Mainstreaming high-throughput biodiversity monitoring using AI

What this was about

Niels Raes, coordinator of the Horizon Europe project Biodiversity Meets Data (BMD), described how the project is scaling AI-based high-throughput monitoring (camera traps, audio and eDNA) to site managers across Natura 2000 and Emerald Network sites. He showed Naturalis's ARISE platform: 4G sensor upload, data management, a library of image and audio AI models, overnight processing on supercomputers, and export as Camtrap DP for publication to GBIF matched to the Catalogue of Life. He also gave compute and cost figures for inference and for training Naturalis's multi-source model. Q&A covered invertebrate monitoring, the cost of deployment on devices, and human-in-the-loop verification.

Why it matters. It shows a working, continental-scale pipeline from automated sensors to standardised GBIF data and management tools, with rarely shared figures on the compute cost of the AI involved.

In the room

• BMD is a Horizon Europe innovation action that mainstreams existing technology to new stakeholders. The primary audience is natural resource site managers, then policymakers, with academia last.
• Scope: better implementation of the EU nature directives across about 27,000 Natura 2000 sites, extended to the Emerald Network (UK, Norway, Switzerland, Ukraine), about 30,000 potential sites, as a pilot. The project is in month 19.
• 'Virtual research environments' were renamed 'biodiversity analysis tools' because the term did not resonate with stakeholders. The tools are co-designed with stakeholders, for example a biodiversity status tool showing which directive-listed species were observed.
• The target workflow is 4G-connected sensors with solar or battery power, automated upload, remote status control, archiving, AI identification and publication to GBIF.
• ARISE (Authoritative and Rapid Identification System for Essential biodiversity information, Naturalis) offers audio models developed under MAMBO (birds, bats, frogs, insects, cicadas/crickets) and image models (MegaDetector, SpeciesNet, and Naturalis's multi-source model).
• MegaDetector runs first to remove humans and licence plates for GDPR. Processing is batched overnight on supercomputers because on-the-fly processing is costly.
• Results are exported as Camtrap DP, mapped to Darwin Core, published to GBIF and matched to the Catalogue of Life, then feed analysis tools such as species distribution models of climate change effects.

Notable moments