Ethics statement

Data was collected by state biologists and made readily available to the community. Observations were made by aerial surveys, and no animals were handled.

Study site

The Chena River is a large clear water tributary (drainage area ~ 5,130km²) of the Tanana River, which is part of the Yukon River drainage in Alaska, U.S.A. (Fig 2). Flow within the Chena River is derived from four major tributaries and hydrology is primarily driven by snow melt, precipitation, and groundwater from unconfined aquifers [42]. Mean annual flow is 38.5 m³ s^-1 over the period of record, although construction of the Moose Creek Dam has prevented flows downstream of the dam from exceeding 340 m³ s^-1 since its completion in 1981.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Map of the Chena River basin and its location in Alaska, USA (inset).

The Core (C1 and C2) in the top panel represents the concentrated spawning area identified by aerial surveys and state biologists during carcass surveys, with the periphery (P1 and P2) located northeast of the core. The known distribution of spawning Chinook salmon in the Chena watershed is depicted by the Anadromous Waters Catalog (https://www.adfg.alaska.gov/sf/SARR/AWC/). The bottom panel shows predicted spawning habitat quality (range 0–1; darker is higher quality) based on resource selection function analysis for each study reach. The Middle Fork confluence is identified by MFC.

https://doi.org/10.1371/journal.pone.0177467.g002

Aerial counts

We compiled aerial survey counts of spawning Chinook salmon conducted during 1971 to 2014 from a database managed by the Alaska Department of Fish and Game (ADFG, S1 Table). Survey conditions (good, fair, poor, or undefined) were noted during each survey, and surveys were conducted up to 7 times within a year (1992) to capture peak spawn timing. We considered the maximum count for each of the four stream study reaches as our data point for a location and year because peak counts provide the most likely scenario under which strong intraspecific competition may occur, as well as being a common response variable modelled in other aerial survey count studies [23, 43]. We reduced the number of years (n = 10) in which peak counts were used for count model development by only considering years in which counts were made within four spatially defined stream reaches. These sections of stream (study reaches from this point forward) were in the core (C1 and C2) and periphery (P1 and P2), and were selected because they were the most consistently flown segments (Fig 2, S1 and S2 Tables).

Count model estimates can be biased by not accounting for heterogeneity in detection efficiency across sampling locations and through time [44]. Detection efficiency of aerial surveys is difficult to measure owing to lack of survey replication and meeting necessary closure assumptions. We explored the effect of survey condition on aerial counts by comparing total aerial count peak densities (total density = summed peak counts¹ stream km^-1) to escapement estimates for the Chena River Chinook salmon population (Supplementary Material, S1 File -Test description for bias in detection efficiency). We found a strong positive relationship between escapement estimates and total peak densities (S1 Fig, p = 0.002, R² = 0.65), as well as slightly higher predicted counts than observed counts during poor quality surveys (S2 Fig). As a result, we assumed spatial and temporal variability in aerial count detection efficiency was minimal for two reasons. First, a strong relationship between total densities and escapement suggested limited temporal variability in detection efficiency, further supported by survey condition via residual plots. Second, within-year survey conditions were consistent among study reaches with the exception of 1999 where conditions were fair for all locations except in P2 which was poor.

Resource selection functions

We used a digital landscape model (NetMap; Earth Systems Institute, Mt. Shasta, CA) parameterized for the Chena River basin to derive hydrologic and geomorphic DI habitat attributes potentially important to spawning Chinook salmon. The NetMap model generates a synthetic digital stream network from a remotely-sensed digital elevation model (DEM) based on flow accumulation and channel delineation algorithms [44,45]. The result is a network of 20–200 m sub-reaches throughout the entire Chena River watershed (2,265 stream-km) to which physical attributes are assigned based on empirical geomorphic relationships [46]. We used three physical attributes that strongly correlated with Chinook salmon spawning habitat suitability in other regions [47,48] to predict spawning suitability for each of our four study reaches: stream gradient (GRAD; %), bankfull width (BFW; m), and valley width index (VWI; a measure of valley constraint; unitless). These three attributes showed low multicollinearity (variance inflation factor, VIF < 3 [49]) and were used to develop a resource selection function model (RSF).

