NEWS & EVENTS

At LivingData 2025, GUARDEN continued to showcase its pionnering approach in AI-assisted biodiversity monitoring with the presentation of the 2025 PlantCLEF challenge, especially through the talk “Advancing AI-based Multi-Species Plant Identification in Vegetation Quadrats for Supporting Environmental Law and Biodiversity Monitoring.” The work, led by Giulio Martellucci (INRAE), Hervé Goëau (CIRAD), Alexis Joly (Inria), Pierre Bonnet (CIRAD), and Fabrice Vinatier (INRAE), highlights how artificial intelligence can bridge the gap between ecological research and environmental law enforcement.

Recent European environmental regulations, such as Directive (EU) 2024/1203 on environmental protection through criminal law and Regulation (EU) 2023/1115 on imported deforestation, are creating a growing demand for robust and scalable tools to efficiently and cost-effectively inventory and monitor plant communities. Vegetation quadrats, the rectangular ground plots widely used by ecologists, are central to this effort. Yet, traditional quadrat-based surveys are time-consuming, thus costly, and require a taxonomic expertise that is not always available and numerous enough to cover the needs of multiple projects and surveys. PlantCLEF 2025 directly addresses this challenge by introducing a new benchmark for automatic identification of all plant species visible in high-resolution quadrat images, which represents a complex multi-label classification problem that also happens to mirror the realities of field monitoring.

The dataset behind the challenge integrates several complementary components: over 1.4 million single-species images for baseline identification, 2,100 annotated high-resolution quadrats capturing realistic multi-species vegetation plots, and an additional 210,000 unlabeled pseudo-quadrats designed for self-supervised learning. Together, these elements form a unique resource connecting controlled imagery from citizen science platforms like Pl@ntNet with real-world ecological scenes, paving the way for scalable AI-based biodiversity monitoring.

The 2025 edition of the PlantCLEF challenge attracted more than 500 participants and 38 research teams worldwide, submitting 659 distinct methods. Results demonstrated significant progress over previous years: top-performing models achieved a sample-averaged F1 score of 0.37, compared to 0.29 in 2024. Most successful approaches leveraged DINOv2-based vision transformers, often paired with multi-scale tiling strategies to handle high-resolution imagery. The findings underscore the importance of rigorous pre-processing, transparent evaluation pipelines, and reproducible methods. All of them are key principles for building trustworthy AI models in operational biodiversity applications.

By framing quadrat-based identification as a global AI challenge, PlantCLEF 2025 contributed, thanks to the collaboration in the framework of GUARDEN’s activities, by providing a controlled testbed for developing and benchmarking methods that can, in fine, support digital evidence pipelines in environmental law enforcement. From detecting habitat degradation to verifying deforestation-free claims, these advances illustrate how open, benchmark-driven research can produce reliable, scalable tools aligned with FAIR data principles and European environmental policy needs.

Hosted on Kaggle and integrated into both the CVPR-FGVC workshop and the CLEF conference, PlantCLEF reflects the open, collaborative spirit of GUARDEN, but also of the entire AI community when it comes to tackling the many challenges posed by conversation issues.