A trial of artificial neural networks for automatically estimating the age of fish

Abstract

Artificial neural networks (ANNs) have the potential to automate routine ageing of fish with the benefit of increased speed in processing, greater objectivity and repeatability of estimates, and a mechanism for quantifying uncertainty of age estimates. ANN models were tested as a means of objectively replicating the age estimates of an experienced human reader. Feed-forward back- propagation ANNs, with three layers of neurons (input, hidden and output), were trained to classify the age of previously aged samples of three temperate species. Three ANN structures, where the number of neurons in the hidden layer was varied, were tested for each species. Inputs to each ANN were pixel brightness values along transects across images of sectioned otoliths. The ANN predicted age-class membership by the position of the neuron in the output layer with the highest value. After training, at least one of the three ANN structures correctly classified the age of fish from unseen transects for two members of the Sparidae family Acanthopagrus butcheri and Pagrus auratus at an accuracy level approaching that of an expert reader. For a member of the Merlucciidae family, Macruronus novaezelandiae, however, which is a species with more complex otolith structure, error rates were high for all three ANN structures tested.