Refuges and ecological traps: Extreme drought threatens persistence of an endangered fish in intermittent streams

Abstract Recent droughts raise global concern over potential biodiversity loss and mitigating impacts to vulnerable species has become a management priority. However, drought impacts on populations are difficult to predict, in part, because habitat refuges can buffer organisms from harsh environmental conditions. In a global change context, more extreme droughts may turn previously suitable habitats into ecological traps, where vulnerable species can no longer persist. Here, we explore the impacts of California's recent record‐breaking drought on endangered juvenile Coho salmon. We estimated the variability of cumulative salmon survival using mark–recapture of nearly 20,000 tagged fish in intermittent stream pools during a 7‐year period encompassing drought and non‐drought conditions. We then determined the relative importance of physical habitat, streamflow, precipitation, landscape, and biological characteristics that may limit survival during drought. Our most striking result was an increase in the number of pools with reduced or zero survival during drought years and a coincident increase in spatial variability in survival among study reaches. In nearly half of the stream pools, salmon survival during drought was similar to mean survival of pools assessed during non‐drought years, indicating some pools had remarkable resistance (ability to withstand disturbance) to extreme drought. Lower survival was most attributable to longer duration of disconnection between upstream and downstream habitats, a consequence of increasing drought severity. Our results not only suggest that many pools sustain juvenile salmon in non‐drought years transition into ecological traps during drought but also highlight that some pools serve as refuges even under extreme drought conditions. Projected increases in drought severity that lead to longer droughts and greater habitat fragmentation could transform an increasing proportion of suitable habitats into ecological traps. Predicting future impacts of drought on Coho salmon and other sensitive species will require identification and protection of drought refuges and management strategies that prevent further habitat fragmentation.


| INTRODUC TI ON
Droughts are an increasing threat to ecosystems worldwide, with unprecedented multi-year droughts recently observed in California, Australia, and South Africa (Robeson, 2015;van Dijk et al., 2013;Vogel & van Zyl, 2016). Harsh environmental conditions associated with droughts can lead to dramatic changes in the composition and abundance of biota, as well as species extirpations (e.g., Bogan & Lytle, 2011;Matusick, Ruthrof, & Hardy, 2012). However, the impact of droughts on ecosystems can vary greatly across the landscape (Nimmo, Haslem, Radford, Hall, & Bennett, 2016). Habitat refuges, which are defined as areas buffered from disturbance relative to their surroundings, can reduce the short-term impacts of drought on biota (Davis, Pavlova, Thompson, & Sunnucks, 2013;Keppel et al., 2012).
Therefore, accurate predictions of drought-induced changes in biota will ultimately depend on identifying drought refuges and assessing their ability to support species survival under more extreme drought conditions.
Human activities can induce ecological traps, or habitats that were once preferentially used by biota that no longer support growth and survival due to sudden environmental changes (Robertson & Hutto, 2006;Schlaepfer, Runge, & Sherman, 2002).
For example, juvenile African penguins follow environmental cues, such as water temperature, to productive feeding habitats; however, climate change and industrial fishing have depleted productivity in these habitats such that temperature cues are no longer correlated with food supplies (Sherley et al., 2017). This ecological trap has contributed to an ~80% reduction in the affected penguin population (Sherley et al., 2017). Biological communities in river ecosystems, such as fish, are also susceptible to ecological traps, especially when river flow regimes are altered by anthropogenic activities. For example, Coho salmon (Oncorhynchus kisutch) in the Shasta River of northern California successfully spawn in river reaches that now become inhospitable during the juvenile rearing period as a consequence of increased human water withdrawals.
These reaches have been transformed from suitable dry-season rearing habitat into ecological traps for this highly endangered fish population (Jeffres & Moyle, 2012). Yet, there remains relatively little empirical evidence on which environmental factors cause suitable habitat to become ecological traps for biota or of the population-level implications of such transformations (Hale & Swearer, 2016).
Residual pools in intermittent streams are a habitat that may be vulnerable to become ecological traps due to human activities, including water withdrawals and habitat modification. Intermittent streams, defined by their predictable cessation of flows during dry periods, are globally ubiquitous, and support a high diversity of freshwater and terrestrial species (Datry et al., 2018), including atrisk species (Labbe & Fausch, 2000;Wall, Berry, Blausey, Jenks, & Kopplin, 2004;Wigington et al., 2006). For example, intermittent and ephemeral streams comprise up to 66% of the river network in California (Levick et al., 2008) and are inhabited by endangered juvenile Coho salmon, a species of cultural and commercial importance (Lichatowich, 1999). Despite the harsh environmental conditions, Coho salmon growth and survival have been found to be higher in intermittent than perennial streams, presumably due to lower fish density and higher food resources (Wigington et al., 2006). During typical summer low-flow conditions, residual pools are important for juvenile salmon as they are the only suitable aquatic habitat remaining within intermittent stream reaches (Hwan & Carlson, 2016). However, extreme drought and human water withdrawals can have the potential to exacerbate stream drying, creating harsher environmental conditions and causing pools to dry.
In this study, we addressed the question of how extreme drought influences species survival in intermittent streams and identified the primary environmental factors controlling the occurrence of refuges and ecological traps. We used mark-recapture techniques to estimate over-summer survival of juvenile Coho salmon in intermittent streams during one of northern California's most severe droughts on record (Robeson, 2015). We measured pool-scale (i.e., habitat-unit scale) survival within four to eight stream reaches across four tributaries over 7 years (2011)(2012)(2013)(2014)(2015)(2016)(2017). We predicted that salmon survival during extreme drought would be reduced compared to survival during non-drought years. We also predicted that variability in survival across the study region would increase during drought due to the transformation of some pools into ecological traps (i.e., habitats with reduced survival during drought years) during extreme drought.
We then determined the relative importance of environmental variables on survival, including precipitation, landscape variables, water quality, streamflow, and physical habitat variables using a mixed modeling framework. We predicted that streamflow and habitat fragmentation would be the most influential variables in explaining variation in salmon survival.  precipitation occurs in the form of rainfall between November and April. This results in peak streamflow during the winter season that slowly recedes through the spring and summer, leading many small tributaries to cease to flow or dry completely. Vineyards and rural residential homes occupy much of the landscape, and streamflow is influenced by water withdrawals from small-scale direct diversions and streamside wells (Deitch, Kondolf, & Merenlender, 2009).

