Evaluation of pre-processing methods for the prediction of cattle behaviour from accelerometer data

Abstract Monitoring livestock behaviour can be a useful way to improve farm animal management and to detect individual health events. The use of automated systems that predict several daily behaviours from accelerometer data is growing and studies have often focused on the comparison of datamining classification methods. However, few studies have attempted to evaluate the effect of the step preceding the classification, namely accelerometer signal pre-processing. This study aimed to evaluate the effect of several pre-processing methods on the prediction of dairy cows behaviour from accelerometer data. Ten Holstein cows equipped with a three-dimensional accelerometer fixed on a neck collar were simultaneously observed by two observers. Observed behaviours were predicted with decision tree using pre-processed accelerometer data as inputs. Different procedures were evaluated for each of these following pre-processing steps: signal filtering, signal segmentation and feature calculation. For signal filtering, low-pass filters and high-pass filters were considered. For signal segmentation, various window sizes and percentages of overlap between windows were implemented. Sixty-one features were computed. This resulted in 150 different combinations of pre-processing steps. For each combination a decision tree model predicting the observed behaviours was trained. The performance of each model to predict the observed behaviours was compared based on accuracy and F-score measures. The relative importance of each pre-processing configuration on the performance of prediction was evaluated with a linear regression model. The best configurations led to an accuracy of 0.95 and a F-score of 0.96 against 0.76 and 0.65 respectively with the worse combinations. The best combinations included a window size of 20 s and 30 s, with an overlap of 90% and no high-pass filter. High-pass filter had the most significant effect on the classification (P