Our RSF was fit using the regression approach outlined by Nielson and Sawyer [50]. The response, Chinook salmon redd counts, was estimated from independent aerial surveys (i.e., the surveys were not included in subsequent count models) conducted in 2005 and 2006 [51]. During each survey, counts of redds were georeferenced and subsequently associated with a sub-reach. We modeled redd counts as a function of GRAD, BFW, and VWI based on a negative binomial distribution using the “glm.nb” function in the R statistical program’s MASS package [52]. Median parameter estimates and 90% confidence intervals for each covariate were estimated using a bootstrapping technique [53] wherein we fit the model for 1000 permutations [54] using all 176 used sub-reaches along with an equal number of randomly selected available sub-reaches from the known spawning distribution of Chinook salmon in the basin (n = 2536 sub-reaches, Anadromous Waters Catalog, Fig 2). The resulting model was used to predict intensity of use (from this point forward “RSF score”) for all sub-reaches within the four study reaches by rescaling predicted counts between 0 and 1 [50]. Each of the four study reaches (C1, C2, P1, and P2; measured at a coarser spatial scale than sub-reaches for RSF analysis) was assigned the mean RSF score from all nested 50–200 m sub-reaches and used as a fixed effect DI covariate for our count model (see below).

Flow model

We used a semi-distributed water-and-energy balance hydrologic model (variable infiltration capacity model, VIC; [55]) parameterized for the Chena River [56] to estimate mean daily flows (m³ s^-1) for the four study reaches across the study period. Predicted flows were used because they provided unique spatial and temporal flow estimates for each study reach. We used the approach proposed by Wenger et al. [57] to downscale daily flow predictions produced by the base model to the stream reach-scale. We constructed hydrographs for each study reach using flow predictions at the farthest downstream end of each study reach. We then developed three DI flow metrics for each study reach during each year in which aerial spawner counts were available, including maximum and mean daily flow (m³ s^-1), and flow variability (C.V. of mean daily flows). All metrics were developed for the interval in which spawning occurs for Chinook salmon of the Chena River (01-July to 31-Aug).

Temperature model

Stream temperature data were collected using HOBO Pro V2 data loggers within each of the four study reaches during June 2005-November 2006 [51]. We built predictive models using multiple stepwise linear regression to predict stream temperature within each of the four study reaches during years in which aerial counts were made. Predictor variables used for model construction were extracted from a local climate station in Fairbanks, AK (wunderground.com). We removed highly correlated climate variables (Pearson’s r > 0.6), which resulted in five climate station variables (mean daily air temperature, maximum daily humidity, square root transformed minimum daily humidity, barometric pressure, and wind velocity) used to predict stream temperature in each study reach. Variables were sequentially removed during model development using deletion tests within a stepwise regression framework. Deletion tests remove the least significant parameter from a model and F-tests (α = 0.15) identify the most parsimonious model that balances complexity and explained variability [58]. Final models were used to backcast mean, minimum, and maximum stream temperatures based on historic climatic variables for each study reach and year combination.

