Implementation of Inertia Sensor and Machine Learning Technologies for Analyzing the Behavior of Individual Laying Hens

SIMPLE SUMMARY: Poultry-welfare regulations have caused a shift from cage housing towards more welfare-friendly systems with more possibilities for the birds to meet their natural behavioural needs. The welfare-friendly systems with litter allow and encourage the hens to perform natural behavior including activities that lead to increases in the amount of airborne dust particles emission from such poultry houses. For successful management of these systems, the behavior of the hens needs to be considered, which is more challenging and time-consuming for the farmer. The main objective of this study was to show a proof of principle to identify, classify and analyze the behaviors of laying hens in three levels of activity by using an inertia sensor and machine learning techniques. The model was able to predict the laying hen behaviors with an accuracy of 90%. The results of such monitoring could be used by farmers in the management of poultry houses. ABSTRACT: Welfare-oriented regulations cause farmers worldwide to shift towards more welfare-friendly, e.g., loose housing systems such as aviaries with litter. In contrast to the traditional cage housing systems, good technical results can only be obtained if the behavior of hens is considered. With increasing flock sizes, the automation of behavioural assessment can be beneficial. This research aims to show a proof of principle of tools for analyzing laying-hen behaviors by using wearable inertia sensor technology and a machine learning model (ML). For this aim, the behaviors of hens were classified into three classes: static, semi-dynamic, and highly dynamic behavior. The activities of hens were continuously recorded on video and synchronized with the sensor signals. Two hens were equipped with sensors, one marked green and one blue, for five days to collect the data. The training data set indicated that the ML model can accurately classify the highly dynamic behaviors with a one-second time window; a four-second time window is accurate for static and semi-dynamic behaviors. The Bagged Trees model, with an overall accuracy of 89% was the best ML model with the F1-scores of 89%, 91%, and 87% for static, semi-dynamic, and highly dynamic behaviors. The Bagged Trees model also performed well in classifying the behaviors of the hen in the validation data set with an overall F1-score of 0.92 (uniform either % or decimals). This research illustrates that the combination of wearable inertia sensors and machine learning is a viable technique for analyzing the laying-hen behaviors and supporting farmers in the management of hens in loose housing systems.