Predicting fish assemblages in France and evaluating the influence of their environmental variables

Fish are valid species for biological monitoring programs, although they have been less widely used than other organisms like diatoms or macroinvertebrates. Fish can be used as indicator organisms for numerous reasons (Karr 1981; Oberdorff et al. 2001): i) they are present in many water bodies, and fish species are relatively easily identifiable; ii) their lifehistories are well-known as are their ecological requirements; iii) they represent a variety of trophic levels in various habitat types and iv) their economic aspect plays an important role in their use in biomonitoring programs. This study was investigated as part of the PAEQANN project (Predicting Aquatic Ecosystem Quality using Artificial Neural Networks, EU project n° EVK1-CT1999-00026, http://aquaeco.ups-tlse.fr/) under the directive of the European Community (European Parliament 2000, directive 2000/60/EC), studying the impact of environmental variables on the structure and the diversity of aquatic communities. Fish assemblages of reference sites were studied over the whole territory of mainland France. As suggested in several studies (Verneaux 1977; Mahon 1984; Oberdorff et al. 1993; Oberdorff et al. 2002a) fish assemblage structures change along an upstream-downstream gradient as proposed by the River Continuum Concept (Vannote 1980). Flow regime, temperature, food availability and substrate conditions vary from upstream to downstream areas. These variations lead to non-linear relationships between the fish assemblage structure and the environmental variables which characterize the river. Due to their efficiency, artificial neural networks (ANN) with the error backpropagation algorithm are appropriate methods to model non-linear data (Rumelhart 1986). Often compared to multiple linear regression, ANN, which can be used without transformation of the variables, shows higher predictive power (Scardi 1996; Paruelo and Tomasel 1997; Guegan et al. 1998; Kemper and Sommer 2002). Moreover, ANN, which was criticized earlier in its development due to its black-box model type and thus lack of explanatory capacity, has been improved by the introduction of sensitivity analysis methods which are increasingly used to define the most influent variables in ANN models (Lek et al. 1996b; Scardi and Harding 1999; Gevrey et al. 2003). In this paper we i) examined the capacity of ANN models to predict French fish species richness, trophic guild richness and the occurrence of five relevant species using 8 environmental variables; ii) identified the importance of the predictive environmental variables on the output variables using the sensitivity analysis; and iii) discussed the potential of ANN methods in fish community prediction.