On the usefulness of prediction intervals for local species distribution model forecasts

Abstract The maps produced by species distribution models (SDMs) are increasingly used by decision-makers for supporting local and regional land-use as well as landscape planning issues. While ecologists generally are interested in large-scale patterns and the overall quality of SDMs, decision-makers and conservationists focus on the reliability of localized predictions relevant for specific projects. Here, we use the machine learning methods Random Forest and Quantile Regression Forest to predict local abundance of the black-tailed godwit Limosa limosa with prediction intervals, a measure of the probability that a future observation will lie between certain limits. Although the confidence intervals for local predictions are very narrow, the corresponding prediction intervals are very wide. Therefore, the actual numbers of the black-tailed godwit expected at a given point in the field may vary from virtually absent to high density. We conclude that practitioners should lower their expectations of maps based on the currently available SDMs and to be careful when utilizing them for supporting local management decisions.