Labeled sensor dataset of beef cattle behavior grazing desert rangelands

Applications of Precision Livestock Management are widely used in intensive production systems, mainly in pasture-based and conventional feeding operations where animal behavior can differ significantly compared to more extensive grazing systems operating on native rangeland. Few studies involved beef cattle production systems operating on rangelands that apply wireless data transmission and ML analytics in real-time.The objective of this study was to evaluate the performance of multiclass ML models for classification of beef cattle behaviors on arid rangeland pastures. The goal was to develop and deploy a real-time animal monitoring system integrating wearable accelerometers, LoRaWAN wireless transmission, ML analytics and a user-friendly dashboard application. This system would allow innovative approaches for beef cattle management on rangelands focusing on production efficiency, sustainability, and improved animal welfare.A case study was conducted to train and test ML classifiers of animal behavior, specifically designed for beef cows managed on large pastures of arid rangeland. Cattle used in this study were born and raised on their respective ranches, and were well adapted to the climatic, edaphic and vegetation conditions of the Northern Chihuahuan Desert. Mature cows (4-9 years old) were used for behavior observations. Focal cows included six Angus-Hereford crossbreed cows randomly selected from a herd of 25 and six Raramuri Criollo cows selected from a herd of 54 at the Jornada Experimental Range (USDA). Additionally, six Brangus cows were selected from a herd of 27 and six Brahman cows were selected from a herd of 22 at the Chihuahuan Desert Rangeland Research Center (NMSU). All cows were equipped with Compact Trackers (CT).The CTs consisted of a multifunctional tracking device equipped with a Global Navigation Satellite System (GNSS) receptor and an embedded triaxial accelerometer sensor. The triaxial accelerometer assessed motion intensity in three dimensions (X, Y, Z). Motion was measured in inertial units at a 12.5 Hz frequency and expressed as <i>g</i> force units. Proprietary onboard processing algorithms summarize the motion data into a one-dimension motion index (MI) aggregated every 1 minute. Cow behavior observation were performed using video recording. All video recordings were reviewed and annotated by the same trained operator and one trained assistant. Annotations included two main motion states, Active (AC) or stationary (ST), two mutually exclusive AC behaviors, walking (WA) and grazing (GR), and two mutually exclusive ST behaviors, ruminating (RU) or resting (RE). AC was considered when cows exhibited horizontal displacement, either while walking with their head up (WA) or grazing with their head down while biting, chewing and searching for forage (GR). ST was considered when cows were either standing still or lying down, either fully resting (RE) or exhibiting gentle head, neck and jaw movements during rumination (RU). Other annotated activities included sleeping, scratching, grooming and nursing. Annotations were summarized to a 1-min interval to match the 1-min sensor MI data. A behavior state or class was annotated if the recorded behavior state or class lasted 30 seconds or more. If the observed behavior lasted less than 30 seconds, the previous label annotation was considered.