Hierarchical patterning of benthic macroinvertebrate communities using unsupervised artificial neural networks

Patterning communities is essential to reveal the ecological states of the target ecosystem effectively and consistently. Especially in aquatic ecosystems, the composition of residential communities rapidly varies in response to various impacts of natural and anthropogenic perturbations such as flooding and pollution (Hawkes 1979, Hellawell 1986, Spellerberg 1991). Particular attention has been recently focussed on properly assessing changes in water quality through community patterning. As well documented, field community data are nonlinear and complex because they involve many species, fluctuating greatly depending upon numerous effects of endogenous (e.g., physiological development, life cycle, etc.) and exogenous factors (e.g., precipitation, pollution, etc.) (Jongman et al. 1995, Legendre and Legendre 1998). A complex system like the responses of communities to their environments usually develops a hierarchical structure (Allen and Starr 1982, O'Neill et al. 1986); in particular, benthic macroinvertebrates in streams clearly develop taxonomic and functional hierarchies that are essential to establish organization in communities (Cummins et al. 1973, Cummins 1974). Additionally, habitats of benthic macroinvertebrates in streams are also classified hierarchically, taking into account the fact that variables are revealed differently across different space and time scales on which a system is viewed (Frissell et al. 1986, Minshall 1988). Since a hierarchical nature is an essential part of stream ecosystems, the determination of the appropriate methods of examination has been a key concept in investigating aquatic ecosystems (Minshall 1993). Consequently, the hierarchical classification approaches could provide in-depth and comprehensive understanding of community organization and water quality in the target ecosystem. Assessment of water quality and prediction of community dynamics in streams are essential for diagnosing ecosystem health and for providing policies of sustainable management of stream ecosystems. Especially benthic macroinvertebrate communities are effective in indicating water quality and could effectively reveal ecological states of the target aquatic ecosystem. They constitute a heterogeneous assemblage of animal phyla, and consequently it is probable that some members will always respond to stresses placed upon them (Hynes 1960, Hawkes 1979, Hellawell 1986). Communities have been analyzed by conventional multivariate statistical methods (Ludwig and Reynolds 1988, Jongman et al. 1995, Legendre and Legendre 1998), however they are limited in extracting information effectively out of complex data. As an alternative tool to deal with this problem of complexity in ecological data, artificial neural networks (ANNs) have been utilized for patterning communities. The ANNs are well known for their ability to extract information from nonlinear and complex systems, and have been well applied to the study of secosystems (Lek and Guégan 2000, Recknagel 2003). Among the ANN techniques, Kohonen's Self- Organizing Map (SOM) (Kohonen 1982, 1989, 2001) is the most popular unsupervised learning algorithm. In aquatic ecosystems, the SOM has been used for classifying communities (Chon et al. 1996, Foody 1999, Park et al. 2001a, 2003a), for water quality assessments (Walley et al. 2000, Aguilera et al. 2001), and for population and community predictions (Céréghino et al. 2001, Obach et al. 2001). The classification by the SOM, however, has the problem of objectivity in finding similarities among the map units (Chon et al. 1996). When the groups are located far apart on the map, it is difficult to judge to what extent they are similar. Furthermore, due to randomness in iterative calculations and variability in determining parameters in the learning process of the network, the grouping presents a slightly different conformation after each training task. Thus, to effectively define clusters among units of the SOM map, differentiation in the degree of clustering per se is additionally required. To divide the map into certain subareas, the unified-matrix algorithm (Ultsch and Siemon 1990, Ultsch 1993) is currently the most often used. However, it is not an easy task to efficiently reveal different degrees of clustering based on this distance matrix. In this study, we propose a combinational method for successively clustering communities through self-organization. The model developed was further evaluated with new data sets to detect the effect of sub-groupings on community development.