Count model

We used a generalized linear model to predict Chinook salmon peak spawner counts within study reaches (i.e. C1, C2, P1, and P2) as a function of DD and DI factors based on a negative binomial distribution. Fifteen candidate models (Table 1) were developed and ranked to determine whether DD, DI, or a combination best explained counts using an information theoretic approach and AIC model-selection [59] with the MuMin package [60] in R. A categorical spatial reference variable (reference to being in the core or periphery, CP) was also included as a fixed effect. Annual carcass surveys conducted by Alaska Fish and Game biologists have identified segments of stream between the lower end of C1 and upper end of C2 to account for the majority of Chinook salmon spawners within the watershed (J. Savereide unpublished data), and therefore aerial surveys outside of this focused length of stream was considered peripheral habitat. We identified and removed multicollinear predictor variables with high (>3) VIF’s which resulted in three DI variables included in candidate models: average flow, minimum temperature, and mean RSF. Annual estimated escapement for the Chena River basin [61] was included to represent DD factors. Models representing DD (models 1:5, Table 1) were based on multiple combinations of additive and interactive effects between escapement and a reference to core or periphery spawning habitat. Models representing DI (models 6:9) were simple models defined previously (Resource Selection Functions, Flow Model, Temperature Model), with one model representing additive effects of all DI parameters. The remaining six models (models 10:15) represent a combination of DD and DI effects. Models 10:12 identify potential interactive effects between DI variables and escapement, where counts were not dependent on reference to core or peripheral spawning habitat. The remaining three models allowed interactive effects between escapement and reference to the core and periphery, and also included an interaction between the core or periphery and DI variables (models 13:15, Table 1). These models were constructed to allow DD and DI factors to vary in the core or periphery separately, testing the abundant center hypothesis. We used the resulting estimates of the relationship between predictor variables and peak counts (linear, nonlinear or interactions), to determine whether habitat selection followed an IFD (linear isodar) or IDD (curved isodar, nonlinear predictor or predictor interaction; [23]). Lastly, we included an offset based on stream length in each candidate model to account for unequal lengths of each study reach.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Model selection results for candidate models predicting Chinook salmon peak spawner counts in the Chena River, Alaska.

https://doi.org/10.1371/journal.pone.0177467.t001

Model validation

Count model predictive performance was assessed using the 0.632+ bootstrap method [62–64] because this approach has been found to be less biased and result in lower variability in predictions compared to other cross-validation techniques, as well as to more reliably approximate true values with small sample sizes. The four bootstrap adjusted metrics used to assess model performance in this study were Pearson’s correlation (r), Spearman’s correlation (ρ), average error (Mean_error), and root-mean squared error (RMSE) [64,65]. Two hundred bootstrap simulations were run for each metric. We considered the best predictive model to have the highest correlation values (r and ρ) and lowest error estimates (Mean_error and RMSE). Predicted counts and subsequent isodar development (see below) were then conducted using the best predictive model determined from model validation. We further assessed the performance of the best predictive model by estimating the same model performance metrics, but compared predicted counts and out-of-sample counts not used in model development (S1 Table). Out-of-sample counts were collected within the same study reaches but during different years, and were excluded from model development because all four study reaches were not monitored within the same year during these surveys.

Isodar development

Predicted counts from the best predictive model were used to construct isodars comparing differences in habitat quantity and quality between the core and periphery. Model selection and validation identified the best predictive model depended on a spatial reference to habitat within the core or the periphery, so isodars were constructed comparing general counts within the core vs. counts in the periphery (Table 1). For example, the best model indicated the only difference between counts in C1 and C2 was based solely on stream length (the offset). Therefore, the isodar was constructed from predicted counts in the core vs the periphery, where the mean length of stream from all four stream segments was used as the offset for count predictions (21.8 km). We predicted counts in each habitat type (core or periphery) for each of the 10 escapement values used for model construction. To propagate error from predicted counts into the isodar, we employed a bootstrap where we selected 10,000 random counts from both the core and the periphery. Random counts were generated from a normal distribution from the mean and standard error of predicted counts for each escapement value from the best predictive model (again generated for the core and periphery). We calculated 90% confidence intervals for the slope and intercept of the 10,000 isodars constructed from the 100,000 bootstrapped counts at each habitat type (10,000 counts generated for each of the 10 escapement values). We considered habitat quantity to differ between core and periphery if intercept confidence intervals did not overlap zero. Habitat quantity, an intercept of zero, defined by isodars would indicate the relative difference in fitness experienced by individuals within the core versus periphery as a result of DI differences between habitats (Fig 1, [16,66]). A slope significantly different than one would indicate differences in habitat quality between the core and periphery, where differences in fitness experienced by individuals within the two habitats become greater with increasing density (Fig 1, [16,66]).

