3d: High-throughput species recognition for ecological research

3d:     High-throughput species recognition for ecological research

Conveners:     Niels Raes (Naturalis Biodiversity Center)

                      Jacob Kamminga (University of Twente)

                      Patrick Jansen (Wageningen University)

1. Inferring plant community phenology via bee-collected pollen

Sydney B. Wizenberg, Mateus Pepinelli, Bao Ngoc Do, Mashaba Moubony, Darya Tamashekan, Ida M. Conflitti, Barbara Gravendeel, Amro Zayed, Naturalis Biodiversity Center

Phenological reshuffling, changes to community-level patterns of developmental timing, is one of the largest expected impacts climate change will have on plant communities.  Phenological timing is a strong determinant of the temporal aspect of plant-pollinator interactions, and thus reshuffling presents a potent evo-eco threat as it runs the risk of destabilizing crucially important pollination networks. Documenting these shifts and their impact on plant-pollinator interactions has been difficult in the past, requiring the development of new techniques to modernize this field. Here, we present a high-throughput approach for inferring the flowering phenology of plant communities through multi-locus metabarcoding of bee-collected pollen. Pollen collected by bees can accurately indicate the on-set of flowering in a wide variety of plant species, providing a new avenue for mapping phenological changes via the plant-pollinator interface.

2. Large-scale animal sounds recognition using deep learning and transfer learning

Burooj Ghani, Vincent Kalkman, Dan Stowell, Naturalis Biodiversity Center

Audio-based animal species recognition plays an invaluable role in the ongoing efforts to monitor and conserve species biodiversity. Leveraging deep learning for audio-based species recognition allows for the rapid and automatic processing of vast amounts of audio data, which is paramount for large-scale ecological studies and real-time monitoring systems. Consequently, advancements in this area have the potential to significantly enhance our understanding of animal populations, their behaviors, and their responses to environmental changes, contributing essential insights for conservation planning and management. Large-scale sound classifiers are becoming available, but in order to use them for species monitoring at the scale of a continent we need systems to have high recognition across many species, and well-validated performance across habitats of interest. In this talk, we delve into the application of advanced deep learning techniques to recognize and classify bird and grasshopper species robustly for the whole of mainland Europe, solely based on their distinct vocalizations. The challenging nature of this large-scale classification problem, compounded by the inherent variability in vocalizations, underscores the need for robust and efficient learning models. We exploit the power of deep learning and transfer learning to design classifiers capable of learning effectively from weakly labeled data. This is particularly beneficial in our context, given the significant challenges associated with obtaining highly accurate labels for large volumes of species vocalizations. Results from our analyses will be presented, highlighting the potential of our approach in handling large-scale, real-world classification tasks. Our research underscores the transformative potential of deep learning in advancing bioacoustics and contributes to the ongoing discourse on the practical applications of AI in wildlife conservation.

3. DIOPSIS: automatic cameras and image recognition for standardized, continuous monitoring of insect populations

Eelke Jongejans, Chantal Huijbers, Stephan Peterse, Laurens Hogeweg, Theo Zeegers, Marten Schoonman, Radboud University

To improve our understanding of trends in insect species occurrence and abundance, automated monitoring systems can provide a non-invasive, cost-effective, and standardized method. Here we present the DIOPSIS camera system, which was developed and tested at many sites in the Netherlands. The system includes a digital camera with a yellow screen that attracts insects which are photographed when their motion is detected. Specialized deep learning software was developed to analyze the images for insect detection, taxonomic classification, deduplication, and biomass estimates. Deployment of 100 cameras in 2021-2023 resulted in millions of pictures and large counts of flies, mosquitoes, moths, true bugs, ants, caddisflies and mayflies. Higher numbers per hour were recorded at night, but insects were also photographed during daytime. Insect abundance was highest in natural, agricultural or urban environments, depending on which insect order was assessed.

4. NestMoni: Advancing Scalable Biodiversity Monitoring with AI and IoT

Jacob Kamminga, Peter Hoekstra, Jeroen Vonk, Vladimirs Popovs, Jesse Visser, and Max Lievense, University of Twente

Effective biodiversity monitoring is essential for understanding ecosystem health, yet scalable and accurate data collection remains challenging. Ecomoni's Nestmoni addresses this with an AI-powered camera system for monitoring birds in nest boxes. It detects and classifies events such as entries, exits, and prey types such as caterpillars and spiders. Configurable via an IoT dashboard, it supports group management, adaptive power schedules, and real-time reporting, enabling efficient and detailed monitoring at scale.

This presentation highlights challenges and shares early results from Nestmoni deployments in Hengelo and Utrecht. In collaboration with Emily Burdfield (UvA), we are developing an open-source dataset to refine prey classification. We use AI to filter common prey such as caterpillars and focus annotation on rarer categories like beetle larvae.

The Nestmoni deployment demonstrates Ecomoni's strong potential for scalable biodiversity monitoring and can extend to other applications, such as bat activity tracking, providing a robust framework for ecological research.

5. Mainstreaming and upscaling of high-throughput species monitoring

Chantal Huijbers, Niels Raes, Patrick Jansen, Naturalis Biodiversity Center

Advanced technologies, such as digital sensors and DNA analysis, offer powerful tools to address the urgent need for better monitoring of biodiversity. However, processing and managing the vast volumes of data generated by these methods and making this data findable, accessible, interoperable and reusable is a challenge. The ARISE infrastructure is being developed to provide easier access to species capture and identification services by providing end-to-end pipelines from sensors to AI-based analysis and from sample collection to DNA analysis. The Biodiversity Meets Data - BMD - project will develop a European Single Access Point to high-throughput biodiversity monitoring tools making use of the ARISE technology, aggregated environmental data, and Virtual Research Environments (VREs) to support biodiversity assessments, trend analyses and future projections which is needed to support EU policy implementation and policy making. Here, we present how ARISE and BMD enable the upscaling and mainstreaming of biodiversity monitoring.