tributaries of Group 4 compared with those of Group 2, and the western fence lizard was more abundant along tributaries of Group 4 compared to all other groups (table 5).Steelhead trout were more abundant in tributaries of Group 4 compared to the other three groups (table 5).Crayfish were more abundant in tributaries of Group 4 compared with Group 1 (table 5).

K-Means cluster groupings
Pooled within-group We sought to relate these reduced sets of variables to the presence and abundance of readily sampled fauna with sensitivity to system degradation (e.g., Welsh & Ollivier, 1998;Lowe & Bolger, 2002;Wilson & Dorcas, 2003), structuring of benthic communi� ties (Parkyn et al., 1997), or commercial value (i.e., steelhead trout).

Discussion
Our objectives were to detect unique sets of head� water tributaries, determine the riverscape patterns, disturbance processes, and environmental gradients associated with each set, and to link the distributions and abundances of riparian and aquatic biota with informative subsets of these attributes (e.g., Dale et al., 1994;Roth et al., 1996;Whittier et al., 2006).
The intent here was that by establishing these link� ages we would provide the basis for employing key elements of this fauna as bio-indicators of ecologi� cal services and network integrity.Any study that Our study differs from previous studies that ex� amined multi-scale environmental relationships of stream-dwelling animals (e.g., Lowe & Bolger, 2002;Roni, 2002;Welsh & Lind, 2002;Stoddard & Hayes, 2005) because many of these studies selected study sites based on categorical distinctions or disjunct dis� tributions.Most studies that claim to examine drivers of environmental suitability for particular taxa at mul� tiple spatial scales, a priori select sites along existing ecological gradients.These studies, therefore, often implicitly substitute anthropogenically forced spatial differences for naturally occurring spatial or temporal differences (Landres et al., 1999).By randomly select� ing sites throughout the SFTR watershed and deter� mining groups a posteriori, our study is unbiased in this respect and thus reveals environmental gradients that occur throughout the SFTR.Our assessment of faunal assemblage responses to this environmental structure was determined at a metacommunity scale (Leibold et al., 2004).
The classification success for ecological com� ponents within spatial scales ranged from 51-75% (sub-basin scale), 40-70% (land-surrounding-thereach scale), and 53-58% (within-reach scale).Within scales, classification success of the composite models improved markedly over the ecological sub-sets, with 85% success at the sub-basin, 78% at land-sur� rounding-the-reach, and 60% at the within-reach scale.The classification success improved even more with the final across-scales watershed level model, achieving 87% correct.As sets of variables were refined at each step, the improved success indicated an enhanced ability to discern a much reduced, yet more informative, set of attributes able to distinguish tributary types.Similar approaches using multivariate analyses have proven useful in other studies seeking to reduce the dimensionality of large data sets by finding the fewest meaningful variables to differentiate sets of sites (e.g., Radwell & Kwak, 2005;Shrestha & Kazama, 2007 and references therein).Fig. 3. Three-dimensional scatter plot of canonical scores from the multi-scale discriminant model (table 4).See appendix 1 for definitions.For illustration we used the highest absolute canonical score to assign each variable to a particular canonical axis, however, each variable loads on each of the axes.Welsh et al.

