Patterns of Synchrony and Environmental Thresholds in the Performance of Forecast Models Used for U.S. West Coast Chinook and Coho Salmon Stocks (original) (raw)
Related papers
Ecological thresholds in forecast performance for key United States West Coast Chinook salmon stocks
ICES Journal of Marine Science, 2019
Preseason abundance forecasts drive management of US West Coast salmon fisheries, yet little is known about how environmental variability influences forecast performance. We compared forecasts of Chinook salmon (Oncorhynchus tshawytscha) against returns for (i) key California-Oregon ocean fishery stocks and (ii) high priority prey stocks for endangered Southern Resident Killer Whales (Orcinus orca) in Puget Sound, Washington. We explored how well environmental indices (at multiple locations and time lags) explained performance of forecasts based on different methods (i.e. sibling-based, production-based, environment-based, or recent averages), testing for nonlinear threshold dynamics. For the California stocks, no index tested explained >50% of the variation in forecast performance, but spring Pacific Decadal Oscillation and winter North Pacific Index during the year of return explained >40% of the variation for the sibling-based Sacramento Fall Chinook forecast, with nonlinea...
Integrated forecasts of fall Chinook salmon returns to the Pacific northwest
Fisheries Research, 2012
Forecasts of returns are used to determine appropriate levels for ocean and in-river harvests of northeast Pacific fall Chinook salmon (Oncorhynchus tshawytscha). The current preseason forecasts are made by February, six months before fish arrive at natal rivers. As such, they do not capture ecosystem changes that occur after February. We incorporated catch and effort data from the Alaska summer commercial fishery to detect unusual ecosystem variability in fish return years and to update forecasts during the ocean fishery season. In addition, the traditional model using the ordinary regression unnecessarily has an intercept term, which often leads to unusual over-forecasts. Further there are concerns about autocorrelation in Chinook salmon returns. Finally the current practice does not provide information about the uncertainty inherent in the forecasts. To address these issues, we developed new forecast models and compared them with the traditional method. The models considered are: I, traditional preseason forecasts; II, autoregressive preseason forecasts without intercept term; III, ocean fishery real-time forecasts; IV, integrated forecasts from models I and III; and V, integrated forecasts from models II and III. We validated our models, using data from three Columbia River stocks [Upriver Bright (URB), Lower River Wild (LRW), and Mid-Columbia Bright (MCB)], and from two Canadian stocks [Fraser River Lates (FRL) and West Coast Vancouver Island (WCVI)].
An Improved Sibling Model for Forecasting Chum Salmon and Sockeye Salmon Abundance
North American Journal of Fisheries Management, 2007
The sibling model is often one of the best methods for calculating preseason forecasts of adult return abundance (recruits) for populations of Pacific salmon Oncorhynchus spp. This model forecasts abundance of a given age-class for a given year based on the abundance of the previous age-class in the previous year. When sibling relations fit historical data well, the sibling model generally performs better than other forecasting methods, such as stockrecruitment models. However, when sibling relations are weak, better forecasts are obtained by other models, such as naïve models that simply use an historical average. We evaluated the performance of a hybrid model that used quantitative criteria for switching between a sibling model and a naïve model when generating forecasts for 21 stocks of chum salmon O. keta and 37 stocks of sockeye salmon O. nerka in the northeastern Pacific Ocean. Compared with the standard sibling model, the hybrid model reduced the root mean square error (RMSE) of forecasts by an average of 27% for chum salmon stocks and 28% for sockeye salmon stocks. Compared with a naïve model, the hybrid model reduced the RMSE of forecasts by an average of 16% for chum salmon stocks and 15% for sockeye salmon stocks. Our results suggest that hybrid models can improve preseason forecasts and management of these two species.
Marine environment-based forecasting of coho salmon (Oncorhynchus kisutch) adult recruitment
Fisheries Oceanography, 2012
Generalized additive models (GAMs) were used to investigate the relationships between annual recruitment of natural coho salmon (Oncorhynchus kisutch) from Oregon coastal rivers and indices of the physical ocean environment. Nine indices were examined, ranging from large-scale ocean indicators, e.g., Pacific Decadal Oscillation (PDO), to indicators of the local ecosystem (e.g., coastal water temperature near Charleston, OR). Generalized additive models with two and three predictor variables were evaluated using a set of performance metrics aimed at quantifying model skill in short-term (approximately 1 yr) forecasting. High explanatory power and promising forecast skill resulted when the spring ⁄ summer PDO averaged over the 4 yr prior to the return year was used to explain a low-frequency (multi-year) pattern in recruitment and one or two additional variables accounted for year-to-year deviations from the low-frequency pattern. More variance was explained when averaging the predictions from a set of models (i.e., taking the ensemble mean) than by any single model. Making multiple forecasts from a set of models also provided a range of possible outcomes that reflected, to some degree, the uncertainty in our understanding of how salmon productivity is driven by physical ocean conditions.
