A comparison of modelled and actual distributions of eleven benthic macroinvertebrate species in a Central European mountain catchment

Abstract

Species distribution Modelling (SDM) constitutes a useful tool to predict the distribution of freshwater species based on selected habitat variables. Model performance (goodness-of-fit) expressed as coefficient of determination is straightforward to judge on SDM quality but does not sufficiently address predictive success. Hence, predictive performance (e.g. AUC) accounts for the correctness of predicted species occurrences. In this study, we compared the model and predictive performance of SDMs on eleven macroinvertebrate taxa in a mountain catchment, with emphasis on species prevalence. SDMs were based on two regression methods using broad-scale environmental predictors (land use, instream habitat quality). We applied a cross-validation and a field validation approach using newly sampled field data. Opposed to other species, SDMs showed acceptable performance (pseudo-R2 > 0.3) for the stonefly Dinocras cephalotes and the caddisflies Silo piceus and Silo pallipes. Model performance was neither positively nor linearly correlated with predictive accuracy. The comparison of cross- and field validation revealed an overestimation of the discriminatory power of cross-validated models. SDMs of less prevalent species tend to over-predict absences rather than presences. Consequently, model performance is decoupled from predictive performance. The validation results suggest the use of new field data providing a more reliable benchmark for SDM assessment.