Anchovy ( Engraulis ringens) and sardine ( Sardinops sagax) abundance forecast off northern Chile: A multivariate ecosystemic neural network approach

Abstract

An evaluation of the performance of artificial neural networks (ANNs) to forecast monthly anchovy ( Engraulis ringens) and sardine ( Sardinops sagax) catches in northern Chile (18°21′S-24°S) is presented, using environmental variables, anchovy and sardine CPUE, fishing effort and catches between 1963 and 2007. An analysis of previous data was carried out, consisting of a non-linear cross-correlation analysis to estimate the time lags for each input variable. A multi-layer perceptron architecture model was used, calibrated with the Levenberg–Marquardt algorithm, thus obtaining anchovy landings and sardine CPUE forecast models. An ecosystemic approach was conducted for both models, considering local and global environmental variables, the anthropogenic effect, and the interaction between species as inputs. The variance explained by both models was slightly higher than 82% and the standard error of prediction was lower than 45%. The strong correlation between the estimated and observed series on the anchovy and sardine models suggests that ANN models capture the trend of the historical data. Furthermore, the generalization capacity along with the sensitivity analysis allowed the identification of high-weight variables in the model, as well as the partial interpretation of the statistical functional relationships between the input variables and the abundance.