AI-based Animal and Biodiversity Monitoring

In the domain of animal and biodiversity monitoring, we investigate computer vision and multi-modal machine learning methods for AI-based automated systems, with the aim of enhancing animal welfare and biodiversity while reducing reliance on manual observation. We currently investigating pigs, cattle, and insects, while remaining open to extending our research to other farm animals, such as poultry and fish.
Pig welfare
When it comes to pig welfare, vocalizations provide direct information about health, stress, and social state, often earlier and more cost-effectively than manual observation, while also offering a non-invasive means of monitoring. The challenge is that recordings on real farms are collected under uncontrolled conditions, with background noise, reverberation, and sensors that can degrade without warning. Moreover, no single acoustic representation captures all relevant information reliably under these conditions. We work on combining complementary acoustic representations while accounting for their reliability in different conditions, with the aim of developing a vocalization-based monitoring tool robust enough for everyday use on real farms.
Insects
When it comes to biodiversity, insects are key indicators, yet their populations are declining and are still monitored largely through slow and costly manual identification by expert entomologists. This is further complicated by the general lack of standardized data. Developing an automated classifier is challenging because of the inherent characteristics of insect populations: some species are much rarer than others in the data, and meaningful classification spans multiple taxonomic levels, from broad groups down to individual species, rather than relying on a single flat label. We work on computer vision methods that classify insects across this hierarchy, with the goal of developing tools capable of keeping pace with the volume of data generated by modern camera traps.
Cattle
When it comes to cattle, non-invasive monitoring through cameras is a common approach. One of the first steps is reliably identifying the same animal across days and months, allowing its health and status to be tracked over time. This makes re-identification a central problem. Tracking individual cows across a farm and detecting the onset of calving still relies heavily on continuous human observation, which is costly and difficult to sustain around the clock. We work on camera-based methods for following individual animals across multiple cameras and combining this information with behavioral cues, with the aim of providing earlier and more reliable warnings while substantially reducing the need for continuous human supervision.
If you would like to apply the results of our research to your farm or business, or if you are interested in collaborating with us on scientific or applied projects related to AI-based animal monitoring and biodiversity, please feel free to contact us. Students interested in completing a final project or thesis with us are also welcome to get in touch.
Lead Researcher: Dr. Ehsan Yaghoubi
Resources:
Github repository:
https://github.com/smAIL-WS/Uncertainty-Aware-Crossmodal-Fusion-for-Classification-of-Animal-Behavior
https://github.com/smAIL-WS/HierarchicalInsectClassification
Publications:
Mahfoud, Z., Chiavassa, J. A., Walther, S., Haselbeck, F., & Yaghoubi, E. (2026). Deep learning-based hierarchical insect classification using camera trap imagery. arXiv preprint arXiv:2607.28005. Under Review
Yaghoubi, E., & Haselbeck, F. (2026). Uncertainty-Aware Crossmodal Fusion for Classification of Animal Behavior. arXiv preprint arXiv:2608.02104. Under Review