Integrating LoRaWAN sensor networks and machine learning models to classify beef cattle behavior on arid rangelands of the southwestern United States

• Wireless tracking of cattle was applied using LoRaWAN sensor technology • Sensor power and data transmission needs were reduced using on-board processing • Cattle behavior was accurately classified using machine learning models • LoRa WAN sensors and machine learning enabled real-time monitoring of cattle behavior Monitoring cattle on large, often rugged, rangelands is a daunting task that can be improved using Long Range Wide Area Network (LoRaWAN) tracking and monitoring technology. This study tested the performance of five machine learning classifiers to discriminate between active and stationary states, and among walking, grazing, ruminating and resting behaviors of cattle. Models were trained and tested using a single motion index (MI) collected at 1-minute intervals by LoRaWAN cattle collars equipped with a Global Navigation Satellite System (GNSS) receptor and triaxial accelerometer. Twenty-four mature cows of four breeds were monitored across four periods between July and November 2022. Behavioral observations were made using 168 h of video records, which resulted in a dataset of 9,222 instances of labeled sensor data. Logistic regression, support vector machine, multilayer perceptron, XGBoost and random forest algorithms were trained and tested. No differences in MI were detected between ruminating and resting; therefore, subsequent model testing considered the combination of rumination and resting as one class. All classifiers correctly differentiated between the two states and among grazing, walking and resting behaviors with an accuracy and macro-F 1 scores of >0.95 and >0.90, respectively. The results suggest satisfactory application of trained models to monitor cattle behavior on desert rangeland. The annotated dataset used in this study is publicly available at Perea et al. [ 1 ].