Predicting Dutch macroinvertebrate species richness and functional feeding groups using five modelling techniques

When establishing a quantitative model to predict macroinvertebrate communities from environmental variables, the variety and complexity of variables often make the process of selecting a modelling method difficult. The most popular prediction method in ecology is multiple linear regression (Holler et al. 1993, Chhetri and Fowler 1996, Green 1996, Quensen and Woodruff 1997, Kolozsvary and Swihart 1999). In spite of numerous qualities, this method has two major weaknesses. Firstly, the parametric approaches involved assume distributions of relationships between data that may or not may hold. Secondly, the assumption of the linearity of the data is often questionable. In recent years, considerable attention has been given to the development of techniques for exploring data sets. New computational methods either overcome the parametric assumption, such as Partial Least Squares (PLS, Wold et al. 1983), or identifying non-linear relationships between the data, such as General Additive Models (GAM, Hastie and Tibshirani 1986, 1990), Regression Trees (RT, Breiman et al. 1984) and Artificial Neural Networks (ANNs, Rumelhart and Mc Clelland 1986). The predictive capacity of Multiple Linear Regression (MLR) was compared with (i) ANNs (Lek et al. 1996b, Brey et al. 1996, Paruelo and Tomasel 1997, Brasquet et al. 1999, Kemper and Sommer 2002), (ii) RT (Rejwan et al. 1999, Boone and Krohn 2000), (iii) PLS (Sanz et al. 1999, Schmilovitch et al. 2000, Dane et al. 2001, Delalieux et al. 2002), and (iv) GAM (Ette and Ludden 1996, Brosse and Lek 2000a). Several studies have compared different modelling techniques; to predict vegetation types (Cairns 2001), to fit the biological structural activity relationship in microbiology (Ramos-Nino et al. 1997), to model fish microhabitats (Brosse and Lek 2000b), to model fish species distributions (Olden and Jackson 2002), to predict the abundance of aquatic insects (Wagner et al. 2000a,b), to develop quantitative inference models in paleolimnology (Racca et al. 2001), to predict forest characteristics (Moisen and Frescino 2002), and to capture ozone behaviour (Gardner and Dorling 2000). Species richness (SR) is an integrative descriptor of the community, as it is influenced by changes of natural environmental variables as well as anthropogenic disturbances (Rosenberg and Resh 1993). Therefore, it is commonly used as an ecological indicator for ecosystem assessments. The functional feeding groups (FFGs) of benthic macroinvertebrates are guilds of invertebrate taxa that obtain food in similar ways, regardless of taxonomic affinities. Therefore, they can represent a taxonomicaly heterogeneous assemblage of benthic fauna as well as a variety of disturbances of their habitats. Moreover, they reflect the food resources available in a given area, therefore their distributions respond mostly to disturbances that alter the food base of the system (e.g., Hershey et al. 1988, Hart and Robinson 1990). The proportion of different groups may change in response to disturbances that affect the food base of the system, thereby offering a means of assessing disruption of ecosystem function. Therefore, the percentages of FFGs have commonly contributed as indicators of rapid bioassessment (Resh and Jackson 1993, Barbour et al. 1999). Predicting SR and FFGs is valuable for aquatic ecosystem management. The primary objective of this work was to compare several recently developed techniques to predict a simple community index, SR. The second objective was to determine the relative strength of these methods while increasing the complexity of the variables i.e. in the prediction of FFGs. The last objective was to compare their ability to select those environmental variables that best predict the macroinvertebrate SR and FFGs.