Application of Artificial Neural Networks for Forecasting Salmon Escapement

Abstract

Forecasts of the abundance and ocean escapement of Pacific salmon Oncorhynchus spp. are essential for fishery management. Based on data availability and quality, several forecasting methods have been used for different stocks. Data limitations have been an impediment to the use of formal statistical forecasting methods for many stocks. To date, the moving average (MA) of historical escapements has appeared to be the best choice for these stocks. I investigated the application of artificial neural networks (ANNs) to salmon escapement forecasting. I constructed ANNs with three layers (input, hidden, and output) and tested them with data from two stocks of Oregon coastal fall chinook salmon O. tshawytscha (the Siletz and Nehalem rivers) from 1986 to 2000. The outputs were compared with those of forecasts made with the traditional MA method. In addition, I analyzed the time series data with an autoregressive integrated moving average method (ARIMA). The ANNs generally outperformed the MA method for both stocks; networks with one hidden neuron slightly outperformed those with two or three hidden neurons. For the Siletz River stock, the one-hidden-neuron ANNs resulted in a mean absolute percent error (MAPE; i.e., the relative error between observed and predicted escapements) of 24.1%, compared with 31.7% for the MA method. For the Nehalem River stock, ANNs result in an MAPE of 27.7%, compared with 34.8% for the MA method. For 1997–2000, ANNs had a lower average forecasting error than the ARIMA for the Siletz River stock but not for the Nehalem River stock.