Identification of suitable unoccupied habitats: direct versus an indirect approach.

Identification of habitats suitable for a species is a key task when studying landscape dynamics. Direct (cultivation experiments) or indirect (species distribution models) methods are employed to identify both suitable but currently unoccupied habitats and habitats that are unsuitable yet occupied (remnant populations). Although both direct and indirect approaches have been used increasingly in recent years the predictive power of cultivation experiments and of models based on different types of data have not been compared. In the present study, we compare the predictive power of distribution models for a short-lived herbaceous plant, Jasione montana. Our models are based on the environmental characteristics of the habitats, on the species composition of the habitats, or on both of these types of data. The predictions of the different models were compared (using AUC values) with the results of our cultivation experiment. We found that the models based on the species composition of the habitats performed better than the model based only on environmental characteristics. The models also differed significantly in the unoccupied habitats they identified as suitable. The most accurate was the model based on both environmental characteristics and species composition. This model also significantly explained both the presence/absence and abundance of J. montana individuals in the cultivation experiment. Nevertheless, the variation in occurrence of J. montana in the cultivation experiment explained by this model was below 50%. We therefore assume that the predictions of this model, in spite of the high AUC values, were inaccurate for at least some habitats. The results of this study are species and landscape specific, so they cannot be generalized. Our study, however, demonstrates that assembling data on both environmental characteristics and species composition of habitats is likely to be useful for predicting habitat suitability at a landscape scale. This study also demonstrates that a high AUC value is not a guarantee that a model’s prediction is reliable because a cultivation experiment may provide different results. When identifying habitats that are suitable for a species (e.g. for the purpose of a metapopulation study), the results should be subjected to a sensitivity analysis.