| Drought classification
The period 2012-2016 was an historic drought that affected most of California, including the study region. We described drought conditions in the Russian River catchment using data from the US Drought Monitor (https://droug htmon itor.unl.edu/), which classifies drought based on key indicators including streamflow, soil moisture, and precipitation.

| Study population
The Russian River catchment once supported a large, self-sustaining population of Coho salmon; however, due to habitat loss and degradation, the population was nearly extirpated by the

| Biological data collection
Each year, approximately 500 hatchery-reared juvenile Coho salmon (n = 19,666 total) implanted with passive integrated transponder (PIT) tags, were released into study reaches in mid-June and tracked using stationary and portable PIT-tag detection systems through early October. Stationary antennas were placed at reach boundaries to document fish moving outside of the study reach and portable antennas allowed us to detect movement within stream pools inside the study reaches. Depending on available resources, between four and eight reaches were stocked and surveyed each year. Fish were not stocked into reaches when observed spring habitat conditions indicated that they would have no chance of survival, such as when pools were already disconnected and drying in June. We revisited these unstocked reaches at the end of summer to confirm the absence of wetted habitat and validate the assumption that fish survival would have been zero.
To estimate the habitat-unit-specific survival over multiple intervals between June and October, we conducted a series of surveys using a portable PIT-tag detection method (O'Donnell, Horton, & Letcher, 2010). A total of three to five surveys were done approximately monthly during each summer, depending on reach and year. Each survey consisted of two site visits on consecutive days allowing estimation of abundance during each survey and survival between surveys while accounting for capture probability following the robust design mark-recapture model (Kendall, Nichols, & Hines, 1997;Obedzinski et al., 2018).

| Environmental data collection
Between June and October (2011-2017), we collected data on physical habitat, streamflow, water quality, precipitation, and landscape characteristics as potential explanatory variables for predicting over-summer survival. Within each habitat unit (i.e., stream pool), we measured physical habitat dimensions (depth, length, width, and volume), water temperature, and dissolved oxygen levels during monthly sampling occasions. Dissolved oxygen and temperature were measured using a handheld device consistently between 08:50 and 11:30 hr to minimize diel variation. We also measured water temperature continuously (60 min intervals) in one habitat-unit per reach, chosen to represent the physical and hydrologic characteristics of all units within the reach. Streamflow was measured approximately once per month to generate reachspecific rating curves based on correlation with continuous stage readings at loggers within each reach, following Rantz (1982).
From these streamflow data, we calculated summary statistics related to minimum, maximum and mean daily flow, and number of days of disconnection among habitat units, estimated to occur in these reaches when streamflow falls below 0.28 L/S (0.01 ft 3 /s; Obedzinski et al., 2018

| Habitat-unit-level fish survival estimates using mark-recapture
To estimate the survival between each sampling occasion in each habitat unit, we first used PIT-tag detections from paired wand surveys to construct an encounter history for each individual, and then applied the robust design mark-recapture model (Kendall et al., 1997) in program MARK (White & Burnham, 1999

| Assessing distribution of survival estimates during drought and non-drought years
We assessed the distribution of survival estimates from drought (2012-2016) and non-drought (2011 and 2017) years using probability density functions to account for non-normal distributions and to avoid binning data. We tested the equality of distributions of the two groups using the sm.density.compare function in the sm package in R (Bowman & Azzalini, 2014). Ecological traps were defined as pools that during drought had survival less than one standard deviation of the threshold value.