Resource selection function

Estimates of regression coefficients for the RSF model (S3 Table) indicated that the intensity of use by spawning Chinook salmon was highest where the valley was unconstrained, the bankfull width wide, and gradients low (β_VWI = -0.008 ± 0.002 s.e.,β_BFW = 0.069 ± 0.005 s.e., β_GRAD = -0.590 ± 0.067 s.e.). Average RSF scores were higher in core relative to periphery habitats (Fig 2), and generally decreased with distance upstream. The C1 study reach had the highest mean RSF (0.259 ± 0.161 Std.) with the lowest values observed at P2 (0.079 ± 0.016 Std., S2 Table).

Stream flow

Mean discharge from the VIC model for the spawning period ranged from 31.6 m³ s^-1 to 64.3 m³ s^-1, whereas average maximum discharge ranged from 155.3 m³ s^-1 to 298.3 m³ s^-1 (S2 Table). In 3 of 4 study reaches, the highest mean discharge occurred in 1998, while the highest discharge in C1 was in 1986. The lowest average flows for all four study reaches were observed in 1993.

Stream temperature

Stream temperature models had strong predictive performance across all study reaches (RMSE = 1.04°C-1.35°C, S4 Table) although all models slightly under-predicted temperatures at the upper bounds of temperature regimes (S3 Fig). Unsurprisingly, the downstream-most study reach (C1) had the warmest average maximum (12.5°C), mean (9.3°C), and minimum (5.9°C) temperatures, whereas the farthest upstream study reach (P2) was coldest (S2 Table). The warmest average predicted stream temperature for each study reach occurred in 1990 (9.0–10.2°C) and the coldest in 1998 (7.7–9.0°C).

Count model and model validation

The top four models in our candidate set accounted for 91% of model weight, and all were DD models (Table 1). The model with the highest weight was an additive model with E, CP, and E² (see Table 1 for definitions, model 4, w_i = 0.36). The next best model also accounted for a substantial amount of model weight (model 2, w_i = 0.35), and was identical to the top model but did not include E². The final two models in the confidence set were similar to the top two models, but included an interaction between E and CP (Table 1). Predictive performance was similar for all of the top four models, when assessed for observed data (0.632+ bootstrap) and out-of-sample data (Table 2). Model selection and model validation indicated that the best predictive model was model 2, which we used to construct and interpret isodars.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Model validation from assessment metrics.

https://doi.org/10.1371/journal.pone.0177467.t002

Modelled peak spawner counts increased in both the core and periphery with escapement (S4 Fig). For the full time-series of escapement estimates, predicted peak spawner counts were highest in 1997 and lowest in 2013 for all study reaches (Fig 3). Consistently low counts for all study reaches were predicted from 2006 to 2013 and matched observed patterns for the Chena River basin.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Predicted peak spawner counts from model 2 (y-axis; Table 1) as a function of annual Chinook salmon escapement estimates (1986–2013; x-axis) for the Chena River, Alaska.

Observed (symbols) and predicted counts (lines and confidence ribbons) through time for C1 and P1 study reaches (top panel) and C2 and P2 study reaches (bottom panel) are shown. Dark and light grey ribbons represent 90% CIs for core and periphery habitats, respectively. Open symbols are observed counts used to fit the model and solid symbols are out-of-sample counts used for model validation. Predicted counts for each study reach are adjusted for the offset (study reach specific stream length).

