Predicting alien herb invasion with machine learning models: biogeographical and life-history traits both matter

Abstract

Identifying the variables associated with invasiveness is a core task for developing risk assessment models to predict invasion potential. However, quantitative models with both biogeographical and life-history variables for invasion risk assessment in China are limited. We hypothesized that (1) compared to statistical algorithms, some machine learning models could offer a promising quantitative approach with high accuracy for potential invader prediction; (2) native range distribution size, origins and life-history traits co-determine an alien plant’s performance in the latter invasion stage. In this study, we used four machine learning models [classification and regression tree (CART), multivariate adaptive regression spline (MARS), random forest (RF) and multiple additive regression tree (MART)] and two traditional statistical algorithms [logistic regression (LR) and linear discriminant analysis (LDA)] to assess the relative importance of biogeographical and trait variables in the naturalized-invasion stage of 150 invasive and 87 non-invasive herb plants in China. Our results showed that good performance was the case for all predictive models (AUROC ranges from 0.68 to 0.87), which had overall mean performance value ranging from 0.66 to 0.82. Compared with traditional statistical algorithms, MART and RF models have a consistently higher accuracy, indicating that these two models could be used as alternative quantitative approaches for risk assessment. Additionally, both biogeographical (native range distribution size) and life-history traits (seed weight) were screened out by the models, suggesting their high correlation with plant invasiveness and important roles in risk assessment.