Fig. 3. Diagrama de dispersión tridimensional de los datos canónicos del modelo discriminante multiescala (tabla 4). Para las definiciones, ver el apéndice 1. Para realizar la ilustración utilizamos el dato canónico absoluto más alto para asignar cada variable a un eje canónico en particular, a pesar de que cada variable carga valores en cada uno de los ejes.
The composite multi-scale model greatly improved on the cluster analysis by using just 10 variables compared to 69, and demonstrating greatly improved separation (compare figures 2 and 3).This model distinguished the four tributary groups along informa� tive environmental gradients based on five sub-basin scale variables (50%), three from land-surround� ing-the-reach (30%), and two at the within-reach scale (20%).Four of the 10 variables represented processes or attributes that respond directly to both anthropogenic modifications and/or natural distur� bances within the landscape (conifer and hardwood counts, riparian width, stream temperature [MWMT] and percent fine substrates).The composite models at each of the scales also contained informative at� tributes that respond directly to land management practices such as forestry and road-building (Tang et al., 1997;Hemstad & Newman, 2006).
Several attributes distinguishing the reach groups overlapped in values (fig.4) indicating that these sets likely represent different positions along a continuum (Vannote et al., 1980), or gradient, within the dendritic network (Benda et al., 2004).Tributar� ies of Group 1 were the lowest order tributaries, at the highest elevation, with the narrowest riparian zones, lowest water temperatures, lowest daily water temperature fluctuations (amplitudes), highest fine sediment loads, and the fewest road crossings, and represented the highest end of the continuum (tables 2-4).Tributaries of Group 2 were the west� ern-and northernmost streams, received the most precipitation, and with the highest mean annual air temperatures.Although the slopes of these reaches were just slightly greater than those of Group 1, they were the lowest in elevation, and transected the most mafic volcanic rock and chert, parent material that appeared to support more hardwood compared to coniferous forest types.These tributaries also had the highest upland and over-stream canopy, with riparian areas being the highest in mesic and hydric plants (tables 2-4).Tributaries of Group 3 were the eastern-and southernmost, with the lowest winter Fig. 4. Six key environmental attributes of the four reach groups illustrating the overlap in physical attributes consistent with a continuum (Vannote et al., 1980) or a hierarchical channel network (Benda et al., 2004).The boxes represent the middle 50% of the data, lines inside are the median, the T-shaped whiskers represent data 1.5 times past the middle 50%, and dots represent outliers.Fig. 4. Seis atributos ambientales clave de los cuatro grupos de cursos que ilustran el solapamiento de los atributos físicos, lo que es consistente con un continuum (Vannote et al., 1980) o con una red de canales jerarquizada (Benda et al., 2004).Los cuadrados representan el 50% medio de los datos, las líneas en su interior son las medianas, los signos en forma de T representan los datos que se hallan a 1,5 veces la media del 50%, y los puntos representan los valores atípicos.

Responses of the bio-indicators
The coastal tailed frog was the only relatively com� mon amphibian associated with tributaries of Group 1.However, the best predictive model (table 6) was relatively weak and uninformative (cf.Welsh & Lind, 2002).This poor performance likely resulted from the uneven distribution and low abundances we found for this species, despite evidence (Bury, 1968) of a once wider distribution throughout this and surrounding major watersheds, including to the east.Such patchy distributions have been observed elsewhere in recent studies, and are likely artifacts of past timber harvesting altering the requisite niche of this ancient frog, a species specifically adapted to conditions that occur most reliably in late succession forests (Welsh, 1990;Welsh & Lind, 2002;Welsh et al., 2005;Spear & Storfer, 2008).Consequently, the tailed frog is an excellent bio-indicator for the more structurally diverse, micro-climatically ameliorated, conditions typical of late seral forests (Welsh, 1990) which also support the highest levels of terrestrial salamander biodiversity (e.g., Davic & Welsh, 2004).Furthermore, the presence of this frog can indicate the potential of streams to support coho salmon (see Welsh & Hodgson, 2008), a threatened salmonid once common in SFTR but that has not been detected in recent times.(Lind, 2005).Differences in the predictive models between this frog and amphibian richness (despite the high values may for both, in this tributary group), is best explained by the comparatively high habitat heterogeneity among these streams, along with the specific and unique behavioral adaptations of the yellow-legged frog.Presence of the southern torrent salamander, the black salamander, the rough-skinned newt, and the Pacific chorus frog, along with the yellow-legged frog, combined to establish the highest amphib� ian richness among the tributary types in Group 2. This high amphibian richness is likely indicative of conditions that also support higher richness of other aquatic taxa associated with the aquatic and riparian habitats.High amphibian richness can function as an easily measured bio-indicator, where greater values indicate enhanced resilience and an improved likeli� hood that reaches can provide and sustain critical ecological services (Dobson et al., 2006).
None of the fauna in tributaries of Group 3 dif� fered significantly in value from the other groups (table 5).These reaches were the most eastern, had lower winter sun exposures, lower precipitation, lower hardwoods 28-60 cm DBH, and lacked mafic volcanic rock and chert, compared to the other groups.This outcome indicates that for the fauna we assessed, the tributaries of Group 3 appear to be relatively impoverished compared to the other groups.
The tributaries of Group 4 supported more steel� head than other groups and more crayfish than Group 1. Crayfish abundance was associated with greater width and less channel embeddedness.Higher steelhead numbers were associated with greater basin area and higher temperatures (both MWMT and MWMT amplitude).The higher abundance in streams with higher temperatures occurred despite potential bioenergetic costs.Employing a subset of streams from each of our three fish-bearing groups (three per group) sampled in 2003, McCarthy et al. (2009) ) showed that individuals in higher temperature streams had lower growth efficiency, with some fish losing weight during the summer months.Bioenergetic models suggested that these fish were feeding at lower rate, 25% (or less) of the maximum consump� tion rate, and that projected future increases in stream temperatures could further exacerbate low growth rates and perhaps have population level effects.The streams of Group 4 also had more road crossings and exposed soil cover in the adjacent upland, a condition which can negatively affect salmonids (e.g., Cederholm et al., 1981).
Influences on the fluvial network are hierarchical, with regional controls such as climate, physiogra� phy, and geology shaping sub-basin conditions, and both sets of attributes acting to shape each sub-basin tributary and its within-reach conditions (Knighton, 1984;Poole, 2002).Given the overlaps in the predictive environmental attributes and the distributions of our bio-indicators, we emphasize that these fauna are not associated just with the particular tributary type where they are most com� mon.Rather the faunal elements generally have peak abundances in particular tributary types, but also occur in lower numbers in adjacent types reflecting the continuous nature of the fluvial network (fig.4) (see Pringle, 2003).It is the collective influences on the greater sub-basin which shape available habitats within tributary types and determine where particular fauna are favored (Gomi et al., 2002;Benda et al., 2004).This conceptualization is supported by the outcome of our predictive modeling where a variety of significant models were derived for co-occurring species.We interpret this outcome as evidence that our set of independent variables represent numer� ous informative environmental gradients within this watershed.Further, because the NPMR models con� sisted of sets of variables acting at different spatial scales, we consider this evidence of the influence of cross-scale interactions (Peters et al., 2007), with attributes acting in unique combinations to influence each bio-indicator depending upon its evolved niche.The multi-scale analysis was informative because it combined variables affecting natural variability (Lan� dres et al., 1999) and land-use history (Foster et al., 2003), and substantiating their combined influence on headwaters.This analysis illuminated variables that can be managed to improved ecological conditions and enhance headwater health, and recognizing that organisms are integrators of all that happens in a watershed (Karr, 2006), the NPMR models indicated several bio-indicators that could be used to track their improving trajectories (Tabor & Aguirre, 2004;Nichols & Williams, 2006).Future papers will address metacommunity dynam� ics (e.g., Welsh & Hodgson, 2010) and fine scale (i.e., microscale) responses of faunal assemblages, which may allow us to elucidate additional factors that affect the spatial patterns of stream-dwelling organisms in this watershed.

