XAI tools to predict biological invasiveness from phenotypes: A case study in plants

Determining which functional traits relate to invasion success is crucial for understanding how biological invasions develop and for predicting which species are likely to become invasive. Yet, studies relating morphological traits to invasiveness remain limited in scope and scale, and evidence for consistent morphological differences between invasive species and their non-invasive relatives is still sparse. Here, by focusing on the plant genus Lythrum as a case study, we develop a deep learning pipeline to distinguish invasive species from their non-invasive relatives by identifying discriminative morphological traits from images. We first developed an image-based classification model. We then applied an explainability pipeline to identify the image regions contributing to model predictions. Finally, we integrated prediction outcomes with region-level attributions to determine which morphological traits influenced invasiveness classification. Results showed that classification was driven primarily by morphological trait combinations characteristic of non-invasive species, rather than by the presence of invasive-specific traits. This supports the idea that invasiveness is not consistently associated with a fixed set of morphological traits, but rather reflects a relaxation or disruption of trait combinations characteristic of their non-invasive relatives. In general, this study showed that image-based models can recover biologically meaningful information from heterogeneous plant photographs and that their predictions are shaped by interactions among morphological traits rather than by single diagnostic features. By integrating an explainability pipeline, embedding analysis, and trait-pair attribution, these findings highlighted both the potential and limits of image-based approaches for invasion assessment, offering a scalable and interpretable method for linking deep learning to trait-based invasion ecology.