https://doi.org/10.1371/journal.pone.0177467.g003

Isodar

A linear isodar curve, where core Chinook salmon spawning habitats were occupied prior to peripheral habitats, was supported by the 0.632+ bootstrapped model validation which indicated model 2 was most accurate at predicting peak spawner counts. Yet, closer inspection of the regression parameters from the isodar bootstrap analysis suggested that habitat quantity was not significantly different between the core and periphery (intercept = 129.3, 90% CI = -76.4 to 297.2, Figs 4 and 5). Significant deviation of the isodar slope from a 1:1 relationship was observed (slope = 0.230, 90% CI = -0.105 to 0.780), indicating habitat quality was significantly different between the two spawning habitats based on isodar interpretation (Fig 5). Similar isodar patterns were demonstrated for isodars constructed for observed aerial counts between each core and periphery study reach (S5 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Isodars for predicted counts between core and periphery Chinook salmon spawning areas in the Chena River, Alaska based on model 2 (see Table 1 for model list) count predictions.

Individual gray lines represent the 10,000 isodars developed from bootstrapped counts from both a core and periphery habitat. The solid black line represents the mean isodar from the 10,000 simulated isodars. Isodars were constructed for the core and periphery, based on count predictions from each habitat of similar length (making the offset for each habitat the mean length of all study reaches).

https://doi.org/10.1371/journal.pone.0177467.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Frequency distribution of slope and intercept estimates from the isodar model comparing counts in the core and periphery.

Each vertical bar indicates the frequency of the coefficient estimate (slope or intercept) from the 10,000 isodar curves. Dark vertical lines indicate the critical value in which isodar coefficients (habitat quality different = slope, habitat quantity = intercept) would favor either the core or periphery. dark gray and light gray bands below vertical bars indicate the 90% and 95% confidence intervals, respectively.

https://doi.org/10.1371/journal.pone.0177467.g005

Mechanisms distributing spawning salmon

High counts in the core spawning area were consistent with the abundant center hypothesis, where strong regulation at a population’s core is predicted to control population dynamics [10]. In our system, densities in peripheral habitats increased concurrent with overall population density suggesting strong competition for core habitats forced spawning salmon into the periphery. Periphery counts were dependent on DD processes which was further supported by the lack of evidence for any model with DI variables explaining salmon distributions at this coarse spatial scale. The prevalence of DD throughout the salmon spawning range in the Chena River provided justification for using DD habitat selection tools (i.e. isodars) to explore realized fitness advantages via habitat selection decisions.

The lack of DI factors influencing count distributions in the periphery did not conform to predictions of the abundant center hypothesis. We expected Chinook salmon spawner counts in the periphery to be more strongly influenced by DI factors such as stream temperature or flow. Chinook salmon, like many salmonids, are known to avoid habitats that exceed critical thermal ranges [69,70]. However, the Chena River is located near the northern extent of known Chinook salmon distributions and consequently supports temperature regimes (S2 Table) well below stressful thermal conditions for adult Chinook salmon (18°C, [70]). Streamflow is an equally important DI stream characteristic affecting spawning Chinook salmon habitat selection [23] and overall population dynamics have been linked to flow variability in the Pacific Northwest [33]. We did not find a significant effect of flow variability on spawner counts, which was likely due to our study reaches being relatively close in proximity to one another with low intra-annual flow variability among the study reaches. However, strong effects of flow on Chinook salmon population dynamics within the Chena River have been observed, particularly on juvenile life stages [36]. Results from that study coupled with this analysis suggest that flow variability in the Chena River has a stronger influence on juvenile salmon population dynamics than adults. Although temperature and flow are arguably the two most influential DI factors structuring aquatic populations, we found them to have relatively limited control on spawning Chinook salmon relative to DD mechanisms.

