A new MAMBO publication investigates the potential of large-scale plant identification models for the early detection of invasive alien plant species using high-resolution roadside imagery. The study, titled “Adapting a global plant identification model to detect invasive alien plant species in high-resolution roadside images”, is the result of a collaboration between MAMBO and the GUARDEN projects.
The research addresses the critical need for early detection of invasive plants, which are a major threat to biodiversity and frequently spread along transport networks such as roadsides. While vehicle-mounted imaging systems offer a promising method for large-scale monitoring, applying artificial intelligence to process this data efficiently remains a challenge. Traditional deep learning methods, such as object detection and segmentation, require extensive, labour-intensive annotation work, making them difficult to scale across the thousands of invasive species present worldwide.
In this study, the authors explore alternative approaches that are less dependent on annotated training data. They assess the potential of two methods: a multi-label classification model and a tiling-based model, both using a vision transformer originally developed by the Pl@ntNet project. These models were tested on high-resolution roadside images, both in their original form and after fine-tuning.
Results show that the tiling-based model performs notably well even without fine-tuning, and slightly better than the classification model after fine-tuning. This suggests a viable path forward for using pre-trained plant identification models to detect invasive species with minimal additional resources or manual adaptation.
This work supports MAMBO’s mission to develop scalable, automated tools for biodiversity monitoring and contributes to the wider goal of improving ecological surveillance through AI and citizen science data integration. The article has been published in the Ecological Informatics journal. Read the full article here.