The effects of true and pseudo-absence data on the performance of species distribution models at landscape scale

The effects of true and pseudo-absence data on the performance of species distribution models at landscape scale

Potential distribution ranges of natural grassland in subtropical humid mountainous areas were predicted by species distribution models (SDMs) to examine the effects of true and pseudo-absence data on model performances that were scarcely assessed by using real data. Climate spaces of potential ranges of natural grassland were then constructed by principal components analysis (PCA). The distribution map projected by six model algorithms based on true absence data had all presented restricted distribution ranges of natural grassland along mountain ridges, whereas that based on pseudo-absence data presented wider distribution ranges. RF model was used to detect the effects of data record number and contribution of climate variables on model performance because of higher True Skill Statistics. Restricted distribution ranges of natural grassland projected by RF based on true absence data were similar to limited climate space quantified by PCA. However, climate variables related to occurrences of natural grassland were not consistent between RF and PCA results. Occurrences of natural grassland associated with treeline at low elevation were presumably determined by multiple climate factors at subtropical mountain ridges, such as relatively lower temperatures, heavy precipitations, and strong winds. Local climate dataset derived from meteorological stations and followed by altitudinal adjustment was available for modeling species distribution range in mountainous areas. Conclusively, true absence data had practically delineated geographical boundaries and characterized the climate environments of natural grassland. True absence data was recommended to collect along a known environmental gradient and used to construct training dataset with pseudo-absence data to improve model performance.