ARTIFICIAL NEURAL NETWORKS AS A TOOL FOR SPECIES IDENTIFICATION OF FISH SCHOOLS

Fish schools of sardine, anchovy, and horse mackerel can be discrimina ted from each other, under given conditions, using a set of parameters extracted from echo-integration data. Trawl sampling and hydroacousti c data were collected in 1992 and 1993 in the Thermaikos Gulf by using a towed dual-beam 120 kHz transducer. The parameters extracted from t he available schools were used to train multi-layered feed-forward art ificial neural networks. Various applied networks easily generated ass ociations between school descriptors and species identity, providing a powerful tool for classification. The expertise of the trained networ k was tested with data from identified schools not used in training. T he use of neural networks cannot replace classical statistical procedu res, but offers an alternative when there are significant overlaps in the school characteristics and the parametric assumptions are not sati sfied. (C) 1996 International Council for the Exploration of the Sea