Comparison among grazing animal behavior classification algorithms for use with open-source wearable sensors

Behavioral monitoring for pasture-based production has the potential to improve the efficiency of livestock operations without increasing labor costs. The application of technologies for remotely monitoring animal behavior has expanded rapidly in the last decade, especially in confinement operations. However, automated systems for extensive operations are limited by the interrelated challenges of power use and data transmission. The objective of this study was to explore behavioral classification techniques suitable for using an open-source wearable sensor system deployed on extensively grazed cows. Behavior classification analyses leveraged simple approaches (analysis of variance and logistic regression), as well as more complex machine learning algorithms (support vector machine (SVM) and random forest (RF)) to better understand the trade-offs between classification approach complexity and accuracy. Behavioral observations were conducted by two independent observers at the field. Algorithms were used to classify four behaviors: grazing, lying, standing, and walking using data aggregated across either 1-second or 1-minute intervals. Algorithms were also compared under situations assuming continuous monitoring compared with periodic snapshots of data representing a scenario where the sensor was only activated every 3 or 5 s. Grazing was the most accurately (93 %) classified behavior followed by laying (92 %) when a 1-minute timestep was used to train the RF model. At both timesteps, SVM and RF were capable of distinguishing among behaviors with improved accuracy compared with simplistic approaches. Similar accuracies were found when evaluating the RF model on the 3 and 5-second iteration, indicating power saving may be achieved by periodic, rather than continuous sampling. As microprocessors continue to advance in terms of their capacity to execute machine learning algorithms, these approaches may help improve the usability of inertial measurement unit sensors for behavioral monitoring in extensive production systems.