At a time when millions of Americans have turkey on the mind, a team of researchers led by an animal scientist at Penn State have tested a way for farmers to keep an eye on their turkeys using machine vision. Crucial for productivity and animal welfare, monitoring the behavior and health of poultry animals on large, commercial farms is costly, time-consuming, and labor-intensive. The new scheme changes that by employ a small drone equipped with a camera and computer vision to automatically recognize what turkeys are doing. The research was the first to test whether a drone combined with a computer vision model could automatically detect different turkey behaviors from overhead video. From the videos, the researchers took individual image frames and manually labeled the turkeys’ behaviors, including feeding, drinking, sitting, standing, perching, huddling and wing flapping. Courtesy of Penn State. The research team used a commercially available drone with a regular color camera to record video four times a day. The test monitored 160 young turkeys from five to 32 days old at the Penn State Poultry Education and Research Center. The drone was designed to ensure full area coverage from the camera footage during each flight. “This work provides proof of concept that drones plus AI can potentially become an effective, low-labor method for monitoring turkey welfare in commercial production,” said Enrico Casella, assistant professor of data science for animal systems in Penn State's College of Agricultural Sciences. “It lays the groundwork for more advanced, scalable systems in the future.” Using the videos, the researchers took individual image frames and manually labeled the turkeys’ behaviors. They created a dataset of over 19,000 instances of labeled behaviors, including feeding, drinking, sitting, standing, perching, huddling, and wing flapping. The images were used to train, test, and validate their computer vision model called YOLO, used to detect objects and actions in images. The researchers tested several YOLO versions and found that the best model could correctly find 87% of all present behaviors and accurately detect specific behavior 98% of the time. These metrics are good, Casella said, especially for behavior classification in a real farm environment, which is often visually messy and challenging. “The study shows that a drone-equipped AI system can accurately detect turkey behaviors,” he said. “This method could reduce labor demands, it could allow continuous, non-invasive monitoring of bird welfare in commercial farms, and it may also reduce the need for constant human presence, lowering training and staffing burdens.” This research was published in Poultry Science (www.doi.org/10.1016/j.psj.2025.106103).