| Determining effects of explanatory variables on survival estimates
We selected 17 candidate explanatory variables (Table S1)  in the data (Harrison, 2015). We determined the effect of each explanatory variable on survival estimates by creating a nested model without that variable using a backward-stepwise regression procedure (Zuur et al., 2009). Each nested model was compared to the full model using chi-square tests of model residual deviances. During this process, we removed non-significant variables (p > .05) from the full model and continued testing individual variable effects by comparing nested models and full models until all non-significant variables were removed (Zuur et al., 2009). We verified the underlying model assumptions and assessed multicollinearity of explanatory variables using the variable inflation factor (vif) function in the R package car (Fox & Weisberg, 2011). We extracted estimated coefficients and calculated 95% confidence intervals using Wald's method from GLMMs as measure of variable effect size. We assessed model spatial and temporal transferability using a non-random, k-fold cross-validation procedure (Wenger & Olden, 2012; see Appendix S4 for more details).

| Cumulative juvenile Coho salmon survival
Mean cumulative survival, averaged across sites and years, during the study was 0.51 ± 0.29 (mean ± SD estimated proportion of fish surviving to the end of the dry season; Figure 3). Survival in pools that were stocked with juvenile Coho salmon during the study period was 0.53 ± 0.26 during non-drought years were identified as drought refuges, pools with survival in drought years that was within or above the system-wide range of survival in

| Effects of explanatory variables on juvenile salmon survival
Our models indicated that days of disconnection had the greatest influence on over-summer survival, exhibiting a negative effect on cumulative salmon survival (

| Key drivers of juvenile salmon survival
Habitat fragmentation is a primary driver of population-level drought impacts in both terrestrial and freshwater ecosystems (Hwan & Carlson, 2016;Oliver et al., 2015;Perkin, Gido, Costigan, Daniels, & Johnson, 2014). For example, Oliver et al. (2015) found a strong positive association between woodland fragmentation and the sensitivity of the ringlet butterfly (Aphantopus hyperantus) to an extreme drought in the UK. In fragmented woodlands, populations were less likely to locate refuges with adequate resources during the drought and thus experienced greater losses and slower recovery (Oliver et al., 2015). Similarly, we found that increased duration of stream pool disconnection negatively influenced pool-scale juvenile salmon survival. Once disconnected, movement of individuals among pools could no longer occur, preventing salmon from relocating to pools that may have had more suitable environmental conditions as drought conditions worsened over the summer. As suggested by Obedzinski et al. (2018), habitat fragmentation is a "master variable" that encompasses the effects of other variables that potentially limit survival, including water quantity, water quality, food availability, competition, and predation. Although days of disconnection does not provide insight into the proximate cause of salmon mortality in F I G U R E 6 Partial dependence from generalized linear mixed models of juvenile salmon survival. Model estimates (solid lines) and 95% confidence intervals (shading) for days of disconnection (a), cropland area (b), and minimum pool volume (c  & Post, 2008) and community responses (e.g., Ledger, Brown, Edwards, Milner, & Woodward, 2013) to river drying. The influence of biotic factors such as competition from other fish (e.g., steelhead trout, Harvey & Nakamoto, 1996) and predation by avian species (Spalding, Peterson, & Quinn, 1995) and mammals (e.g., river otter, Dolloff, 1993) could also be tested using enclosures in a field-based experimental approach. Moreover, manipulative experiments that allow for testing interactions among biotic and abiotic variables could provide insight into their synergistic or antagonistic effects on populations (Vander Vorste et al., 2017). Interaction between landuse and habitat fragmentation represents a particularly important variable to test considering that we found a negative association between cropland land-use and survival, even at <10% cropland area.
In our study, we did not quantify interactions among variables owing to sample size limitations; however, we see this as an important next step to identify the drivers of salmon survival.

