Optimization of artificial neural networks for predicting fish assemblages in rivers

Predicting the structure of fish assemblages in rivers is a very important goal in ecological research, both from a purely theoretical point of view and from an applied one. Moreover, it will play a relevant role in the definition of reference conditions in the light of the EU Directive 2000/60/EC (i.e. the Water Framework Directive). Estimates of the probability of presence/absence of fish species have been obtained so far using different approaches. Although conventional statistical tools (e.g. logistic regression) provided interesting results, the application of artificial neural networks (ANNs) has recently outperformed those techniques. ANNs are especially effective in reproducing the complex, non-linear relationships that link environmental variables to fish species presence and/or abundance. In this chapter some new developments in ANN training procedures will be presented, which are specifically aimed at solving ecological problems related to the way the errors are computed in species composition models. The resulting improvements in species prediction involve not only the accuracy of the models, but also their ecological consistency. A case history about fish assemblages in the rivers of the Veneto region (NE Italy) is presented to demonstrate how the enhanced modelling strategy improved the accuracy of the predictions about fish assemblages.