Patterning riverine fish assemblages using an unsupervised neural network

In temperate river systems, the uni-directional character of water tends to give them a linear structure along a gradient of environmental conditions. In these systems, biological assemblages are organised longitudinally and there is generally an increase in species richness from the source to the river mouth (e.g. measured by the river width, the distance from the source, the stream order, and the size of the watershed). Longitudinal changes in local assemblage richness have usually been attributed to one of two processes: biotic zonation or continual addition of species downstream. Biotic zonation corresponds to discontinuities in river geomorphology or abiotic conditions promoting distinct assemblages along the longitudinal gradient (e.g. Huet 1959, Schlosser 1982, Balon et al. 1986, Rahel and Hubert 1991, Oberdorff et al. 1993, Belliard et al. 1997). For example, species replacement may occur as a result of physiological specialization for temperature. In contrast to the advocates of zonation, additions of species are usually related to environmental gradients having smooth transitions of abiotic factors contributing to nested patterns of assemblage composition along the longitudinal gradient (e.g. Sheldon 1968, Rahel and Hubert 1991). Whatever the process (i.e. biotic zonation or species addition) local species richness usually increases along the upstream-downstream gradient (Huet 1959, Sheldon 1968, Schlosser 1982, Balon et al. 1986, Rahel and Hubert 1991, Belliard et al. 1997, Oberdorff et al. 2001, 2002a). The environmental factors that have been identified to explain this increase in species richness are generally linked (i) to upstream-downstream differences in local habitat characteristics defined by depth, slope, current velocity, temperature and substrate composition (Huet 1959, Gorman and Karr 1978, Schlosser 1982, Angermeier and Schlosser 1989, Rahel and Hubert 1991, Oberdorff et al. 2001, 2002a) or by "dimensionless" hydraulic characteristics such as the Froude number or the Reynolds number (Lamouroux and Souchon 2002, Lamouroux and Capra 2002) and (ii) to an upstream-downstream increase in environmental stability (Horwitz 1978, Schlosser 1982, Schlosser and Ebel 1989). Resulting from this common feature in riverine fish ecology (i.e., the longitudinal change in fish assemblage structure along the upstream-downstream gradient of a river), some authors have attempted to classify river basins into different biotic zones. The classical studies include the work of Thienemann (1925) who proposed six zones for continental European rivers: spring brook, trout zone, grayling zone, barbel zone, bream zone and brackish-water, each based on the presence of a specific fish species. This elegant concept persisted in the systems devised by Huet (1949, 1954), who proposed longitudinal zonations of rivers based on the occurrence of key species. The Huet zonation consists of four zones, beginning with the headwater and moving to the lowlands (i.e., the trout zone, the grayling zone, the barbel zone, and the bream zone). To visualize the organization and structure of fish assemblages, several multivariate techniques have been used depending on the aim of the studies, including multivariate analysis of variance (Bendell and McNicol 1987, Jackson and Harvey 1989), factor analysis (Stevenson et al. 1974, Oberdorff et al. 1993), correspondence analysis (Hughes and Gammon 1987, Strayer 1993, Pusey et al. 1995, Vila-Gispert 2002), cluster analysis (Hughes et al. 1987, Poff and Ward 1989, Johnson and Wichern 1992), principal component analysis (Matthews 1985, Matthews and Robinson 1988, Paller et al. 1994, Vila-Gispert 2002) and canonical correspondence analysis (Taylor et al. 1993, Copp 1992, Koel 1997). These methods are all adversely affected by the non-linear nature of the ecological data, whereas the methods identified subsequently (i.e., adaptive learning algorithms) are not. As alternative methods, adaptive learning algorithms such as artificial neural networks (ANNs) are becoming more and more popular in ecological studies (Lek and Guégan 2000, Rekgnagel 2003). Among the algorithms of the ANNs, the self-organizing map (SOM) shows an ability for classification, abstraction, and visualization, the idea of which is to show the data set in another, more usable, representation (Kohonen 2001), and to efficiently determine patterns of aquatic ecological assemblages (Chon et al. 1996, Brosse et al. 2001, Park et al. 2003a). In this study, we propose a SOM model as an alternative method to display patterns of fish species distribution in French rivers, and to evaluate the relative importance of several environmental factors in influencing organization and structure of fish assemblages.