PSXI-2 A computer vision system for feed bunk management in beef cattle feedlot

Abstract

Feed bunk scoring is a common management practice in feedlots. Usually, the bunk score is assigned visually by a trained person. However, the subjectivity of bunk scoring and inconsistency across bunk readers can result in excessive variation on feed delivery. Such variation can result on feed waste, sub-optimal animal performance, and increased incidence of metabolic disorders. The objective of this study was to develop an artificial intelligence system to perform bunk management in real-time. RGB-cameras were installed above the feed bunk in a commercial feedlot, and a total of 4,280 images were acquired, together with visual bunk scores with four categories: empty (no feed remaining), low (scattered feed remaining), medium (30–50% of feed remaining), and full (> 50% of feed remaining). Cattle behavior at the feed bunk was also classified into four classes: empty (no cattle at the feed bunk); low (< 30% bunk occupied); medium (30–70% feed bunk occupied); full (above 70% feed bunk occupied). The labeled images were then used for model training and a new set of 105 images were used for validation. A deep neural network (DNN) called ResNet was implemented to generate the predictions using a transfer learning with weights from the ImageNet dataset. A cloud computing system was developed to acquire, process and store images every 15 minutes, and implement real-time predictions of bunk score and cattle behavior. Prediction accuracies across bank score categories were: 81.8% (empty), 82.4% (low), 88.8% (medium), and 90% (full). For cattle behavior, accuracies were: 83.7% (empty), 66.6% (low), 71.4% (medium), and 86.6% (full). Combining feed bunk score and cattle behavior can provide an important decision-making tool to improve nutritional management in beef cattle feedlot. The use of artificial intelligence can allow the development of fully automated real-time systems to enhance livestock operations.