| Variability in survival
High variability in survival within and across study reaches was the most striking results of this study. Variability in survival more than doubled in drought years compared to non-drought years, highlighting that extreme drought did not uniformly affect habitats across the study area. Some reaches maintained pools with relatively consistent survival estimates (e.g., MIL upper) throughout the study period (CV = 22%), whereas other reaches (e.g., GRP lower) experienced much greater survival variability among pools (CV = 352%; Figure 4).
These patterns in survival variability are potentially explained by the high landscape heterogeneity within our study area, which has also been shown to explain variation in species persistence in other ecosystem types (Godfree et al., 2011;Schwantes et al., 2018).
Heterogeneity in physical catchment characteristics is a defining feature of Mediterranean climate regions (Cid et al., 2017). Indeed, adjacent catchments within our broader study region have remarkably distinct hydrologic characteristics owing to differences in lithology that influence subsurface water storage (Dralle et al., 2018) and oversummer streamflow conditions (Larsen & Woelfle-Erskine, 2018).
However, there are currently no regional data available that describe variation in catchment lithology in relation to catchment hydraulics (water storage and runoff properties) and field methods to characterize local substrate and subsurface hydrologic properties within streams remain time-and resource-intensive. As the understanding of catchment and hydrology advances, we expect that the ability to predict summer low-flow conditions, and therefore salmon survival, will also improve. In the meantime, however, the limited transferability of our model in predicting survival at individual study reaches or years emphasizes that assessing drought impacts on regional populations of sensitive species will remain a challenging task (Moritz & Agudo, 2013;Thuiller et al., 2004).

| Drought refuges and ecological traps
Until recently, intermittent streams have been neglected as ecologically important components of freshwater ecosystems (Datry et al., 2018;Marshall et al., 2018); however, our results highlight their importance as drought refuges for an endangered species.
Despite unprecedented drought conditions, we estimated that nearly half of the pools were refuges during drought years because they had similar survival to pools assessed in non-drought years.
This result, along with previous findings of increased growth and survival of juvenile Coho salmon in intermittent compared to perennial streams (Wigington et al., 2006) Labbe & Fausch, 2000), perhaps because they encounter less competition and/or predation compared to perennial streams.
Our results emphasize the value of these unique habitats as drought refuges but also underscore their vulnerability to future global change.
For freshwater rivers, increases in flow intermittency and drying associated with extreme drought, in combination with human water demand, likely play an important role in transforming once-suitable pool habitats into ecological traps. Within our studies sites, we observed anecdotally that pools with alluvial substrate were more likely to fragment and dry and had lower survival compared to pools underlain by bedrock or clay, similar to findings of May and Lee (2004) in Oregon streams. Stream habitats with alluvial substrate are commonly selected for spawning because they contain gravel; however, these habitats are particularly sensitive to water withdrawals from diversions and groundwater pumping, increasing the risk of dewatering redds and stranding juvenile fish (Reiser & White, 1983).
Adult salmon preferentially spawn in alluvial substrates where their offspring are more prone high water temperatures and reduced low flows resulting from agricultural water withdrawals (Jeffres & Moyle, 2012

| Management implications
A continued rise in the number of extreme weather events will no doubt intensify threats to biodiversity (IPCC, 2012) and predicting and mitigating biodiversity loss remains a central challenge to species conservation efforts (Morelli et al., 2017;Thuiller et al., 2004). Our results demonstrate that accurately predicting drought impacts on a species will require going beyond regional-scale climatological assessments of drought severity. In our case, the Russian River catchment experienced extreme to exceptional drought conditions between 2014 and 2015; however, salmon survival was not uniformly affected by drought and local environmental variables were more influential than watershed-scale climate and physical features. The type of long-term, intensive survival studies performed here are expensive and difficult to implement; thus, indicators or proxy variables are needed for assessing drought and climate-change impacts on sensitive species. Wet-dry mapping has shown to be a particularly cost-efficient and effective method for assessing spatial patterns of stream habitat fragmentation (Hwan & Carlson, 2016) and readily lends itself to citizen-science data collection efforts (Turner & Richter, 2011). As the spatial and temporal extent of such data grows, wet-dry mapping also holds promise for gaining new insights into where drought ref- uges and ecological traps occur on the landscape, and the underlying physical mechanisms that influence their persistence.
Developing predictive relationships between widely available metrics (e.g., climatic, geological, land-use) and the occurrence of habitat fragmentation in space and time could then prove a low-cost means of identifying areas of impairment at a much larger scale, allowing resource managers to better focus limited resources for species recovery.

ACK N OWLED G EM ENTS
We thank field crews, hatchery staff, interns, and Watershed Stewards Project AmeriCorps members for their assistance with hatchery releases and data collection. We appreciate Matt Deitch and Mia vanDocto for providing their streamflow datasets, William Boucher for operating the PIT-tag detection systems, Andrew Bartshire for providing spatial data, and Zachary Reinstein for creating Figure 1. Andrew McClary was helpful in accessing raw data used in this study, and Gregg Horton provided valuable support in developing the robust design models. We are grateful to the many landowners who provided us access to the streams on their properties.
Funding support for this study was provided by the National Fish and Wildlife Foundation, the US Army Corps of Engineers, and the California Wildlife Conservation Board. We also thank four anonymous reviewers for their insightful comments that greatly improved this manuscript.

DATA AVA I L A B I L I T Y S TAT E M E N T
R code and data from this study are available upon request to M.O. (mobedzinski@ucsd.edu).