Fig. 1 .
Fig. 1.The South Fork Trinity River Watershed, California, USA, with sampling locations of the 60 headwater reaches.Circles represent reach Group 1 (n = 13), triangles represent reach Group 2 (n = 16), squares represent reach Group 3 (n = 11), and diamonds represent reach Group 4 (n = 20).See methods for details on the determination of group membership.

Fig. 2 .
Fig. 2. Scatter plots of Euclidian distances showing six views of the separation of the four group solution from the K-Means cluster analysis (R Development Core Team, 2009).

Table 5 .
ANOVA tests of faunal assemblages and individual amphibian and reptile species, steelhead trout, and crayfish abundances among four stream groups.Data used for individual species were the sums of VES, seeps, and 10 m 2 of belts (see text), (n = 60) or electrofished reaches (n = 55).For assemblages, we used numbers of species detected per tributary, richness including incidentals.Lagarto de vientre azul del oeste, Sceloporus occidentalis; lagarto de Sagebrush, S. graciosus; lagarto aligator del norte, Elgaria coerulea.

Table 6 .
Non-parametric multiplicative regression (NPMR) models for five bio-indicators whose distributions varied significantly among stream groups.The data used in the modeling were those from just the reach groups where each metric was observed.Tolerance is in the units of the response variable and refers to the niche width along that variable; ecological neighborhood size refers to that portion of the data used to determine tolerances for each variable in the model: xR 2 .Leave-one-out cross-validation R 2 ; Ns.Neighborhood size.(See methods for more details.)Tabla 6. Modelos de regresión multiplicativa no paramétrica (NPRM) para cinco bioindicadores, cuyas distribuciones variaban significativamente entre los grupos de cauces.Los datos utilizados en la modelización fueron los de los grupos de tramos, en los que se observó todo parámetro métrico.La tolerancia está en las unidades de la variable respuesta y se refiere a la anchura del nicho a lo largo de dicha variable; el tamaño de la vecindad ecológica se refiere a la porción de los datos usados para determinar las tolerancias para cada variable del modelo: xR 2 .R 2 por validación cruzada dejando uno afuera; Ns.Anchura de nicho.(Para más detalles, ver los métodos.)