Abstract

In this study the performances of computational neural networks (CNNs), multiple linear regressions (MLRs) and generalised additive models (GAMs) to predict Pacific sardine (Sardinops sagax) landings and to analyse their relationships with environmental factors in the north area off Chile were studied. For this purpose several local and global environmental variables and indexes (sea surface temperature, sea level and Ekman transport index in the Chilean coast and, sea surface temperature in the area Niño 3 + 4 and Niño 1 + 2, and the south oscillation index) were considered as inputs or independent variables. Additionally, several CNNs were calibrated and validated adding the anchovy (Engraulis ringens) landings in the same area as model inputs. The time lags of the variables considered were selected through analysis of the non-linear cross-correlation functions and an alternative form of sensitivity analysis based on the approach of the missing value problem. The analysis of error measures with validation data set showed that the best results were obtained when local and global variables were used separately and combined with anchovy landings. Globally, the best result was given by a CNN with 18 input variables (model CNN 6(II) which only considered global variables and anchovy landings) and 10 neurons in a hidden layer. For this configuration the explained variance was slightly higher to 86% which supposed a standard error of prediction of 7.66%. These results were significantly better than those obtained with MLRs and GAMs. The strong correlation between predicted and observed sardine landings suggests that CNNs captured the trend of the historical data. Also, the generalisation capacity together the sensitivity analysis allowed us to identify the variables with a high weight in the model and partially to interpret the statistical functional relationships between these environmental variables and sardine landings. © 2009 Elsevier B.V. All rights reserved.