Camera, AI assess turkeys, predict future body weight in study

UNIVERSITY PARK, Pa. — Body weight monitoring is one of the most important tools for poultry farmers. It provides valuable information about growth, uniformity, feed conversion efficiency — how well birds turn feed into meat — and the occurrence of diseases in a flock. But obtaining frequent and accurate body weight measurements is a significant challenge in commercial poultry production, according to researchers at Penn State who may have developed a solution. They recently found that a camera paired with artificial intelligence (AI) could monitor individual turkeys and make reasonably accurate bodyweight predictions up to three weeks into the future.

The team, who published their findings in Frontiers of Animal Science, said computer vision and AI can potentially be used not only to estimate a turkey’s current body weight, but also to predict its future body weight with approximately 93% accuracy.

“Traditional approaches to individual body weight monitoring require extensive manual labor and frequent animal handling, creating both economic and animal welfare concerns,” said study senior author Enrico Casella, assistant professor of data science for animal systems in Penn State’s College of Agricultural Sciences. “In the poultry industry, weight information is essential for maximizing the value of each individual animal, as the average weight of a flock determines equipment settings for processing operations.”

The study, conducted at the Penn State Poultry Education and Research Center, involved 30 male turkeys that were housed together and observed them from day 37 to day 133 of age, almost 14 weeks of growth monitoring. The researchers used a camera positioned above the birds to capture two types of information: normal color images and depth images, which provided information about how far different parts of the turkey were from the camera. This information revealed more details about the three-dimensional shape and size of the bird, which are informed by its weight.

Because there often were multiple birds in the camera’s view, the AI had to learn that a group of pixels — the smallest single points or dots that make up a digital image on a screen — belonged to one turkey, while another group of pixels showed a different turkey. Distinguishing between the two is called instance segmentation, and it’s different from simply recognizing “there is a turkey,” Casella said. The system needed to identify each individual pixel belonging to the individual animal, as opposed to pixels that belong to the background or a different animal.

To train and test the AI, the researchers manually weighed the turkeys five times per week. These actual weights served as references for the AI. They employed a type of deep-learning neural network called ResNet commonly used for analyzing images. In this study, ResNet was trained to look at information from the turkey images and learn relationships between the birds’ appearance, size and shape and relate them to actual body weights.

The camera-AI system’s prediction of future body weight was nearly as accurate as estimating the turkey’s current weight. That's significant, Casella explained, because previous computer-vision systems were not capable of temporal forecasting across time. That has major implications for poultry production, he noted, because similar forecasting had not previously been explored in the industry.

“Imagine a commercial turkey farm with thousands of birds … If farmers could reliably estimate weights using cameras, they wouldn’t need to catch and weigh large numbers of birds manually,” Casella said. “This technology could potentially help with monitoring growth, identifying birds that aren't growing normally, planning feed strategies, predicting when birds will reach market weight, planning processing schedules, reducing animal handling and labor, and improving production efficiency.”

The study is promising, but it doesn't mean the technology is ready for every commercial turkey farm, Casella pointed out. The experiment involved only 30 turkeys under relatively controlled conditions. Larger studies are needed, he said, to determine how well the model works across different farms, breeds, environments, lighting conditions, poultry densities and much larger populations.

First author Mireia Molins, who graduated with a master’s degree in animal science earlier this year, was first author on the study. She is now an animal science field engineer with NovaTech Engineering in Willmar, Minnesota.

The research was supported by Penn State Institute for Computational and Data Sciences, which provided access to computational research infrastructure; and the U.S Department of Agriculture’s National Institute of Food and Agriculture and Hatch Appropriations under project number PEN05072 and accession number 7009799. This content is solely the responsibility of the authors and does not necessarily represent the views of the funders.