Field validation shows bias-corrected pseudo-absence selection is the best method for predictive species-distribution modelling

Abstract

Aim: To determine the performance of different pseudo-absence selection strategies on the prediction of species-distribution models after 30 years of regional climatic and land use changes. Location: Continental France and the Iberian Peninsula. Methods: In this study, we used a large database of Coprophagous Scarabaeidae beetle records collected between 1970 and 1980 in continental France and the Iberian Peninsula to assess the relative performance of different modelling methods in predicting species distributions using current climate and land use information. We used maxent with standard settings and boosted regression trees with three different approaches to generate pseudo-absences. We used historical data to model species distribution and then projected the models into the present. Each method's performance was then assessed by specific field sampling conducted at 20 different sites. Results: Field validation demonstrated that model predictions were more accurate when pseudo-absence data were selected from a sampling bias grid and that model evaluations based on test datasets can lead to false conclusions if not correctly calibrated. The study also demonstrated that the method in which pseudo-absences are dealt with has a major impact on ecological conclusions. Main conclusion: Correcting for spatial bias in collections datasets is of great importance for predicting future trends in species distributions. Uncorrected models showed a strong bias in their predicted species richness patterns.