When the lights go out and the pigs go to sleep, the nightlife cranks up at a different part of the farm – the deadboxes and composting structures. A farm can enforce detailed biosecurity protocols, but at 2 a.m., none of those rules apply to a hungry coyote.
“We can make management strategies and tell our workers to shower in and shower out, but we can’t tell a coyote to do that,” says Gabriella Cattani, a student at the University of Minnesota College of Veterinary Medicine. “He’s just going to do what he’s going to do.”
Across the livestock industry, mortality is routinely handled using compost piles, rendering areas or deadboxes. While essential, these sites act as an open buffet for scavenging wildlife. At a time when wild-wild-animal-borne threats like African swine fever (ASF) and Highly Pathogenic Avian Influenza (HPAI) loom, unmonitored visitors represent a biosecurity blind spot.
“Wildlife have a strong potential to transport pathogens mechanically—whether that’s carrying disease from the carcass to the wild, or from the wild back to the carcass,” Cattani explains.
Picture the chain reaction: A coyote scavenges a carcass, dragging it out of the designated bin into high-traffic gravel. The next morning, a rendering truck rolls over that exact ground, picking up fomites on its tires and carrying them directly to the next farm down the road.
To break that transmission chain, researchers first needed to quantify how often wildlife actually visit these sites. But gathering that data revealed a massive bottleneck.
The Big Data Bottleneck
To monitor farm boundaries, researchers installed 30 solar-powered, motion-activated Reolink cameras facing deadstock structures at swine farms across the U.S.
The cameras worked—a little too well.
Triggered by blowing grass, shadows, and subtle weather changes, the devices uploaded more than 400,000 video clips between December 2025 and June 2026. Manually logging each clip into a spreadsheet was nearly impossible.
“This project came out of pure necessity,” Cattani says. “Reviewing videos by hand quickly became unmanageable. At the time I became involved, we had almost 400,000 videos to sort through.”
Despite having no background in swine research or computer coding, Cattani took on the challenge. She spent the first half of her summer teaching herself Python, diving into machine learning architecture, and turning the video archive into an automated AI detection pipeline.
Teaching AI to See
Using the Minnesota Supercomputing Institute (MSI) and the Ultralytics platform, Cattani built a binary classification system using the YOLO (You Only Look Once) neural network framework.
To train the system, she curated a balanced dataset of 500 representative videos—yielding roughly 5,000 individual frames split into training, validation, and testing sets.
The biggest hurdle? Teaching the model the difference between the swine carcass and the live pest.
“During the training process, I specifically trained the models to ignore the motionless pig carcasses,” Cattani says. “That way, we knew with high confidence that the model was flagging the active scavengers—like coyotes, rats and birds—rather than the mortality itself.”
Gold-Standard Accuracy
The top-performing model, YOLO26n, delivered an overall accuracy of 95.3% and an F1 score of 92.4%. It correctly identified 92.2% of clips containing wildlife (sensitivity) and 96.6% of clips with no animals present (specificity).
Equally important, YOLO26n showed minimal difference between its training and validation loss curves—a technical indicator that the model was genuinely understanding visual patterns rather than simply “memorizing flashcards.”
“In machine learning, over 95% accuracy is considered gold standard,” Cattani notes. “When we deployed the model on unseen footage, we saw the expected minor dips that come with variable field conditions, but performance remained high. If I went through 400,000 videos by hand, they wouldn’t match that consistency.”
What It Means for Producers
When analyzing the positive footage, Cattani found that vultures accounted for nearly 90% of all animal visits, followed by coyotes, rodents and occasional smaller species like armadillos and small birds.
For producers, having automated, real-time alerts transforms farm security from guesswork into targeted, cost-effective infrastructure upgrades:
- Capacity Issues: If cameras catch mortalities spilling over bin walls, it signals the need for a larger deadbox.
- Ground Scavengers: Frequent coyote or rodent detections justify installing heavier locking lids, fences, or reinforced gates.
- Avian Pressure: Heavy vulture roosting points to the need for overhead netting or roof structures.
“It’s not a question of whether or not wildlife are present on farms; they are,” Cattani says. “The real question is: How do we adapt our structures so they can’t make contact in the first place? In the face of an emerging disease outbreak, small modifications might make all the difference in keeping a herd safe.”
