Classifier’s Performance for Detecting the Pecking Pattern of Broilers during Feeding

Abstract

Broiler feeding is an efficient way of evaluating growth performance, health, and welfare status. This assessment might include the number of meals, meal period, ingestion rate, meal intervals, and the proportion of time spent eating. These parameters can be predicted by studying the birds’ pecking activity. The present study aims to design, examine, and validate classifying algorithms to determine individual bird pecking patterns at the feeder. Broilers were reared from 1 to 42 days, with feed and water provided ad libitum. A feeder equipped with a force sensor was installed and used by the birds starting at 35 days of age, to acquire the pecking force data during feeding until 42 days. The obtained data were organized into two datasets. The first comprises 17 attributes, with the supervised attribute ‘pecking detection’ with two classes, and with ‘non-pecking’ and ‘pecking’ used to analyze the classifiers. In the second dataset, the attribute ‘maximum value’ was discretized in three classes to compose a new supervised attribute of the second dataset comprising the classes’ non-pecking, light pecking, medium, and strong. We developed and validated the classifying models to determine individual broiler pecking patterns at the feeder. The classifiers (KNN, SVM, and ANN) achieved high accuracy, greater than 97%, and similar results in all investigated scenarios, proving capable of performing the task of detecting pecking.

Keywords

• broiler feeding
• data mining
• precision livestock farming