Case study: Comparing the use of nonlinear discriminating analysis and Artificial Neural Networks in the classification of three fish species: acaras (Geophagus brasiliensis), tilapias (Tilapia rendalli) and mullets (Mugil liza)

In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier's archiving and manuscript policies are encouraged to visit: Case study: Comparing the use of nonlinear discriminating analysis and Artificial Neural Networks in the classification of three fish species: acaras (Geophagus brasiliensis), tilapias (Tilapia rendalli) and mullets (Mugil liza) a b s t r a c t a r t i c l e i n f o This study used the Discriminant Analysis statistical technique and Artificial Neural Networks, multilayer perceptron, in the classification of three fish species sampled in the state of Rio de Janeiro, Brazil: Geophagus brasiliensis (acaras), Tilapia rendall (tilapias) and Mugil liza (mullets). These fish were sexed when possible, weighed, measured, and had their Gonadosomatic and Hepatosomatic Indices calculated, as well as their Condition Factor. The use of an Artificial Neural Network (ANN) presented satisfactory results, even though the groups were composed of very diverse-sized animals. Without the need for non-violation assumptions and other considerations, the Artificial Neural Network was found to be the excellent alternative to classification problems of unbalanced data, such as the one presented in this study. The Discriminant Analysis is a multivariate statistical technique, destined to treat classification problems when the aim is to establish relations between non-metric and metric variables. The discrimination is obtained by pondering the variables, in order to maximize variance between groups and minimize variance within groups. This pondering is done by means of a function. According to Hair et al. (1998), the successful application of this technique requires the consideration of several aspects, including the selection of the variable response, the size of the sample for the estimate of the discriminant functions and sample division for validation purposes. In the classification stage, larger groups tend to have an unproportionally higher classification probability. One alternative is the extraction of a random sample from larger groups followed by a size reduction to a level comparable to the smaller groups. Several procedures have been suggested for sampling division in order to estimate the discriminant functions and for validation purposes. The most common is the crossed validation, which avoids the " overad-justment " of the discriminant function, since the validation is conducted on a totally separate sample (Hair et al., 1998). Statistical tests such as the Wilks λ statistic, …