Classification of stream diatom communities using a self-organizing map

One of the major axes of the new E.U. Water Framework Directive is to assess the deviation of an ecosystem with respect to the highest ecological quality awaited (non-perturbed or reference conditions), thanks to the responses of aquatic communities. By comparing diatom communities in natural and disturbed sites, indicators for different types of anthropogenic disturbance can be found. But, since diatom species composition varies among streams due to natural as well as anthropogenic factors, we should be able to increase the accuracy of assessing anthropogenic impact by first accounting for natural variability among sites. Kociolek (2000) argue that the proportion of geographically restricted diatom species is high and spatial distribution patterns of species are still poorly understood. Although many studies have focused on the effect of human pressure on diatom communities (Pan et al. 1999, Licursi and Gomez 2002, Winter and Duthie 2000, Potapova and Charles 2002, Soininen 2002), the number of studies attempting to characterise the natural patterns and the relative weights of environmental parameters influencing this natural variability is limited (Descy 1984, Leclercq and Depiereux 1987, Sabater and Roca 1992, Stevenson 1997, Pan et al. 2000). Ordination techniques are a useful way to explore the characteristics of datasets and to find relationships between variables. Diverse linear ordination methods have been used to simplify the data including polar ordination, principal components analysis (PCA), correspondence analysis (CA) (Pearson 1901, Hill and Gauch 1980, Beals 1984, Jongman et al. 1995). The limitations are well-known, e.g. all of them present strong distortions with nonlinear species abundance relations (Kenkel and Orloci 1986); horseshoe effects due to unimodal species response curves in PCA and arch effects, outliers, missing data, disjointed data matrix in CA (Giraudel and Lek 2001). Recently, as an alternative tool to deal with this problem of complexity in ecological data, artificial neural networks (ANNs) have been utilized for patterning communities in various ecosystems (i.e., aquatic, forest, agriculture, etc.) (Lek and Guégan 2000, Recknagel 2003). Among the ANN techniques, Kohonen's self-organizing map (SOM) (Kohonen 1982, 2001) is the most popular unsupervised learning algorithm, allowing the classification of data without prior knowledge and the visualisation of species assemblages in a two-dimensional space (Giraudel and Lek 2001). The SOM has been used for the classification of communities (Chon et al. 1996, Foody 1999, Park et al. 2001a, 2003a), for water quality assessment (Walley et al. 2000, Aguilera et al. 2001), and for prediction of populations and communities (Céréghino et al. 2001, Obach et al. 2001). The ability of the SOM for classification and ordination in ecology has also been compared with conventional multivariate analysis (Chon et al. 1996, Foody 1999). In particular, Giraudel and Lek (2001) compared the SOM with several different multivariate analysis methods including PCA and CA, and concluded that the SOM seems fully usable in ecology and can be a perfect complement to classic techniques for exploring data and for achieving community ordination. Nijboer et al. (see # 4.5) also compared the SOM with a canonical correspondence analysis and cluster analysis for the ordination and classification of macroinvertebrate communities, showing that each method has its own strengths and weakness depending on the objectives of the study. Our study, run on a pilot dataset (Adour-Garonne stream system, South-West of France), was the first attempt to highlight the natural spatial distribution scheme of benthic diatoms on a regional scale, and explored the performance of the SOM in diatom community studies. The main purpose is to give a practical application of unsupervised neural networks for patterning diatom community structure in reference situations, sustaining the WFD implementation. In this study, we developed an innovative methodological approach to establish a first diatom-based bio-typology of an Adour-Garonne basin stream system which makes clear headway in ecoregional zoning. The research was carried out in the framework of the European Research Program PAEQANN (Predicting Aquatic Ecosystem Quality using Artificial Neural Networks - 5^th PCRD), aiming to develop general methods, based on advanced modelling techniques, for predicting the structure and diversity of key aquatic communities under natural conditions and subjected to man-made disturbances.