Spatial predictions of Baltic phytobenthic communities: Measuring robustness of generalized additive models based on transect data

The spatial distributions of benthic surface sediments and phytobenthic plant species were modelled at a high spatial resolution using generalized additive models together with field data from diving transects. The efficiency of different modelling options was validated using independent datasets, and model fit versus predictive power was analysed. For rock/boulder, sand and mud/clay increasing complexity of the model resulted in higher Reciever Operating Characteristics (ROC) values for the model fit, but lower ROC values for the independent validation. The same pattern was found for hard substrate algae species, whereas it was not true for the rooted plant species. As high model ROC values were often found to be connected to low predictive power of the models, this implies that internal model validation results should be treated cautiously. In general, the models should be kept simple, as the performance of the explanation model increases with increasing complexity, while the predictive power of the model generally decreases. Only by using external validation datasets, the true predictive capacity of an explanation model can be reliably measured, as internal validation schemes tend to over-estimate model performance. Our results also indicate that the Akaike Information Criterion is a more reliable model selection method than Cross-selection when there are few predictor variables.