Northwest Fisheries Science Center

Forecast of Adult Returns for coho salmon and Chinook Salmon

Similar to 2015, many of the ocean ecosystem indicators suggest 2016 to be another poor year for juvenile salmon survival, the second consecutive year of poor ocean conditions. The PDO was strongly positive (warm) throughout 2016, coinciding with the continuing of the anomalously warm ocean conditions in the NE Pacific initiated by the “The Blob” that began in the fall of 2013. Strong El Niño conditions at the equator also persisted throughout 2015 until May of 2016. Sea surface and upper 20 m water temperatures off Newport Oregon remained warmer than usual (+2°C) throughout most of 2016 continuing two consecutive years of anomalously warm ocean conditions. The zooplankton community remained in a lipid-deplete state throughout 2016, with the lowest biomass of lipid rich northern copepods and the highest species richness observed in the 19 year time series. Our annual summary of ecosystem indicators during 2015 is here, and our "stoplight" rankings and predictions are shown below in Table SF-01, Table SF-02, and Figure SF-01.

Ocean ecosystem indicators of the Northern California Current. Colored squares indicate positive (green), neutral (yellow), or negative (red) conditions for salmon entering the ocean each year. In the two columns to the far right, colored dots indicate the forecast of adult returns based on ocean conditions in 2016 (coho salmon) and 2015 (Chinook salmon).

Table SF-02 Rank scores derived from ocean ecosystem indicators data found in Table SF-03 and color-coded to reflect ocean conditions for salmon growth and survival (green = good; yellow = intermediate; red = poor). Lower numbers indicate better ocean ecosystem conditions, or "green lights" for salmon growth and survival. To arrive at these rank scores for each ocean ecosystem indicator, all years of sampling data from Table SF-03 were compared (within each row).

Table SF-03.   Data for rank scores of ocean ecosystem indicators.

Data for rank scores of ocean ecosystem indicators. Click HERE to download the data as a *.csv.

Figure SF-01 shows correlations between adult Chinook salmon counts at the Bonneville Dam and coho salmon smolt to adult survival (%) (PFMC 2016a) versus a simple composite integrative indicator – the mean rank of all the ecosystem indicators (the second line from the bottom) in Table SF-02. This index explains about 50% of the variance in adult returns. A weakness of this simple non-parametric approach is that each indicator is given equal weight, an assumption that may not be true. Therefore, we are exploring a more quantitative analysis of the ocean indicators shown in Table 3, using principal component analysis (PCA).

Scatter plot showing relationships between ocean indicators and counts of adult salmon at Bonneville Dam. Figure SF-01.  Salmon returns versus the mean rank of ecosystem indicators. Arrows show the forecasted returns for Chinook salmon in 2017 (solid line) and 2018 (dashed line). The mean rank of the ocean ecosystem indicators was tied at 15.25 in 2015 and 2016, forecasting a return of 70,600 and 227,700 adult spring and fall Chinook salmon to Bonneville Dam respectively in 2017 and 2018 (top two panels). Using the rank of 15.25 from 2016, the forecast of the smolt to adult survival of coho salmon to Oregon coastal streams is 1.7 percent in 2017. However, the relationship between the ocean ecosystem indicators and coho salmon survival was not very strong (R2 = 0.32), so this forecast should be used with caution.

Principal component analysis (PCA) was run on the indicator data. This procedure reduces the number of variables in the dataset as much as possible, while retaining the bulk of information contained in the data (a sort of weighted averaging of the indicators). Another important feature of PCA is that the principal components (PCs) are uncorrelated. This eliminates one of the original problems with the indicator data set (i.e., multi co-linearity). The first principal component (PC1) explains 55% of the ecosystem variability among years while the second principal component explains only 12%. Therefore, PC1 is used as a new predictor variable in a linear regression analysis of adult salmon returns (this process is termed principal component regression, or PCR) and those results are shown below in Figure SF-02.

Salmon returns versus the axis 1 scoreFigure SF-02. Salmon returns versus the first principal axis scores (PC1) from a Principal Component Analysis on the environmental indicators from Table SF-02.

*outliers were excluded using Cook's Distance

In addition to correlating PC1 with salmon returns, we incorporated this metric into a more formal modeling structure. Specifically, we used sibling regression and dynamic linear modeling (DLM; Scheuerell & Williams 2005) to relate PC1 to returns. DLMs are similar to linear regression, but allow the regression coefficient(s) to vary over time, effectively allowing for a shift in the magnitude of response to ocean conditions. In all models, we allowed the coefficients for siblings and PC1 to vary, but kept a constant model intercept.

The best model for both spring and fall Chinook salmon showed support for a dynamic effect of jack counts, but not of PC1. For coho salmon, there was no support for any parameter to vary, resulting in a simple linear regression model (of logit-transformed SAR).

Figure SF-03. Time series of observed spring Chinook salmon adult counts (top), fall Chinook salmon adult counts (middle), and coho salmon SAR (bottom) by out-migration year. In each plot, the dark line represents the model fit and lighter lines represent plus or minus 2 standard deviations from the mean. Forecasts were created from a DLM model with log of sibling counts (for the Chinook models only) and PC1 as predictor variables.

Although the PCA and DLM analyses represent a general description of ocean conditions, we must acknowledge that the importance of any particular indicator will vary among salmon species/runs. We are working towards stock-specific salmon forecasts by using methods that can optimally weight the indicators for each response variable in which we are interested (Burke et al. 2013).