Habitat selection within the watershed

Spawning habitat selection by Chena River Chinook salmon was better explained by IFD than a dominance hierarchy. This was surprising given what is known about aggressive defense of spawning locations by adult salmon [37,71]. However, evidence of IFD has been documented in other fish species [72,73] as well as other salmonids [20,74–76], although IDD in these studies was not directly tested. Therefore, it is not unrealistic to find evidence of IFD influencing spawning Chinook salmon distributions in the Chena River, especially at the spatial scale at which our analysis was conducted. Indeed, Morris [29] emphasized how interpretation of habitat selection can be altered simply by the spatial scale of a study. For Chinook salmon, homing can occur at spatial scales as fine as 1 km [77]. Habitats at the reach scale (our analysis; 10.6–40.5 km stream length) would be filled independent of competition, due to homing, and then competition within each reach (core or periphery) would structure how fish occupy habitats of varying quality [71]. Our RSF analysis demonstrated variability in the quality of habitat available within each study reach. Furthermore, the intercept of our isodar analysis suggested that neither the core nor periphery were more densely occupied during low escapements. This provides support for the hypothesis that Chena River Chinook salmon might home at spatial resolutions finer than the watershed scale (e.g., stream reach). However, the scale of homing by Chena River Chinook salmon is currently unknown, and further study is warranted.

Isodars indicated habitat quality was significantly better in the core than the periphery. The position of the isodar relative to the 1:1 line, being significantly lower than one, indicated the relative difference in fitness for spawning salmon between the two habitats became greater with increasing densities in both the core and periphery. If there is fine-scale homing in the Chena River Chinook salmon population, our results would suggest that greater spawning activity in the core relative to the periphery may be a result of higher demographic success of early life-history stages (e.g. egg incubation, juvenile rearing). Strong site fidelity in salmonids is common when reared in higher quality habitat (e.g. higher food productivity), while straying rates become more common when rearing habitat quality is low [48]. Intrinsic rearing habitat potential models for juvenile Chinook salmon in the Chena River predicted higher quality rearing habitat between the Moose Creek Dam and the Middle Fork confluence [78]. All core spawning habitat, as well as much of P1, is located within this same area (Fig 2). However, confirming differences in fitness between these two locations as well as fine-scale site fidelity (at least in the core) would require more research.

Alternatively, if fine-scale site fidelity was less influential in habitat selection decisions, our results would lend support to source-sink spillover effects. Density-dependent habitat selection theory has strong application to source-sink theory, in which IDD behavior leads to source-sink structuring within a metapopulation [79,12]. Increased densities in the source (core) would lead to exceeding carrying capacities and subordinate movement into lower quality sink (periphery) habitat. Spillover, or mass effects, in other fish populations [80,81], provide evidence for this hypothesis and may help explain elevated salmon densities in peripheral habitat during high escapement years. Currently, few studies have identified vital rates in Chena River Chinook salmon [36], and none to our knowledge have identified vital rates at fine spatial scales within the watershed (e.g. among stream reaches or tributaries). Quantification of egg and juvenile survival rates, as well as successful spawning dynamics (redd establishment, mate acquisition, etc.) would be required in core and periphery habitats before source-sink models could be applied. This is a dilemma throughout Alaska where knowledge of demographic rates is strongly needed during freshwater phases of salmon life-cycles [38].

Although spawner counts did not entirely conform to the abundant center hypothesis, a system that does would have unique consequences for the resulting isodar. Counts in the periphery being less dependent on DD than the core would result in counts from both habitats being out of synchrony. If mean counts in both habitats are equivalent, this would result in an isodar with a large range of potential intercept and slope estimates. The more out of phase counts in both habitats become, the more negative an isodar slope and positive an intercept. However, the abundant center hypothesis also predicts that counts should be much greater in the core than the periphery, therefore the resulting isodar would be similar to our current isodar. Specifically, the resulting isodar would demonstrate a horizontal line with a slope closer to 0 than 1. This scenario would then lead one to ask, would an isodar be appropriate to infer the relationship between habitat selection and fitness among habitats when DD does not influence habitat selection in all habitats? A similar question has been asked by Halliday and Blouin-Demers [82] when exploring the importance of DI mechanisms on habitat selection by ectotherms. Regardless, we agree with Halliday and Blouin-Demers [82] that this is an interesting avenue of future research.