Approaches of machine learning and validation strategies to predict grazing behavior in beef cattle using sensors

Abstract

Machine learning approaches have been crucial for addressing current challenges in precision livestock, as it presents new tools for developing large scale predictive analytics in many fields including the area of sensor technology. In this context, the objectives of our study were to evaluate the following strategies of cross-validation used to predict grazing and not-grazing activities in grazing cattle. The machine learning approaches were generalizer linear regression (GLR), random forest (RF) and artificial neural network (ANN) as well as the cross-validation strategies evaluated were: 20% of the dataset randomly exclude to build the validation dataset (holdout), leave-one-animal-out (LOAO), and leave-one-day-out (LODO). Six Nellore bulls, 345 ± 21 kg body weight, were kept on pasture of Marandu Palisadegrass and had accelerometer and gyroscope sensor attached on neck. Animal behavior was registered through visual observation within a period of 10 hours for 15 days. The gyroscope record data were not used because a larger gap in a datapoint was observed. The overall accuracy of GLR, RF, and ANN were respectively 57.1%, 76.9%, and 74.2% in holdout validation, 53.1%, 58.7% and 72% in LOAO and 47.4%, 58.8% and 59.7% in LODO. GLR was not adequate model to predict animal behavior using our dataset. RF and ANN are more efficient to process complex dataset as these. Clearly, the validation strategy inferring in accuracy results and this is an important point in data analysis. Low values validation accuracy results in LODO shown us that predictive models are not adequate to use in different conditions of pasture. LOAO with ANN was the best validation strategy and it could predict animal behavior of different animals without used in predict model. Holdout validation, widely used in several similar studies, present an inflate accuracy values due to environmental conditions (e.g. animal or grazing conditions) that influence in dataset using in training and the validation dataset of the model.