Spatial ensemble learning for predicting the potential geographical distribution of invasive species

Abstract

Understanding the geographical distribution of invasive species is beneficial for preventing and controlling biological invasions. A global model is often constructed with existing species distribution models (SDMs) to describe the relationships between environmental characteristics and species distributions. Because of the spatial variations in environmental characteristics, it may be difficult for a single SDM to obtain an accurate result in any given location or area. Therefore, a spatial ensemble learning method for predicting the potential geographical distribution of invasive species is presented in this study. The method mainly includes two types of learners: one learner is a base learner used to predict the geographical distribution of invasive species, and the other learner is a spatial ensemble learner for combining predictions from different base learners. In this research, spatial ensemble learning is used to predict the geographical distribution of Erigeron annuus in the Yangtze River Economic Belt, China. The kappa coefficient and AUC (area under the receiver operating characteristic curve) obtained with the spatial ensemble learner are 0.88 and 0.94, respectively, and these values are greater than those obtained using three base learners and other ensemble strategies. This demonstrates the feasibility and effectiveness of spatial ensemble learning.