Macroinvertebrate assemblages in glacial stream systems: A comparison of linear multivariate methods with artificial neural networks

Abstract

The distribution of 19 macroinvertebrate taxa was related to 36 environmental variables in 3 Alpine glacial streams. Principal component analysis (PCA) and a self-organising map (SOM) were used to ordinate sample sites according to community composition. Multiple linear regression (MLR) was carried out with the environmental variables as predictors and each macroinvertebrate taxon as criterion variable, a multilayer perceptron (MLP) used the environmental variables as input neurons and each taxon as output neuron. The contribution of each environmental variable to macroinvertebrate response was quantified examining MLR regression coefficients and compared with partial derivative (Pad) and connection weights approach (CW) methods. PCA and SOM emphasized a difference between glacial (kryal) and non-glacial (krenal and rhithral) stations. Canonical correlation analysis (CANCOR) confirmed this separation, outlining the environmental variables (altitude, distance from source and water temperature) which contributed most with macroinvertebrates to site ordination. SOM clustered kryal, rhithral and krenal in three well separated group of sites. MLR and MLP detected the best predictors of macroinvertebrate response. Pad sensitivity analysis and CW method emphasized the importance of water chemistry and substrate in determining the response of taxa, the importance of substrate was overlooked by linear multivariate analysis (MLR).