Species-specific traits associated to prediction errors in bird habitat suitability modelling

Abstract

Although there is a wide range of empirical models applied to predict the distribution and abundance of organisms, we lack an understanding of which ecological characteristics of the species being predicted affect the accuracy of those models. However, if we knew the effect of specific traits on modelling results, we could both improve the sampling design for particular species and properly judge model performance. In this study, we first model spatial variation in winter bird density in a large region (Central Spain) applying regression trees to 64 species. Then we associate model accuracy to characteristics of species describing their habitat selection, environmental specialization, maximum densities in the study region, gregariousness, detectability and body size.

Predictive power of models covaried with model characteristics (i.e., sample size) and autoecological traits of species, with 48% of interspecific variability being explained by two partial least regression components. There are species-specific characteristics constraining abundance forecasting that are rooted in the natural history of organisms. Controlling for the positive effect of prevalence, the better predicted species had high environmental specialization and reached higher maximum densities. We also detected a measurable positive effect of species detectability. Thus, generalist species and those locally scarce and inconspicuous are unlikely to be modelled with great accuracy. Our results suggest that the limitations caused by those species-specific traits associated with survey work (e.g., conspicuousness, gregariousness or maximum ecological densities) will be difficult to circumvent by either statistical approaches or increasing sampling effort while recording biodiversity in extensive programs.