Retrospective Evaluation of Preseason Forecasting Models for Sockeye and Chum Salmon
Http Dx Doi Org 10 1577 M06 287 1, 2011
Using comprehensive data sets for chum salmon Oncorhynchus keta (40 stocks) and sockeye salmon O. nerka (37 stocks) throughout their North American ranges, we compared the retrospective performance of 11 models in preseason forecasting of adult abundance. Chum and sockeye salmon have more complicated age structures than pink salmon O. gorbuscha, which we investigated previously (Haeseker et al. 2005), and this complexity presents new challenges as well as opportunities for forecasting. We extended our previous work to include two new forecasting models that make use of leading indicators: either the survival rate of earlier-maturing pink salmon from the same brood year (Ricker pink salmon index model) or the abundance of earlier-maturing siblings (hybrid sibling model, a new version of the standard sibling model). No single forecasting model was consistently the best for either chum or sockeye salmon, but the hybrid sibling model frequently performed best based on mean absolute error, mean percent error, and root mean square error. As was observed for pink salmon, several naïve models (i.e., simple time series models without explicitly modeled mechanisms) also performed well, as did forecast averaging models composed of two models with the least-correlated forecasting errors. In general, model ranking depended on the particular stock and performance measure used. However, even the top-ranked model for each stock explained on average only 21% of the observed interannual variation in chum salmon recruitment and only 36% of the variation in sockeye salmon recruitment. Although improvements may be possible for some stocks in specific circumstances, a major breakthrough in general forecasting ability seems unlikely given the breadth of stocks and models examined to date. Therefore, better in-season updates and adjustments to fishing regulations and a cautious approach to opening and closing fisheries should remain high priorities.
Retrospective Evaluation of Preseason Forecasting Models for Pink Salmon
North American Journal of Fisheries Management, 2005
Using comprehensive data sets for chum salmon Oncorhynchus keta (40 stocks) and sockeye salmon O. nerka (37 stocks) throughout their North American ranges, we compared the retrospective performance of 11 models in preseason forecasting of adult abundance. Chum and sockeye salmon have more complicated age structures than pink salmon O. gorbuscha, which we investigated previously , and this complexity presents new challenges as well as opportunities for forecasting. We extended our previous work to include two new forecasting models that make use of leading indicators: either the survival rate of earlier-maturing pink salmon from the same brood year (Ricker pink salmon index model) or the abundance of earlier-maturing siblings (hybrid sibling model, a new version of the standard sibling model). No single forecasting model was consistently the best for either chum or sockeye salmon, but the hybrid sibling model frequently performed best based on mean absolute error, mean percent error, and root mean square error. As was observed for pink salmon, several naïve models (i.e., simple time series models without explicitly modeled mechanisms) also performed well, as did forecast averaging models composed of two models with the least-correlated forecasting errors. In general, model ranking depended on the particular stock and performance measure used. However, even the top-ranked model for each stock explained on average only 21% of the observed interannual variation in chum salmon recruitment and only 36% of the variation in sockeye salmon recruitment. Although improvements may be possible for some stocks in specific circumstances, a major breakthrough in general forecasting ability seems unlikely given the breadth of stocks and models examined to date. Therefore, better in-season updates and adjustments to fishing regulations and a cautious approach to opening and closing fisheries should remain high priorities.
North Pacific Anadromous Fish Commission Bulletin, 2016
Researchers in Alaska have provided accurate pre-season annual forecasts of pink salmon harvest to resource stakeholders of Southeast Alaska (SEAK) in times of climate change. Since 1997, the Southeast Alaska Coastal Monitoring project has collected biophysical data associated with seaward migrating juvenile salmon from May to August, and it has used these data along with larger basin-scale indexes to forecast SEAK pink salmon returns via regression and ecosystem metric models. In nine of the past eleven years (2004-2014), predictions from linear regression models ranged 0-17% of actual harvests, an average absolute forecast deviation of 10%. The primary explanatory variable was juvenile pink salmon peak catch. In some years, models included secondary variables to improve fi t. A supplemental modeling approach was tested recently to better inform stakeholders. This approach incorporated both an annual rank score forecast outlook based on ecosystem metrics, and a visual stoplight color-code graphic. Accurate pre-season salmon forecasts and descriptive outlooks from this applied research has increased economic effi ciency of the fi sh processing industry, enabled managers and resource stakeholders to anticipate harvest with more certainty, helped promote resource sustainability, and provided insight into ecosystem mechanisms related to pink salmon production in a changing climate.
