Migratory species are particularly vulnerable to climate change because habitat throughout their entire migration cycle must be suitable for the species to persist. For migratory species in rivers, predicting climate change impacts is especially difficult because there is a lack of spatially-continuous and seasonally-varying stream temperature data, habitat conditions can vary for an individual throughout its life cycle, and vulnerability can vary by life stage and season. To predict thermal impacts on migratory riverine populations, we first expanded a spatial stream network model to predict mean monthly temperature for 465,775 river km in the western U.S., and then applied simple yet plausible future stream-temperature change scenarios. We then joined stream temperature predictions to 44,396 spatial observations and life stage-specific phenology (timing) for 26 ecotypes (i.e. geographically distinct population groups expressing one of four distinct seasonal migration patterns) of Chinook salmon (Oncorhynchus tshawytscha), a phenotypically diverse anadromous salmonid that is ecologically and economically important but declining throughout its range. Thermal stress, assessed for each life stage and ecotype based on federal criteria, was influenced by migration timing rather than latitude, elevation, or migration distance, such that sympatric ecotypes often showed differential thermal exposure. Early-migration phenotypes were especially vulnerable due to prolonged residency in inland streams during the summer. We evaluated the thermal suitability of 31,699 stream km which are currently blocked by dams to explore reintroduction above dams as an option to mitigate the negative effects of our warmer stream temperature scenarios. Our results showed that negative impacts of stream temperature warming can be offset for almost all ecotypes if formerly occupied habitat above dams is made available. Our approach of combining spatial distribution and phenology data with spatially- and temporally-explicit temperature predictions enables researchers to examine thermal exposure of migrating populations that use seasonally-varying habitats.

Monthly stream temperature

Seasonal stream temperature is essential for modelling differential vulnerability of contrasting ecotypes. We expanded a pre-existing spatial stream network (SSN) model, currently only available for summer months, to all months of the year in order to have a full picture of year-round thermal habitat. SSN models require a stream network, observed water temperatures at discrete locations, and spatially and/or temporally explicit covariates. This model uses the National Stream Internet (NSI) network, which was derived from the NHDPlus dataset and prepared for use with SSNs. We included observed stream temperatures from 1993 on, with the end year varying by region (2011-2015). Within each month, we averaged observed temperatures for sites that had multiple observations within a day and observations on at least 90% of days; we then used this monthly mean for the model. Mean monthly temperature was modeled as a function of ten spatial and two temporal covariates. The spatial covariates are elevation ('ELEV'; m), canopy ('CANOPY'; %), slope ('SLOPE'; m/m), annual precipitation ('PRECIP'; mm), cumulative drainage area ('CUMDRAINAG'; km²), North American Albers northing coordinate ('Y_COORD'; m), upstream watershed area that is lake or reservoir ('NLCD11PC'; %), the amount of flow that is base flow ('BFI'; %), upstream watershed area that is glacier ('GLACIER'; %), and tailwater ('TAILWATER'; binary – 0, 1). Each covariate was spatially linked with the stream network at a 1 km interval. The temporal covariates are historical air temperature (°C) and flow (m³/s). We filled in gaps in the time series using an iterative PCA approach, then calculated monthly means and linked values to water temperature observations by year. Mean monthly temperature was modelled by fitting the SSN linear mixed model for each month in each of 8 sub-regional watersheds, spanning most of California, Oregon, and Washington. See our paper for additional modeling details.

To validate each monthly model for each sub-regional watershed, we randomly split the water temperature data into a training dataset and a testing dataset based on spatial location, such that approximately 80% of the data were used for model fitting and 20% of the data were used for model validation. When a site had data in multiple years, we assigned all data for that site to either the training or testing dataset rather than being split between them so that our out-of-sample metrics would be an estimate of how the model performed in areas where we had no data. We performed leave-one-out cross validation on the training dataset and calculated three performance metrics on each training dataset and testing dataset: the square of the correlation coefficient between observations and predictions (r²), the root mean square prediction error (RMSPE), and the mean absolute prediction error (MAPE). We then used the model to predict mean monthly water temperature at 1 km resolution for the period 2002-2011. The years 2002-2011 represent the most recent time period when most temperature records were collected. Predictions used the universal kriging equation which accounts for both the model predictors and spatial autocorrelation. Finally, we assigned temperatures in reaches representing manmade lakes and reservoirs ('FTYPE' = "ArtificialPath" & 'WATERBODY' = 1) as "-9999"; these were not included in our analyses.

Chinook salmon spatial distribution dataset

We defined the spatial distributions of freshwater life stages for each Chinook salmon ecotype. Chinook salmon consist of federally recognized Ecologically Significant Units (ESU) based on geography and genetic relatedness at neutral markers, but an ESU can contain multiple run types, the trait that specifies the peak seasonal timing (i.e. spring, summer, fall, winter) of adult migration into freshwater and is used most frequently to define populations. Chinook salmon observations were extracted from eight field-observational and distributional data sources (1. Aquatic Species Observation Database, obtained via California Dept. of Fish and Wildlife; 2. http://www.calfish.org/DataandMaps/CalFishDataExplorer.aspx; 3. www.gbif.org; 4. www.iobis.org; 5. https://www.streamnet.org/; 6. https://www.epa.gov/national-aquatic-resource-surveys/data-national-aquatic-resource-surveys; 7. https://aquatic.biodata.usgs.gov; 8. http://vertnet.org) from August 2017-January 2018. We divided linear distribution datasets (StreamNet and CalFish) into points separated by 1 km and then added the points to the observational dataset. Observations have a georeferenced accuracy to 500 m, and each observation was linked with the appropriate ecotype, life stage, and month. We removed fish of unknown ecotype, unknown life stage, or known hatchery origin. To match the temporal limit of the stream temperature modeling project, we removed observations prior to 1993. We specified the ecotype and/or ESU (if not specified) by spatially merging observation locations with ESU distribution shapefiles (http://www.westcoast.fisheries.noaa.gov/maps_data/Species_Maps_Data.html). Due to low numbers and ambiguity of other life stages (e.g. “juvenile” could indicate rearing or outmigration), we focused our efforts on observations listing spawning location, such as those based on redd counts or spawner surveys. We supplemented our dataset with additional spawning and redd observations from the upriver bright fall-run ecotype in the ~90 km Hanford Reach of the Columbia River, digitized from several references. We removed duplicate observations from the entire spatial distribution dataset using R. Finally, we removed observations that georeferenced > 500m from a stream. Ecotypes with fewer than 50 observations were not examined. This filtering resulted in a total of 44,396 spawning site observations. We defined pre-spawn holding habitat from spawning and redd locations. For additional details and a map figure, see the published paper. 

Attached are files for data reproduction. The stream temperature modeling results are attached as 8 shapefiles (.shp), corresponding to each sub-regional watershed. The Chinook salmon spatial distribution dataset and associated data is attached as a .csv. See associated README_StreamTemp and README_Chinook files.