Development of a device to record and classify behaviour on grazing goats.

GONZALEZ, R.; VALLI, F.; PAEZ LAMA, S.; SBRIGLIO, L.; CATANIA, C.; ALLEGRETTI L.

Mendoza

Otro; XXVI Reunión Científica Anual de la Sociedad Biológica de Cuyo; 2018

Behavioral studies of grazing animals allow a better understanding of how animals use the resources of grasslands. They also help to develop herd management tools to improve the sustainability and productivity of goat production systems. However, behavioral assessment is difficult and, until recently, animal activity was only quantifiable through direct observation or video monitoring, very laborious and time consuming methods. Recent advances in sensor technology allow predicting behaviour from information recorded on devices carried by animals. In this work, the objective was to build a device to record and classify the behavior of goats grazing in a desert rangeland. A Pixhawk autopilot was used as a platform to store the measurements of different sensors: 3 accelerometers, 3 gyroscopes and a GPS receiver. GPS and Pixhawk were powered by a 4000 mAh battery. All sensors, Pixhawk, GPS and battery were stored in a plastic box custom built using 3D printing techniques. This box was fixed to the head of a goat with a camera to record the behavior. Then the goat grazed freely in a natural pasture of the Monte desert (Mendoza, Argentina). After grazing, the device was recovered and the information processed. An activity predictor was built using the recorded data. The video record was used to classify the behaviour in four states: resting in the pen (RP), resting in the field (RF), walking (W) and grazing (G). Time-series corresponding to each sensor was splitted in a one-minute time window and labeled according to the four types of activities. A procedure known as bag-of-features was then applied to extract the predictor variables used by the model. The bag-of-features is a method that consists of applying a clustering algorithm on the vector composed of the sensors measures and then building a histogram of the number of vectors belonging to each cluster. This is used to train a statistical learning model known as Random Forest (RFO). The evaluation of the RFO performance was conducted on a dataset of 777 one-minute time windows. The 70% of the original dataset was used for building and fine tuning the model. Then, the model was evaluated on the remaining 30% of the dataset. The total time of data collection was 12 hours and 57 minutes, from which goat spent 15.5, 18.3, 34.3, and 32.0 % in the activities of W, RP, RF and G, respectively. The precision achieved with the device to correctly detect and classify the activities was 96.1, 73.0, 89.0, and 75.9 % for W, RP, RF and G, respectively. These results show that it is possible to register and classify, with an acceptable precision, the behaviour of a goat grazing in the Monte desert.