MSI: Forecasting Pink Salmon Harvest in Southeast Alaska
Auke Bay Laboratories (ABL, NOAA) has been using juvenile pink salmon catch and associated environmental data to forecast adult pink salmon harvest in Southeast Alaska (SEAK). The Alaska Department of Fish and Game (ADF&G) is also incorporating juvenile pink salmon catch data from ABL into their forecast model; however, rather than forecasting from a direct relationship between harvest and juvenile catch, ADF&G uses the juvenile data to modify an exponential smoothing model of past harvests. The 2008 pink salmon commercial harvest forecasts (with 80% confidence intervals) from both ABL and ADF&G are shown above.
|
Purse seiners harvesting pink salmon |
| |
|
Juvenile and adult pink salmon captured during SECM sampling |
| |
|
Fig. 1. Percent juvenile pink salmon surviving after 40 days at sea |
Pink salmon returns are notoriously difficult to forecast, because their two-year life history cycle precludes the use of returns of younger age classes to predict cohort abundance. Pre-season forecasts using spawner/recruit relationships modified by indices of environmental conditions have not been very useful, because escapement data are imprecise and high variation in freshwater and marine survival makes parental abundance a weak predictor of returns.
Understanding factors that influence survival of juvenile salmon in marine waters is a major objective of the Southeast Alaska Coastal Monitoring (SECM) project. Since 1997, SECM has collected a time series of indexes of juvenile salmon abundance and associated biophysical data in neritic habitats of northern SEAK.
Direct measures of the abundance of juvenile pink salmon may provide a method to improve forecasts. Mortality of juvenile pink salmon is high and variable during their initial marine residency (Fig. 1), and is considered a major determinant of year-class strength. Sampling juveniles after this period of high initial mortality may provide information that can be used with environmental and other data time series to forecast subsequent harvests.
To forecast SEAK pink salmon harvest, we used the SECM juvenile catch and biophysical data and the ADF&G commercial catch data. For 1998 through 2007, the harvest of pink salmon in SEAK was highly correlated with the peak average catch per unit effort (PeakCPUE) of juvenile pink salmon sampled by SECM the previous year in northern SEAK (Fig. 2) strait habitat. PeakCPUE is the average (CPUE+1) for either June or July, using the month with the highest average per year. Other biophysical parameters by themselves, including data on growth and environmental conditions collected by SECM, and large-scale climate indexes (PDO, El Niño), were not significantly correlated with the pink salmon harvest data.
|
Fig. 2. Correlation of SEAK pink salmon harvest and PeakCPUE of juveniles |
| |
|
Fig. 3. Actual harvest compared to forecast (with 80% CIs) from CPUE forecast models |
Linear models using the relationship between harvest and prior year juvenile CPUE were used to forecast pink salmon harvests in SEAK from 2004-2007. Step-wise multiple regression determined if environmental parameters significantly improved model fit. The Akiake Information Criterion (AICc) was used to test for over-parameterization of multiple regression models, and “Jackknife” hindcasting and bootstrap confidence intervals were used to evaluate model performance and choose the “best” model. For 2004-2006, the simple linear relationship between the PeakCPUE and harvest was selected as the best model. In 2007, a multiple regression model including May sea surface temperature data as well as PeakCPUE was selected as the best model.
The forecast model performed very well for 2004, 2005, and 2007 harvests, giving estimates within 12% of actual harvests (Fig. 3). However, in 2006 the actual harvest was well below the forecast, and was the weakest return of pink salmon to SEAK since 1988. Although the juvenile CPUE did indicate a low return for 2006 relative to recent year average harvests, the poor performance of the single-parameter CPUE model in 2006 shows that juvenile CPUE alone will not provide consistently accurate harvest forecasts.
To forecast 2008 returns, we evaluated PeakCPUE models incorporating juvenile CPUE data, associated biophysical parameters, and harvest data through 2007; the forecast from each model was then projected from the 2007 juvenile CPUE data. We also considered August CPUE as an auxiliary model parameter that could indicate delayed migration or anomalous distribution (see last page for table of all parameters considered for the models, and their correlation with harvest).
PeakCPUE was the only parameter that had a significant bivariate correlation with harvest. In the stepwise regression analysis, a 4-parameter model including CPUE, June 20-m temperature, El Niño Southern Oscillation (ENSO), and July 3-m salinity explained 99% adjusted R²), of the variability in the harvest data as compared to 80% for the simple linear regression with Peak CPUE. The 2-parameter model included May 20-m temperature, similar to the 2007 forecast model, and increased the adjusted R² to 93% (see table below). When ENSO entered the model, the May temperature parameter was replaced by June temperature. The AICc decreased at each model step, and was lowest for the 4-parameter model, indicating that this model is also the most parsimonious.
|
Fig. 4. Interannual catches (average CPUE) of juvenile pink salmon by month in northern SEAK |
| |
|
Fig. 5. Effect of process and measurement error on 2008 forecasts using juvenile CPUE models |
In 2007, the CPUE for juvenile pink salmon was the lowest over the 11-year time series for juvenile salmon sampling in northern Southeast Alaska (Fig. 4). As a result, all of the CPUE models indicated a poor return to SEAK in 2008.
The performance of four CPUE models was evaluated by a process that included: 1) using jackknife analyses (drop one year of data, forecast that year from remaining years); 2) forecasting the 2005-2007 harvests using truncated data series; and, 3) comparing bootstrap and parametric regression confidence intervals (CIs) (Fig. 5). We selected the 2-parameter model (PeakCPUE + May20-m temperature) forecast of 16.1 million as the “best” forecast for 2008 for the following reasons:
- The model explains over 90% of the variation (R² = 0.93);
- Bootstrap confidence intervals are narrower than for 3- or 4- parameter models;
- Truncated model performance of the 2-parameter model is similar to 3-parameter model;
- The 2-parameter model forecast is intermediate between 3- and 4- parameter models.
There are certainly outstanding questions on the validity and application of the juvenile catch data for forecasting year-class strength. Is harvest an appropriate proxy for total return? Is the scope of spatial and temporal coverage adequate? Is variation in marine survival after the early marine periods causing the “misses”, for example, in the 2006 return? Can sufficient auxiliary biological and environmental data be incorporated to explain the variation in “downstream” survival? We will be examining these issues as long as the juvenile CPUE models show promise for forecasting pink salmon abundance and for improving our understanding of factors affecting survival in the marine ecosystem. Ultimately, the utility and continued development of these forecast models will be determined by their performance over a number of years.
Contact:
Alex Wertheimer
Auke Bay Laboratories
Alaska Fisheries Science Center, NOAA Fisheries
Ted Stevens Marine Research Institute
17109 Pt Lena Loop Rd
Juneau AK 99801
(907) 789-6040
Alex.Wertheimer@noaa.gov
|