Dynamical seasonal ocean forecasts to aid salmon farm management in a climate hotspot
Marine aquaculture businesses are subject to a range of environmental conditions that can impact on day to day operations, the health of the farmed species, and overall production. An understanding of future environmental conditions can assist marine resource users plan their activities, minimise risks due to adverse conditions, and maximise opportunities. Short-term farm management is assisted by weather forecasts, but longer term planning may be hampered by an absence of useful climate information at relevant spatial and temporal scales. Here we use dynamical seasonal forecasts to predict water temperatures for south-east Tasmanian Atlantic salmon farm sites several months into the future. High summer temperatures pose a significant risk to production systems of these farms. Based on twenty years of historical validation, the model shows useful skill (i.e., predictive ability) for all months of the year at lead-times of 0-1 months. Model skill is highest when forecasting for winter months, and lowest for December and January predictions. The poorer performance in summer may be due to increased variability due to the convergence of several ocean currents offshore from the salmon farming region. Accuracy of probabilistic forecasts exceeds 80% for all months at lead-time 0 months for the upper tercile (warmest 33% of values) and exceeds 50% at a lead-time of 3 months. This analysis shows that useful information on future ocean conditions up to several months into the future can be provided for the salmon aquaculture industry in this region. Similar forecasting techniques can be applied to other marine industries such as wild fisheries and pond aquaculture in other regions. This future knowledge will enhance environment-related decision making of marine managers and increase industry resilience to climate variability.
The Use of Generalized Additive Models for Forecasting the Abundance of Queets River Coho Salmon
North American Journal of Fisheries Management, 2009
We examined three types of models for preseason forecasting of the abundance of Queets River coho salmon Oncorhynchus kisutch: (1) a simple model in which estimates of smolt production are multiplied by projected marine survival rates, (2) a Ricker spawner-recruitment model, and (3) a regression model relating log-transformed adult recruitment to smolt production. Each type of model was formulated with and without environmental variables that influence production and survival. We attempted to use a nonparametric generalized additive model (GAM) to guide the selection of the environmental variables and the form of the regression model. The GAM model was derived through a stepwise selection strategy based on the Akaike information criterion. Parametric approximate models were developed for each selected GAM model, and their performance was compared with postseason estimates of abundance using three criteria: the mean absolute percentage error, the largest absolute percentage error, and the probability of being included in the 90% prediction interval. This paper shows that the GAM approach is useful in constructing forecasting models by identifying promising relationships with predictor variables and improving abundance forecasts through the incorporation of environmental variables.
Statistical models of Pacific salmon that include environmental variables
Past attempts to improve population models of Pacific salmon Oncorhynchus spp. by adding indices of freshwater or marine conditions have shown mixed success. To increase chances that such models will remain reliable over the long term, we suggest adding only environmental covariates that have a spatial scale of positive correlation among monitoring locations similar to, or greater than, that of the salmon variables that scientists are trying to explain. To illustrate this approach, we analyzed spawner and recruit data for 120 populations (stocks) of pink O. gorbuscha, chum O. keta, and sockeye O. nerka salmon from Washington, British Columbia, and Alaska. Salmon productivity of a given species was positively correlated across stocks at a spatial scale of about 500-800 km. Compared to upwelling and sea-surface salinity, summer sea-surface temperature (SST) showed a more appropriate spatial scale of positive covariation for explaining variation in salmon productivity, and was a significant explanatory variable when added to both single-stock and multi-stock spawner-recruit models. This result suggests that future models of these salmon populations should possibly include stock-specific, summer SST. To further explore our understanding of salmon population dynamics, we developed 24 alternative stock-recruitment models. We compared these models in three ways: (1) their fit to all past data, (2) their ability to forecast recruitment, and (3) their performance inside an "operating model," which included components for dynamics of the natural ecological system, stock assessments based on simulated sampling of data, regulationsetting based on those assessments, and variation in implementing those regulations (reflecting noncompliance or other sources of outcome uncertainty). We also compared single-stock models with multi-stock models (meta-analyses). The latter led to more precise estimates of the effects of SST on log e (recruits/spawner) and greater accuracy of preseason forecasts for some stocks. Analyses with the operating model show that reducing outcome uncertainty should be a top management priority.