Document type: scientific article published in *Smart Agricultural Technology*
Authors: Panqi Pu, Junge Wang, Yue Gao, Dapeng Li, Geqi Yan, Hongchao Jiao, Xianyao Li, Hai Lin
Abstract in French (translation): Recognition of multiple behaviors and assessment of the health status of laying hens using the RSA-YOLO algorithm in the context of animal welfarefarming
In animal welfarefarming, the behavior of laying hens is a key indicator of health, but its automated assessment often remains limited. This study presents a framework for recognizing multiple behaviors and assessing health status using an enhanced RSA-YOLO model. By integrating RFAConv, C3k2_KSFA, and the ATFL loss function into YOLOv11n, this lightweight model (2.75 million parameters) achieved an mAP@50 of 88.3% for six behaviors (standing, feeding, perching, preening, feather pecking, and resting), thereby outperforming baseline models. A composite health index (HCI) was developed, weighting feeding, movement, and resting behaviors. A pilot study revealed that a healthy reference cage had a higher HCI (98.2) and feeding rate (32.6%) than a cage with mortality cases (70.1 and 12.4%, respectively). In this pilot study, it is important to note that behavioral changes, such as a decrease in feed intake, preceded a decline in egg production by approximately one week, thereby demonstrating the early-warning potential of this analytical framework. This research proposes an automated and quantitative tool for on-farm health monitoring in the poultry industry.
Preview: In welfare-oriented farming, laying hen behavior is a critical health indicator, yet automated assessment is often limited. This study introduces a framework for multi-behavior recognition and health assessment using an improved RSA-YOLO model. By integrating RFAConv, C3k2_KSFA, and ATFL loss into YOLOv11n, the lightweight model (2.75 M parameters) achieved 88.3% mAP@50 for six behaviors (standing, feeding, roosting, preening, feather pecking, resting), outperforming baseline models. A Health Composite Index (HCI) was developed, weighting feeding, movement, and resting behaviors. A pilot study revealed a healthy reference cage had a superior HCI (98.2) and feeding ratio (32.6%) compared to a cage with mortality (70.1 and 12.4%, respectively). In this pilot study, significantly, behavioral changes, such as decreased feeding, preceded a drop in egg production by approximately one week, demonstrating the framework’s early-warning potential. This research offers an automated, quantitative tool for on-farm health monitoring in poultry farming.
