NEWS & EVENTS

GUARDEN and MAMBO highlighted in LivingData 2025

Last week, Bogotá hosted LivingData 2025, an event jointly organized by TDWG, GEO BON, GBIF, and OBIS. This major scientific conference has become a space where researchers, data specialists, AI developers, and policy actors come together to rethink how open biodiversity data can serve real-world action. And of course, GUARDEN was there! You can check out our previous article on Rémi Palard’s presentation of GeoPl@ntNet right here.

Artificial Intelligence, as a key component of GUARDEN, took center stage through the presentation of Pierre Bonnet, researcher at CIRAD and coordinator of the GUARDEN project. His talk, “AI-based citizen-generated plant observations as forensic evidence in biodiversity investigations,” explored how the billions of plant observations gathered via citizens through platforms like Pl@ntNet, iNaturalist, or Flora Incognita can be leveraged as forensic evidence for biodiversity monitoring and research.

Drawing on vast datasets of geolocated and timestamped images enhanced by artificial intelligence, the presentation illustrated how participatory platforms could support efforts to monitor deforestation, combat illegal plant trade, or even track land-use changes. Within the framework of the EU Deforestation Regulation, such information could strengthen traceability and accountability mechanisms, ensuring deforestation-free sourcing and compliance with environmental standards.

To be legally credible however, citizen-generated data must meet strict criteria of traceability, reliability, and transparency from image capture to analysis. Artificial intelligence plays a key role in species recognition, anomaly detection, and cross-checking against lists of protected or invasive taxa, but expert validation remains indispensable. Provenance frameworks are essential for ensuring such transparency, reproducibility, and ultimately, the credibility of these observations. As demonstrated by concrete examples, like the monitoring of invasive species in Réunion Island, well-structured citizen science can effectively complement institutional monitoring systems. At the same time, its biases and data quality limitations are carefully addressed.

Beyond its technical dimension, one of GUARDEN’s core ambitions over the past three years has been to improve the trustworthiness of citizen science-produced data for researchers and policymakers, while simultaneously providing wider access to large-scale, reliable information. This effort has enabled the establishment of cutting-edge, AI-powered data streams that transform observations into useful information for monitoring biodiversity and helping managers implement environmental policies, from local to global levels.

Extending the Vision: The MAMBO Project and AI for Invasive Species Monitoring

One of GUARDEN’s sister projects, MAMBO, shares the same ambition, working hand in hand with GUARDEN to develop AI-driven methods for large-scale biodiversity monitoring. The presentation highlighted a joint study by French and Danish teams showing how a globally trained plant‑identification AI can be adapted to detect multiple invasive plants in roadside images, advancing GUARDEN’s goal of faster, scalable monitoring and early management action.

The research tackled one of the pressing challenges in biodiversity monitoring: detecting invasive alien plants, which are one of the main drivers of biodiversity loss and ecosystem degradation. Roadsides, in particular, act as major pathways for their spread, yet monitoring them efficiently at a large scale remains a logistical and computational challenge. The study explored how Pl@ntNet’s trained AI model, originally built for identifying individual species from citizen-contributed images, can also be adapted to detect multiple invasive taxa in complex, real-world environments such as roadside imagery.

Using the Danish Road Dataset, more than 14,000 high-resolution images collected with a vehicle-mounted camera, two model adaptation strategies were compared: ‘Variable Model Input Size’ or VaMIS, and ‘Tiling’, a method that uses dividing images into smaller sections for multi-scale analysis. Both approaches yielded remarkable performance after fine-tuning, reaching impressive results, going over 90% accuracy in identifying some invasive species. The Tiling method, in particular, showed strong generalization potential even without retraining, offering an efficient solution for large-scale, annotation-free detection.

By reusing Pl@ntNet-trained models, MAMBO demonstrated how open AI infrastructures can reduce annotation costs, improve on the challenge of scalability, and extend biodiversity monitoring to imagery sources such as roadsides.

Together, GUARDEN and MAMBO illustrate how AI-driven tools, open data, and citizen science can, once used together, support biodiversity monitoring. By connecting field observations, artificial intelligence, and regulatory frameworks, these projects continue of improve on a new generation of digital tools that make environmental protection both scalable and accountable on an unprecedented degree.