Can bottom-up procedures improve the performance of stream classifications?

Abstract

Top-down methods for defining stream classifications are based on a conceptual model or expert-defined rules, whereas bottom-up methods use biological training data and statistical modelling. We compared the performance of six classification methods for explaining the taxonomic composition of invertebrate and fish assemblages recorded at 327 and 511 sites, respectively, distributed throughout France. Classification 1 and 2 were top-down classifications; The European Water Framework System A (WFDa,) and the French Hydro-ecoregions (HER 2). Four bottom-up classification procedures of increasing complexity were defined based on 11 variables that included watershed characteristics describing climate, topography, and geology, and site characteristics including elevation, bed slope and temperature. Classification 3 was defined using matrix correlation (MC) to select a combination of variable categories that produced the best discrimination of the observed taxonomic composition. Classification 4 and 5 were defined by clustering the sites based on their taxonomic data and then using linear discriminant analysis (LDA) and Random forests (RF) to discriminate the clusters based on the environmental variables. Classification 6 was defined using generalized dissimilarity modelling (GDM). Our hypothesis was that the bottom-up classifications would perform better because they flexibly accommodate complex relationships between compositional and environmental variation. We tested the classifications using the classification strength statistic (CS). The RF-based classification fitted the taxonomic patterns better than GDM or LDA and these latter classifications generally fitted better than the MC, WFDa or HER classifications. Cross validation analysis showed that differences in predictive CS (i.e. the CS statistics produced from sites not used in defining the classifications) were often significant. However, these differences were generally small. Gains in predictive performance of classifications appear to be small relative to the increase in complexity in the manner in which environmental variables are combined to define classes.