The stakes are far from academic. Around 14,000 plant species have already naturalized worldwide, roughly four percent of all vascular plants, and approximately six percent of those have become invasive. Invasive plants displace native species, contribute to extinctions, disrupt energy and nutrient cycles, and generate billions of dollars in annual costs to agriculture, health, and ecosystem services. Predicting which species will cross the critical threshold from naturalization to invasion has therefore become a central goal of invasion biology, and a prerequisite for prevention rather than expensive after-the-fact control. Yet more than 30 competing hypotheses have been proposed to explain why some introduced species thrive while others fade, ranging from propagule pressure to biotic interactions to climatic filtering, and untangling their relative influence has proven notoriously difficult.

The new study, published in Discover Conservation, approached the problem by treating each plant species as a data point and each invasion hypothesis as a measurable variable. The researchers compiled an extraordinary 49 predictors for every species, spanning functional traits such as leaf area, seed mass, and plant height; phylogenetic distances measuring how closely related each newcomer is to the native flora; metrics of climatic similarity between a species’ native range and the Caatinga; minimum residence time since first introduction as a proxy for propagule pressure; and the output of ecological niche models predicting how much suitable habitat each species could occupy. The species were classified into three stages following established criteria: non-native, meaning introduced taxa without reproductive capacity; naturalized, meaning those with self-sustaining populations; and invasive, meaning those capable of spontaneous spread.

The Caatinga itself provides a demanding test bed. Home to approximately 3,347 plant species, fifteen percent of which are found nowhere else on Earth, the biome is characterized by spiny shrubs, deciduous trees, cacti, and bromeliads under an open canopy, all shaped by water scarcity, high temperatures, and pulsed resource availability. These strong environmental filters make the region an ideal natural laboratory for testing whether invaders succeed by bringing novel strategies or by resembling the hardy natives already adapted to drought. Previous work had suggested the latter, and the new machine-learning results lend that idea substantial quantitative support.

Methodologically, the team built separate binary models for the two key transitions: from non-native to naturalized, and from naturalized to invasive. Because invasive species were underrepresented in the data, they applied the synthetic minority over-sampling technique, known as SMOTE, to generate synthetic observations of the minority class in the multivariate variable space. A preliminary random forest ranked the importance of all 49 predictors, after which the researchers tested subsets of the top-ranked variables, tuned hyperparameters through exhaustive and randomized searches with 200 iterations, and validated everything with five-fold cross-validation. Three tree-based algorithms competed in the final round: random forest, CatBoost, and XGBoost. XGBoost delivered the best performance, though its optimal settings led to overfitting, so the team adopted the next-best configuration to keep the models honest.

The results revealed a striking asymmetry. The invasion models performed well, achieving an area under the ROC curve of 0.811, while the naturalization models struggled, reaching only 0.717 and proving too weak for reliable classification. For invasion, just three predictors carried nearly all the predictive power: the pseudo-F statistic, a multivariate measure of climatic niche dissimilarity between a species’ native range and the Caatinga derived from PERMANOVA; the abundance-weighted mean phylogenetic distance to native species, or MPDw; and the predicted distribution area within the biome. In other words, the species most likely to become invasive are those whose native climates differ markedly from or span broadly beyond the Caatinga’s conditions, those that are closely related to the native flora, and those with vast stretches of potentially suitable habitat.

To peer inside the black box of these predictions, the researchers turned to Shapley additive explanations, or SHAP, a framework borrowed from game theory that quantifies exactly how much each variable pushes an individual prediction up or down. The SHAP analysis confirmed the global rankings and exposed their nonlinear, interacting effects. Low pseudo-F values were generally associated with naturalized species, while low MPDw values combined with very large predicted areas were linked to invasiveness. Species with native environments similar to the Caatinga tended to be classified as naturalized, whereas those both ecologically similar to native species and equipped with large potential ranges were flagged as invasive. Notably, no invasive species displayed an introduced niche broader than its native one, a pattern that argues against rapid niche evolution and instead supports climatic preadaptation combined with functional similarity to natives as the dominant mechanism.

The misclassifications proved almost as informative as the correct predictions. Several notorious invaders, including Leucaena leucocephala and Prosopis pallida, were predicted as merely naturalized, a sobering reminder that even well-performing models can underestimate the threat posed by harmful species, particularly when applied to newly introduced plants. Conversely, species such as Psidium guajava and Amaranthus spinosus were flagged as invasive despite their current naturalized status. Rather than dismissing these errors, the authors suggest some may signal emerging invasions or outdated status designations: guava, for instance, disperses prolifically through human activity and frugivorous animals and is already invasive elsewhere, while Amaranthus spinosus is already spreading unintentionally near the Caatinga’s borders. In this sense, the model functions not just as a classifier but as an early-warning system whose disagreements with the record deserve field verification.

The weak performance of the naturalization model carries its own ecological message. Plant height and seed mass, along with rainfall-related bioclimatic variables, emerged as the most influential predictors, but they apparently capture only part of what governs whether an introduced species establishes self-sustaining populations. The authors point to the possibility that key drivers were missing from the dataset altogether, such as the availability of mutualistic partners like pollinators required for successful reproduction, consistent with the so-called missed mutualism hypothesis. Combined with the limited number of introduced species and strong class imbalance, these gaps likely blunted the model’s accuracy. Crucially, however, naturalization is the less dangerous transition; invasion is where the real ecological damage begins, and that is precisely where the framework excelled.

The broader implications reach well beyond one Brazilian biome. Most invasive plants in the Caatinga belong to the Poaceae and Fabaceae, two of the most successful lineages in the ecosystem, and grasses as a family show a strong phylogenetic signal of invasiveness globally. This supports the view that preadaptation, rather than novel traits, drives the transition from naturalization to invasion: species evolutionarily equipped for drought and disturbance simply need an opening. By integrating machine learning with interpretable SHAP values, the study offers a template that conservation managers in drylands worldwide could adapt, screening newly arrived species for climatic tolerance, relatedness to the native flora, and predicted range size before they spread. The researchers have made their raw data and analysis notebooks openly available, inviting replication and refinement. For a biome that remains one of Brazil’s most neglected, the message is clear: the invaders most likely to succeed are the ones that look like they already belong, and now, for the first time, science can spot them before they take over.

Subject of Research: Machine-learning prediction of naturalization and invasion of non-native plants in the Brazilian Caatinga dry forest

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

Article References: Almeida, T. S., de Paiva Silva, D., Martinez, P. A., Cruz, A. B. S., da Paixão Menezes, D., Oliveira, E. V. D. S., & Gouveia, S. F. (2026). Machine-learning classification of non-native plant status in the Brazilian dry forest, Caatinga. Discover Conservation, 3(1), Article 1. https://doi.org/10.1007/s44353-025-00073-9

Image Credits: AI Generated

DOI: 10.1007/s44353-025-00073-9

Keywords: invasive species, Caatinga, machine learning, SHAP values, plant ecology, biological invasions, dryland ecosystems, phylogenetic similarity, climatic niche, ecological niche models, XGBoost, Brazil