Machine-learning classification of non-native plant status in the Brazilian dry forest, Caatinga

Biological invasions are a major driver of global biodiversity loss, yet predictive frameworks remain underdeveloped, particularly in underexplored ecosystems such as the Caatinga dry forest in Brazil. Here, we evaluated the factors influencing the naturalization and invasion of 159 non-native plant species in the Caatinga using machine learning (ML) classification models. We aimed to identify the key ecological, evolutionary, and environmental predictors most strongly associated with each invasion stage and to assess their global and local influence using Shapley additive explanation (SHAP) values. We compiled 49 predictors, including functional traits, phylogenetic distances, niche-similarity metrics, and distribution-model outputs, to classify species into non-native, naturalized, or invasive categories. Separate models were developed for naturalization and invasion transitions, and their performance was evaluated using multiple metrics. Naturalization models had lower accuracy, with height, seed mass, and rainfall-related variables emerging as the most influential predictors. In contrast, invasion models performed well, with climatic niche dissimilarity (pseudo-F), phylogenetic similarity (MPDw), and predicted distribution area as key predictors. SHAP values revealed nonlinear effects and misclassifications that, in some cases, aligned with known ecological behaviors. Our findings highlight the importance of functional similarity, broad climatic tolerance, and niche position in shaping plant invasions in the Caatinga. Despite limitations in predicting naturalization, integrating ML and SHAP provides a promising framework for the early detection and management of invasive species in dryland ecosystems.