Patterning community changes in benthic macroinvertebrates in a polluted stream by using artificial neural networks

Data for community dynamics are complex and difficult to analyze, since communities consist of many species varying in a non-linear fashion in spatial and temporal domains. However, investigation of community changes in disturbed aquatic ecosystems is critical for diagnosing temporal community responses to stressful sources and for establishing sustainable management policies to solve the problems of polluted aquatic systems. Although there have been numerous accounts of conventional multivariate analyses on community patterning through clustering and ordination (e.g., Bunn et al. 1986, Legendre and Legendre 1987, Ludwig and Reynolds 1988, Quinn et al. 1991) or on community-environment relationships (e.g., van Dobben and ter Braak 1998), the conventional statistical methods are limited in the sense that they are mainly applicable to linear data and are less flexible for handling ecological data with missing values, noise, etc. Additionally, the studies have mostly been carried out on static community patterns from single samplings. Artificial neural networks (ANNs) are an alternative tool for solving the problem of complexity residing in community data. They are problem oriented and are adaptively flexible for applications (Lippmann 1987, Zurada 1992, Haykin 1994). The multi-layer perceptron has been extensively used in prediction of communities by revealing complex relationships between communities and environmental factors, such as algal bloom (e.g., Recknagel et al. 1997) and establishment of grasslands (e.g., Tan and Smeins 1996). Additionally, temporal networks have been developed to predict community dynamics in a time-delayed manner. The partially and fully connected recurrent ANNs have been utilized to predict short-term community changes (Chon et al. 2000c, 2001). In this case, however, the models were only used for predicting the occurrence of communities. Grouping of community changes has rarely been conducted using artificial neural networks. Recently, however, patterning of community changes has been focused on ecological water quality assessment. Successful management of aquatic ecosystems requires better understanding of the patterns of community development, i.e., either progression of pollution or recovery from the stressful agents. In conventional methods, however, not many studies have focused on grouping of community changes in the temporal domain per se. Mostly communities were clustered in static terms. Since Legendre et al. (1985) and Legendre (1987) discussed chronological clustering in multivariate datasets to represent the succession of species within a community by using ordination and segmentation techniques, specific results concentrating on either the methods for clustering community changes or groupings from field data have not been reported. Similarly, in ANNs, not many studies have been carried out on grouping of the temporal development of communities. By implementing the Self-Organizing Map (SOM) (Kohonen 1989), groupings on communities have been conducted for clustering and ordination on static community patterns (e.g., Chon et al. 1996, Foody 1999, Kwak et al. 2000). Grouping of community changes is not an easy task since patterning of temporal developments has the problem of the unlimited increase in the number of variables as the sampling periods are increased. Chon et al. (2000a) recently grouped community changes by the combined use of unsupervised ANNs, the Adaptive Resonance Theory (ART) plus the SOM. In this case, however, only noise was filtered through the ART, and dimension reduction from the original datasets was not carried out. Additionally, community changes were grouped in a relatively short period of less than six months. In this study, we further address the feasibility of the SOMs in dimension reduction of sequential datasets of community changes and grouping of community changes over a longer period. The proposed model could be used for detecting community changes commonly- occurring in the survey area.