Behavior classification of cows fitted with motion collars: Decomposing multi-class classification into a set of binary problems

Precision management systems for livestock offer the potential to monitor and manage animals on an individual basis. A key component of these sensor based systems are the analytical models that automatically translate sensor data into different behavioral categories. A new methodology was proposed for multi-class behavior modeling based upon the "one-vs-all" framework. This methodology differs from the standard approach to behavior classification where a single classifier is trained to discriminate between multiple behaviors. Instead a set of binary classifiers are trained to each discriminate one of the behavior classes against a combined class of all the remaining behaviors. The confidence scores from the set of binary classifiers are then combined to generate a behavioral estimate. The performance of this new modeling approach is validated across a study involving 24 Holstein Friesian dairy cows that were each fitted with an Inertial Measurement Unit (IMU) on a collar upon their neck. Five general classes of cow behavior grazing, walking, ruminating, resting and "Other" were classified. Binary time series classifiers were tailored to each of the five behaviors through training and validation of each model across a range of configurations including nine window sizes, five classifiers and a feature selection process of 84 candidate input features. Results revealed that the proposed model classified grazing behavior with an extremely high classification accuracy (F-score of 0.98), whilst ruminating and resting behaviors were also classified with a high accuracy (F-scores of 0.87 and 0.85, respectively), and walking was classified with a lower accuracy (F-score of 0.73). The proposed model offered a 5% performance improvement over standard multi-class time series classifiers that was attributed to the "one-vs-all" framework training a classifier for each behavior independently; this created diversity in the behavior model. The feature selection process used in developing each of the binary classifiers found that features representing the motion intensity and pitch of the cow's head were most important to each of behavior's classification. Whilst a minor performance improvement was obtained using the proposed methodology, it is suggested that further performance improvements could be obtained by increasing the diversity of the classifier's inputs. Diversity could be created by fusing the data of other sensors that can be fitted to the cows i.e. GPS tracking unit, pressure sensor and microphone. Crown Copyright (C) 2016 Published by Elsevier B.V. All rights reserved.