Wildlife Scientists Boost AI Accuracy for Species ID With Smarter Image Selection

Wildlife Researchers Enhance AI for Species Identification in Camera Trap Photos

Motion-activated cameras are widely used in wildlife monitoring, but manually reviewing the vast number of images they produce is time consuming and inefficient. Current AI models designed to assist with this task often fall short in accuracy, especially when applied to new locations they were not trained on.

Scientists at Oregon State University have developed a more effective AI training method that improves species recognition in wildlife photos. Their approach focuses on training models with less data but more targeted and varied environmental information, resulting in higher accuracy and reduced computational demands.

Improving AI Accuracy with Focused Training

One key challenge in wildlife AI applications is maintaining accuracy when analyzing images from previously unseen locations. According to Christina Aiello, a research associate involved in the study, their method improved classification accuracy both at familiar and novel sites by carefully selecting training data.

The research team, including undergraduate Owen Okuley and professor Clinton Epps, tested their approach using bighorn sheep as a model species. They found that training AI models on a single species with diverse background environments yielded better results than training on multiple species at once.

Less Data, Better Results

Contrary to the common assumption that more training images always improve AI accuracy, the study showed diminishing returns beyond a certain point. Okuley explained that using about 10,000 well-selected images — far fewer than typical datasets — produced nearly 90% accuracy in identifying bighorn sheep across different locations.

This reduction in required data also lowers computing power and energy consumption, which benefits both researchers and the wildlife habitats they study.

Practical Experience Fuels Research Innovation

Okuley’s involvement began through a mentoring program, where he gained hands-on experience with camera trap data and genetic sampling. Leading this AI project provided him with a comprehensive view of research processes, from conceptualization to publication.

He plans to pursue a PhD focused on expanding AI applications to identify waterfowl species and hybrids, aiming to refine wildlife monitoring tools further.

Collaborative Efforts and Support

This research involved experts from Johns Hopkins University, the California Department of Fish and Wildlife, and the National Park Service. Funding and support came from the National Park Service’s Pacific Northwest Cooperative Ecosystem Studies Unit, Oregon State University, and the College of Agricultural Sciences Continuing Researcher Support Program.

Reference

Okuley OS, Aiello CM, Glad W, et al. Improving AI performance in wildlife monitoring through species and environment-specific training: A case study on desert Bighorn sheep. Ecological Informatics. 2025. doi: 10.1016/j.ecoinf.2025.103179

For professionals interested in AI applications in environmental research, exploring targeted training approaches can enhance model performance while reducing resource demands. More training options and resources on AI in science can be found at Complete AI Training.