Patterning, predicting stream macroinvertebrate assemblages in Victoria (Australia) using artificial neural networks and genetic algorithms

Abstract

Macroinvertebrate assemblages are widely used for biomonitoring of stream ecosystems. Several modern assessment concepts and approaches have been desribed. The so-called referential approach (Parsons and Norris 1996, Marchant et al. 1999, Smith et al. 1999) is based on the comparison of macroinvertebrate communities between potentially impacted sites and reference sites considered to be pristine. Knowing the relationships between environmental variables and macroinvertebrate occurrence at reference sites, it is possible to predict species or taxa, which should occur at the remaining sites in the absence of anthropogenic stress. The ratio of observed/expected (O/E) families is used as a measure for sitespecific ecological conditions. Statistical and computational techniques have been successfully integrated into the referential approach facilitating stream site classification and prediction of macroinvertebrate assemblages. Classification or grouping of macroinvertebrates into assemblages is sometimes criticized as an arbitrary procedure as they are usually distributed in continuous gradients rather than well defined separate groups (Chessman 1999). However in order to deal with large numbers of macroinvertebrate taxa it is often crucial to consider groups instead of individual taxa provided appropriate classification techniques are available. Widely used statistical methods for data classification and ordination are cluster and principal component analysis. Both methods have shortcomings in coping with heterogeneous and nonlinear data, and results can be confounded by outliers and missing data. Artificial neural network (ANN) based classification techniques such as Kohonen or SelfOrganizing Maps (SOM) may help to overcome these shortcomings. A number of ecological case studies have shown that SOM are an efficient classification tool (Chon et al. 1996, 2003, Cereghino et al. 2001, Park et al. 2001a, 2003a, Brosse et al. 2001, Giraudel and Lek 2001). ANN as well as genetic algorithms (GA) prove to be appropriate for the prediction of macroinvertebrate and fish assemblages in streams. Multi-layer perceptron ANN were successfully applied to predict the occurrence of stream macroinvertebrates from environmental variables (Walley and Fontama 1998, Schleiter et al. 1999, Pudmenzky et al. 1998, Hoang et al. 2001). GA were used to predict fish distribution from physical characteristics of streams (d'Angelo et al. 1995) and to select input variables of classification tree models predicting benthic macroinvertebrate communities in Belgian watercourses (Goethals et al. 2003).