Prediction of pigs' activity level using activity index with environmental and physical features through recurrent neural network models

Abstract

With the increasing attention toward smart farming and animal welfare, the study of automated animal behavior prediction becomes essential for the improvement of the previous studies on livestock welfare and farm management. This study aims to use Recurrent Neural Network models to predict pigs' activity level which is represented as the Activity Index extracted from video clips recorded on a smart farm in South Korea. The Recurrent Neural Network models are compared with models including one-dimension Convolutional Neural Network model, Autoregressive model, Support Vector Machine, and Random Forest. Evaluation Metrics including Mean Absolute Error, Mean Absolute Percentage Error, and Root Mean Square Error are used to compare the models performances. Environmental data such as temperature (°C) and relative humidity (\%) are collected. Physical data of the pigs including age (days) and weight (kilograms) are collected. Both environmental and physical features are combined with the Activity Index for the improvement of prediction accuracy. This study has found that Recurrent Neural Network models perform better than the other predictive models. The input of the Activity Index with the environmental feature yields better prediction with lower errors overall. This study benefits further research in the field of animal activity monitoring and prediction. Also, this study implements an unsupervised method of exploring livestock activity levels, which is time-efficient and better in generalization.