Developing serially dependent classifiers for animal behavioral analysis

Abstract

The increasing global demand for meat production necessitates more sustainable and data-driven livestock management practices. Precision Livestock Farming (PLF) leverages sensor technologies and computational methods to monitor individual animals, enhancing productivity and welfare. In this study, we investigate the use of supervised serially dependent classifiers for predicting free-range cattle behavior using time-series data from GPS and tri-axial accelerometers. Over a nine-week winter period, 22 Angus-based cows were monitored at the Red Bluff Research Ranch located in Montana, USA. Four features--step size, turning angle, and the mean and variance of Signal Vector Magnitude (SVM)---are engineered and used to classify grazing, resting, and traveling behaviors via Hidden Markov Models (HMMs) and Long Short-Term Memory (LSTM) recurrent neural networks. Model development was conducted using leave-one-out cross-validation to assess generalizability across individuals. The HMM achieved an average classification accuracy of 92.7%, and the two types of LSTMs achieved accuracies of 94.3%, and 95.4%. Each of the three types of models revealed ecologically meaningful patterns, including diurnal activity cycles, temperature-dependent behavioral shifts, and increased nighttime grazing during full moon phases. Generalized Linear Mixed Models (GLMMs) were used to expose the interaction effects of time of day and temperature on behavior. The findings support the utility of HMMs and LSTMs as effective tools for cattle behavior classification in extensive rangeland settings. The approach holds promise for real-time PLF applications such as early anomaly detection, resource allocation, and adaptive grazing management.